diff --git a/.github/workflows/pages.yml b/.github/workflows/pages.yml deleted file mode 100644 index db2b399f..00000000 --- a/.github/workflows/pages.yml +++ /dev/null @@ -1,58 +0,0 @@ -name: Deploy Off Grid Docs - -on: - push: - branches: [main] - paths: ['website/**'] - workflow_dispatch: - -permissions: - contents: read - pages: write - id-token: write - -concurrency: - group: "pages" - cancel-in-progress: false - -jobs: - build: - runs-on: ubuntu-latest - steps: - - name: Checkout - uses: actions/checkout@v4 - - - name: Setup Ruby - uses: ruby/setup-ruby@v1 - with: - ruby-version: "3.3" - bundler-cache: true - working-directory: website - - - name: Setup Pages - uses: actions/configure-pages@v5 - - - name: Build with Jekyll - run: bundle exec jekyll build - working-directory: website - env: - JEKYLL_ENV: production - - - name: Index with Pagefind - run: npx pagefind --site website/_site - - - name: Upload artifact - uses: actions/upload-pages-artifact@v3 - with: - path: website/_site - - deploy: - environment: - name: github-pages - url: ${{ steps.deployment.outputs.page_url }} - runs-on: ubuntu-latest - needs: build - steps: - - name: Deploy to GitHub Pages - id: deployment - uses: actions/deploy-pages@v4 diff --git a/website/.gitignore b/website/.gitignore deleted file mode 100644 index 30b09485..00000000 --- a/website/.gitignore +++ /dev/null @@ -1,6 +0,0 @@ -# Jekyll build output and local bundler state - never commit these. -_site/ -.jekyll-cache/ -.jekyll-metadata -.bundle/ -vendor/ diff --git a/website/CNAME b/website/CNAME deleted file mode 100644 index 2386ada3..00000000 --- a/website/CNAME +++ /dev/null @@ -1 +0,0 @@ -offgridmobileai.co diff --git a/website/Gemfile b/website/Gemfile deleted file mode 100644 index 8e6923f6..00000000 --- a/website/Gemfile +++ /dev/null @@ -1,12 +0,0 @@ -source "https://rubygems.org" - -gem "jekyll", "~> 4.3" -gem "jekyll-seo-tag" -gem "jekyll-sitemap" - -platforms :mingw, :x64_mingw, :mswin, :jruby do - gem "tzinfo", ">= 1", "< 3" - gem "tzinfo-data" -end - -gem "wdm", "~> 0.1.1", :platforms => [:mingw, :x64_mingw, :mswin] diff --git a/website/Gemfile.lock b/website/Gemfile.lock deleted file mode 100644 index f2acc212..00000000 --- a/website/Gemfile.lock +++ /dev/null @@ -1,243 +0,0 @@ -GEM - remote: https://rubygems.org/ - specs: - addressable (2.9.0) - public_suffix (>= 2.0.2, < 8.0) - base64 (0.3.0) - bigdecimal (4.1.2) - colorator (1.1.0) - concurrent-ruby (1.3.6) - csv (3.3.5) - em-websocket (0.5.3) - eventmachine (>= 0.12.9) - http_parser.rb (~> 0) - eventmachine (1.2.7) - ffi (1.17.4) - ffi (1.17.4-aarch64-linux-gnu) - ffi (1.17.4-aarch64-linux-musl) - ffi (1.17.4-arm-linux-gnu) - ffi (1.17.4-arm-linux-musl) - ffi (1.17.4-arm64-darwin) - ffi (1.17.4-x86-linux-gnu) - ffi (1.17.4-x86-linux-musl) - ffi (1.17.4-x86_64-darwin) - ffi (1.17.4-x86_64-linux-gnu) - ffi (1.17.4-x86_64-linux-musl) - forwardable-extended (2.6.0) - google-protobuf (4.35.1) - bigdecimal - rake (~> 13.3) - google-protobuf (4.35.1-aarch64-linux-gnu) - bigdecimal - rake (~> 13.3) - google-protobuf (4.35.1-aarch64-linux-musl) - bigdecimal - rake (~> 13.3) - google-protobuf (4.35.1-arm64-darwin) - bigdecimal - rake (~> 13.3) - google-protobuf (4.35.1-x86-linux-gnu) - bigdecimal - rake (~> 13.3) - google-protobuf (4.35.1-x86-linux-musl) - bigdecimal - rake (~> 13.3) - google-protobuf (4.35.1-x86_64-darwin) - bigdecimal - rake (~> 13.3) - google-protobuf (4.35.1-x86_64-linux-gnu) - bigdecimal - rake (~> 13.3) - google-protobuf (4.35.1-x86_64-linux-musl) - bigdecimal - rake (~> 13.3) - http_parser.rb (0.8.1) - i18n (1.14.8) - concurrent-ruby (~> 1.0) - jekyll (4.4.1) - addressable (~> 2.4) - base64 (~> 0.2) - colorator (~> 1.0) - csv (~> 3.0) - em-websocket (~> 0.5) - i18n (~> 1.0) - jekyll-sass-converter (>= 2.0, < 4.0) - jekyll-watch (~> 2.0) - json (~> 2.6) - kramdown (~> 2.3, >= 2.3.1) - kramdown-parser-gfm (~> 1.0) - liquid (~> 4.0) - mercenary (~> 0.3, >= 0.3.6) - pathutil (~> 0.9) - rouge (>= 3.0, < 5.0) - safe_yaml (~> 1.0) - terminal-table (>= 1.8, < 4.0) - webrick (~> 1.7) - jekyll-sass-converter (3.1.0) - sass-embedded (~> 1.75) - jekyll-seo-tag (2.9.0) - jekyll (>= 3.8, < 5.0) - jekyll-sitemap (1.4.0) - jekyll (>= 3.7, < 5.0) - jekyll-watch (2.2.1) - listen (~> 3.0) - json (2.19.9) - kramdown (2.5.2) - rexml (>= 3.4.4) - kramdown-parser-gfm (1.1.0) - kramdown (~> 2.0) - liquid (4.0.4) - listen (3.10.0) - logger - rb-fsevent (~> 0.10, >= 0.10.3) - rb-inotify (~> 0.9, >= 0.9.10) - logger (1.7.0) - mercenary (0.4.0) - pathutil (0.16.2) - forwardable-extended (~> 2.6) - public_suffix (7.0.5) - rake (13.4.2) - rb-fsevent (0.11.2) - rb-inotify (0.11.1) - ffi (~> 1.0) - rexml (3.4.4) - rouge (4.7.0) - safe_yaml (1.0.5) - sass-embedded (1.101.0) - google-protobuf (~> 4.31) - rake (>= 13) - sass-embedded (1.101.0-aarch64-linux-android) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-aarch64-linux-gnu) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-aarch64-linux-musl) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-arm-linux-androideabi) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-arm-linux-gnueabihf) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-arm-linux-musleabihf) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-arm64-darwin) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-riscv64-linux-android) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-riscv64-linux-gnu) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-riscv64-linux-musl) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-x86_64-darwin) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-x86_64-linux-android) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-x86_64-linux-gnu) - google-protobuf (~> 4.31) - sass-embedded (1.101.0-x86_64-linux-musl) - google-protobuf (~> 4.31) - terminal-table (3.0.2) - unicode-display_width (>= 1.1.1, < 3) - unicode-display_width (2.6.0) - webrick (1.9.2) - -PLATFORMS - aarch64-linux-android - aarch64-linux-gnu - aarch64-linux-musl - arm-linux-androideabi - arm-linux-gnu - arm-linux-gnueabihf - arm-linux-musl - arm-linux-musleabihf - arm64-darwin - riscv64-linux-android - riscv64-linux-gnu - riscv64-linux-musl - ruby - x86-linux-gnu - x86-linux-musl - x86_64-darwin - x86_64-linux-android - x86_64-linux-gnu - x86_64-linux-musl - -DEPENDENCIES - jekyll (~> 4.3) - jekyll-seo-tag - jekyll-sitemap - tzinfo (>= 1, < 3) - tzinfo-data - wdm (~> 0.1.1) - -CHECKSUMS - addressable (2.9.0) sha256=7fdf6ac3660f7f4e867a0838be3f6cf722ace541dd97767fa42bc6cfa980c7af - base64 (0.3.0) sha256=27337aeabad6ffae05c265c450490628ef3ebd4b67be58257393227588f5a97b - bigdecimal (4.1.2) sha256=53d217666027eab4280346fba98e7d5b66baaae1b9c3c1c0ffe89d48188a3fbd - colorator (1.1.0) sha256=e2f85daf57af47d740db2a32191d1bdfb0f6503a0dfbc8327d0c9154d5ddfc38 - concurrent-ruby (1.3.6) sha256=6b56837e1e7e5292f9864f34b69c5a2cbc75c0cf5338f1ce9903d10fa762d5ab - csv (3.3.5) sha256=6e5134ac3383ef728b7f02725d9872934f523cb40b961479f69cf3afa6c8e73f - em-websocket (0.5.3) sha256=f56a92bde4e6cb879256d58ee31f124181f68f8887bd14d53d5d9a292758c6a8 - eventmachine (1.2.7) sha256=994016e42aa041477ba9cff45cbe50de2047f25dd418eba003e84f0d16560972 - ffi (1.17.4) sha256=bcd1642e06f0d16fc9e09ac6d49c3a7298b9789bcb58127302f934e437d60acf - ffi (1.17.4-aarch64-linux-gnu) sha256=b208f06f91ffd8f5e1193da3cae3d2ccfc27fc36fba577baf698d26d91c080df - ffi (1.17.4-aarch64-linux-musl) sha256=9286b7a615f2676245283aef0a0a3b475ae3aae2bb5448baace630bb77b91f39 - ffi (1.17.4-arm-linux-gnu) sha256=d6dbddf7cb77bf955411af5f187a65b8cd378cb003c15c05697f5feee1cb1564 - ffi (1.17.4-arm-linux-musl) sha256=9d4838ded0465bef6e2426935f6bcc93134b6616785a84ffd2a3d82bc3cf6f95 - ffi (1.17.4-arm64-darwin) sha256=19071aaf1419251b0a46852abf960e77330a3b334d13a4ab51d58b31a937001b - ffi (1.17.4-x86-linux-gnu) sha256=38e150df5f4ca555e25beca4090823ae09657bceded154e3c52f8631c1ed72cf - ffi (1.17.4-x86-linux-musl) sha256=fbeec0fc7c795bcf86f623bb18d31ea1820f7bd580e1703a3d3740d527437809 - ffi (1.17.4-x86_64-darwin) sha256=aa70390523cf3235096cf64962b709b4cfbd5c082a2cb2ae714eb0fe2ccda496 - ffi (1.17.4-x86_64-linux-gnu) sha256=9d3db14c2eae074b382fa9c083fe95aec6e0a1451da249eab096c34002bc752d - ffi (1.17.4-x86_64-linux-musl) sha256=3fdf9888483de005f8ef8d1cf2d3b20d86626af206cbf780f6a6a12439a9c49e - forwardable-extended (2.6.0) sha256=1bec948c469bbddfadeb3bd90eb8c85f6e627a412a3e852acfd7eaedbac3ec97 - google-protobuf (4.35.1) sha256=a3a6471331d918f58dfa4d014a8f6286f0af2cf4840216bde52fcf2ea3fe3726 - google-protobuf (4.35.1-aarch64-linux-gnu) sha256=50ca44d0eeff3f8475e630a1accdd974256f3510694d574e2c9d6119ea8bc9e1 - google-protobuf (4.35.1-aarch64-linux-musl) sha256=d5c65cef6bd6498a9e5ed5f88cf6cf7e341c10b0a005e32137d5d1a2b6e8c18a - google-protobuf (4.35.1-arm64-darwin) sha256=d9c957df04fa89c749fa9a72a7b383eb4296efc9b2303dc6fd6fbe39c698ad6b - google-protobuf (4.35.1-x86-linux-gnu) sha256=cc7492566b27ad8b5dfa3a7b6f9b11e905050cc53c2fa8ff22de375a9e7a70fb - google-protobuf (4.35.1-x86-linux-musl) sha256=ab403790b59e4dc588ffbed1eaacf05d4ca2f0d12ac9c13d6c64e69380d8c99e - google-protobuf (4.35.1-x86_64-darwin) sha256=66b62b4df00931018a692806df66393efa960d6d2b7da69735187249f950d3ee - google-protobuf (4.35.1-x86_64-linux-gnu) sha256=c786439087512a3fbd199e9897d265b855f951d4027e218ea55e858d45969edd - google-protobuf (4.35.1-x86_64-linux-musl) sha256=91890eb0002934a339fdb7d77a147c46b7474b6799db27872b747b905837f744 - http_parser.rb (0.8.1) sha256=9ae8df145b39aa5398b2f90090d651c67bd8e2ebfe4507c966579f641e11097a - i18n (1.14.8) sha256=285778639134865c5e0f6269e0b818256017e8cde89993fdfcbfb64d088824a5 - jekyll (4.4.1) sha256=4c1144d857a5b2b80d45b8cf5138289579a9f8136aadfa6dd684b31fe2bc18c1 - jekyll-sass-converter (3.1.0) sha256=83925d84f1d134410c11d0c6643b0093e82e3a3cf127e90757a85294a3862443 - jekyll-seo-tag (2.9.0) sha256=0260015a8e1df9bf195cdfb0c675b7b2883fd8cbf12556e1c1cbe36a831c6852 - jekyll-sitemap (1.4.0) sha256=0de08c5debc185ea5a8f980e1025c7cd3f8e0c35c8b6ef592f15c46235cf4218 - jekyll-watch (2.2.1) sha256=bc44ed43f5e0a552836245a54dbff3ea7421ecc2856707e8a1ee203a8387a7e1 - json (2.19.9) sha256=9b9025b7cdddafa38d316eca0b2358488e42d417045c1b90d216a9fefe46b79a - kramdown (2.5.2) sha256=1ba542204c66b6f9111ff00dcc26075b95b220b07f2905d8261740c82f7f02fa - kramdown-parser-gfm (1.1.0) sha256=fb39745516427d2988543bf01fc4cf0ab1149476382393e0e9c48592f6581729 - liquid (4.0.4) sha256=4fcfebb1a045e47918388dbb7a0925e7c3893e58d2bd6c3b3c73ec17a2d8fdb3 - listen (3.10.0) sha256=c6e182db62143aeccc2e1960033bebe7445309c7272061979bb098d03760c9d2 - logger (1.7.0) sha256=196edec7cc44b66cfb40f9755ce11b392f21f7967696af15d274dde7edff0203 - mercenary (0.4.0) sha256=b25a1e4a59adca88665e08e24acf0af30da5b5d859f7d8f38fba52c28f405138 - pathutil (0.16.2) sha256=e43b74365631cab4f6d5e4228f812927efc9cb2c71e62976edcb252ee948d589 - public_suffix (7.0.5) sha256=1a8bb08f1bbea19228d3bed6e5ed908d1cb4f7c2726d18bd9cadf60bc676f623 - rake (13.4.2) sha256=cb825b2bd5f1f8e91ca37bddb4b9aaf345551b4731da62949be002fa89283701 - rb-fsevent (0.11.2) sha256=43900b972e7301d6570f64b850a5aa67833ee7d87b458ee92805d56b7318aefe - rb-inotify (0.11.1) sha256=a0a700441239b0ff18eb65e3866236cd78613d6b9f78fea1f9ac47a85e47be6e - rexml (3.4.4) sha256=19e0a2c3425dfbf2d4fc1189747bdb2f849b6c5e74180401b15734bc97b5d142 - rouge (4.7.0) sha256=dba5896715c0325c362e895460a6d350803dbf6427454f49a47500f3193ea739 - safe_yaml (1.0.5) sha256=a6ac2d64b7eb027bdeeca1851fe7e7af0d668e133e8a88066a0c6f7087d9f848 - sass-embedded (1.101.0) sha256=57dbc3409e2c0a2c581a4c9945c2bd72ec88e71ec98017bd02dd1da8a76b22f4 - sass-embedded (1.101.0-aarch64-linux-android) sha256=ad0a35c6ff5cdc4e31c5d8261a768ef460c6c7d1458081183e42170e1474c4b5 - sass-embedded (1.101.0-aarch64-linux-gnu) sha256=8f6926349e880dbb4fda75fb182086fb60aa821b756bcbfc8e5b2b23b1f269d6 - sass-embedded (1.101.0-aarch64-linux-musl) sha256=0a02b4b75db305160db1b0512f040762242128a9b5c01b0fa7a504f2b24557f2 - sass-embedded (1.101.0-arm-linux-androideabi) sha256=88e762531f90d4a2eef55ffafeba337585e1bb24b22307ebd7fe5505de9a4476 - sass-embedded (1.101.0-arm-linux-gnueabihf) sha256=358df7f98d13ff53739ebc45f5c5acd4dd96833bf87f39c74d11f798081d4158 - sass-embedded (1.101.0-arm-linux-musleabihf) sha256=d993a3ad017250d46d1312bc10cfded5ab4585906d37a60a56494bd662182d3a - sass-embedded (1.101.0-arm64-darwin) sha256=9fed684380b49499dfc856aba0d026a0748f924bd78044ffff2bae1537aea73e - sass-embedded (1.101.0-riscv64-linux-android) sha256=1bd35527e1ca24af6aa2ba9db797bc9b72331ebbec9ff6e99ea02b24f83b3da1 - sass-embedded (1.101.0-riscv64-linux-gnu) sha256=7805a249ed4c31b58b2f7d054d14f0ed24c61a836f579cd76ed2367088142bcf - sass-embedded (1.101.0-riscv64-linux-musl) sha256=8bd60eb9ebe6f9f2740a4e9ddb792a3db1fbba282e3375103844b4979eb7c731 - sass-embedded (1.101.0-x86_64-darwin) sha256=20ff4afd7c052b3f8d7b4abe511e8114985a07dff7616750a43ee1fc2759fb28 - sass-embedded (1.101.0-x86_64-linux-android) sha256=472b72114ededa00a22e13553dd31f98d1a92d52bdbe2e738a4dbac2dd50c779 - sass-embedded (1.101.0-x86_64-linux-gnu) sha256=ff48452b2351eaaf4e6e02f59de0e9e35f6f163e8456d520db49b4d87cf73c27 - sass-embedded (1.101.0-x86_64-linux-musl) sha256=2286feef3c7a8d9bbb075aa1651e1a81942aca1213f7c2d30cc1d6f4eb5b4104 - terminal-table (3.0.2) sha256=f951b6af5f3e00203fb290a669e0a85c5dd5b051b3b023392ccfd67ba5abae91 - unicode-display_width (2.6.0) sha256=12279874bba6d5e4d2728cef814b19197dbb10d7a7837a869bab65da943b7f5a - webrick (1.9.2) sha256=beb4a15fc474defed24a3bda4ffd88a490d517c9e4e6118c3edce59e45864131 - -BUNDLED WITH - 4.0.11 diff --git a/website/_config.yml b/website/_config.yml deleted file mode 100644 index 329d5927..00000000 --- a/website/_config.yml +++ /dev/null @@ -1,37 +0,0 @@ -title: Off Grid -description: Run powerful AI models directly on your iPhone or Android phone. No internet required, no subscriptions, no cloud. Your data never leaves your device. -url: https://offgridmobileai.co -baseurl: "" - -permalink: pretty - -markdown: kramdown -kramdown: - input: GFM - syntax_highlighter: rouge - syntax_highlighter_opts: - css_class: highlight - -plugins: - - jekyll-seo-tag - - jekyll-sitemap - -posthog_token: "phc_u7cnV9P3cTovsfPTE2nhB7g5G4qdtZJ5dgMzv7ryfKWs" - -# RevenueCat Web Purchase Link id (the path segment in pay.rev.cat//...). -# This is public by design - it appears in every customer's checkout URL, so it -# is safe to commit. It is a link path id, not a secret/auth token. The pay page -# builds: https://pay.rev.cat//?email= -revenuecat_link_id: "avvnmcnfsgbmjaee" - -logo: /assets/logo.png -cover: /assets/cover.png - -exclude: - - README.md - - Gemfile - - Gemfile.lock - - node_modules - - vendor - - .bundle - - _site diff --git a/website/_layouts/default.html b/website/_layouts/default.html deleted file mode 100644 index 6345e712..00000000 --- a/website/_layouts/default.html +++ /dev/null @@ -1,461 +0,0 @@ - - - - - - {% if page.title == "Home" %}Off Grid — Run AI Locally on Your Phone{% else %}{{ page.title }} — Off Grid{% endif %} - - - - - - - - - - - - - - - {% seo %} - - - - - - - - - {% if page.parent %} - {% assign parent_page = site.pages | where: "title", page.parent | first %} - - {% endif %} - - - {% if page.faq %} - - {% endif %} - - - -
- -
- - - -
-
- - - - -
- -
- - - - -
- {% if page.parent %} - - {% endif %} - -
- Did this land? -
- - -
- -
- -
- {{ content }} -
- - {% if page.parent == "Perspectives" %} -
-
- Run the personal AI OS on your phone. - Off Grid Pro is $50 one-time. No subscription. -
- Get Pro -
- {% endif %} - - {% if page.parent %} - {% assign siblings = site.pages | where: "parent", page.parent | sort: "nav_order" %} - {% assign current_index = 0 %} - {% for s in siblings %} - {% if s.url == page.url %}{% assign current_index = forloop.index0 %}{% endif %} - {% endfor %} - {% assign prev_index = current_index | minus: 1 %} - {% assign next_index = current_index | plus: 1 %} - - {% endif %} -
-
- - - - - - - - - - - - - - - - - - diff --git a/website/assets/apple-touch-icon.png b/website/assets/apple-touch-icon.png deleted file mode 100644 index 96a35493..00000000 Binary files a/website/assets/apple-touch-icon.png and /dev/null differ diff --git a/website/assets/cover.png b/website/assets/cover.png deleted file mode 100644 index 7564786a..00000000 Binary files a/website/assets/cover.png and /dev/null differ diff --git a/website/assets/css/main.css b/website/assets/css/main.css deleted file mode 100644 index 2e79fd9e..00000000 --- a/website/assets/css/main.css +++ /dev/null @@ -1,821 +0,0 @@ -/* ─── Reset & Base ─────────────────────────────────────────────────────────── */ -*, *::before, *::after { box-sizing: border-box; margin: 0; padding: 0; } - -:root { - --accent: #16a34a; - --accent-hover: #15803d; - --accent-subtle: #f0fdf4; - --accent-subtle-border: #bbf7d0; - --text-primary: #111827; - --text-secondary: #374151; - --text-muted: #6b7280; - --border: #e5e7eb; - --border-light: #f3f4f6; - --bg: #ffffff; - --bg-subtle: #f9fafb; - --bg-hover: #f3f4f6; - --code-bg: #f3f4f6; - --shadow-sm: 0 1px 3px rgba(0,0,0,0.08); - --shadow-md: 0 4px 16px rgba(0,0,0,0.08); - --sidebar-width: 256px; - --font: "Inter", -apple-system, BlinkMacSystemFont, "Segoe UI", system-ui, sans-serif; - --font-mono: "SF Mono", ui-monospace, "Cascadia Code", "Fira Code", monospace; - --radius: 8px; -} - -@media (prefers-color-scheme: dark) { - :root:not([data-theme="light"]) { - --text-primary: #f9fafb; - --text-secondary: #d1d5db; - --text-muted: #9ca3af; - --border: #374151; - --border-light: #1f2937; - --bg: #111827; - --bg-subtle: #1f2937; - --bg-hover: #374151; - --code-bg: #1f2937; - --accent-subtle: #052e16; - --accent-subtle-border: #14532d; - --shadow-sm: 0 1px 3px rgba(0,0,0,0.3); - --shadow-md: 0 4px 16px rgba(0,0,0,0.4); - } -} - -[data-theme="dark"] { - --text-primary: #f9fafb; - --text-secondary: #d1d5db; - --text-muted: #9ca3af; - --border: #374151; - --border-light: #1f2937; - --bg: #111827; - --bg-subtle: #1f2937; - --bg-hover: #374151; - --code-bg: #1f2937; - --accent-subtle: #052e16; - --accent-subtle-border: #14532d; - --shadow-sm: 0 1px 3px rgba(0,0,0,0.3); - --shadow-md: 0 4px 16px rgba(0,0,0,0.4); -} - -html { font-size: 16px; -webkit-font-smoothing: antialiased; } -body { font-family: var(--font); color: var(--text-primary); background: var(--bg); line-height: 1.6; transition: background 0.2s, color 0.2s; } - -/* ─── Layout ────────────────────────────────────────────────────────────────── */ -.layout { display: flex; min-height: 100vh; } - -/* ─── Sidebar ────────────────────────────────────────────────────────────────── */ -.sidebar { - width: var(--sidebar-width); - flex-shrink: 0; - border-right: 1px solid var(--border); - background: var(--bg-subtle); - display: flex; - flex-direction: column; - position: sticky; - top: 0; - height: 100vh; - overflow-y: auto; -} - -.sidebar-logo { padding: 18px 16px 12px; display: flex; align-items: center; justify-content: space-between; } -.sidebar-logo a { display: flex; align-items: center; gap: 9px; text-decoration: none; color: var(--text-primary); } -.sidebar-logo img { border-radius: 7px; flex-shrink: 0; } -.logo-text { font-size: 0.9375rem; font-weight: 600; letter-spacing: -0.01em; color: var(--text-primary); } - -/* Theme toggle */ -.theme-toggle { - display: flex; - align-items: center; - justify-content: center; - width: 28px; - height: 28px; - background: none; - border: 1px solid var(--border); - border-radius: 6px; - cursor: pointer; - color: var(--text-muted); - flex-shrink: 0; - transition: border-color 0.15s, color 0.15s, background 0.15s; -} -.theme-toggle:hover { border-color: var(--accent); color: var(--accent); background: var(--accent-subtle); } -.theme-toggle .icon-sun { display: none; } -.theme-toggle .icon-moon { display: block; } -[data-theme="dark"] .theme-toggle .icon-sun { display: block; } -[data-theme="dark"] .theme-toggle .icon-moon { display: none; } -@media (prefers-color-scheme: dark) { - :root:not([data-theme="light"]) .theme-toggle .icon-sun { display: block; } - :root:not([data-theme="light"]) .theme-toggle .icon-moon { display: none; } -} - -.search-trigger { - display: flex; - align-items: center; - gap: 7px; - width: calc(100% - 32px); - margin: 0 16px 8px; - padding: 7px 10px; - background: var(--bg); - border: 1px solid var(--border); - border-radius: var(--radius); - color: var(--text-muted); - font-size: 0.8125rem; - font-family: var(--font); - cursor: pointer; - text-align: left; - transition: border-color 0.15s; -} -.search-trigger:hover { border-color: var(--accent); color: var(--text-secondary); } -.search-trigger span { flex: 1; } -.search-trigger kbd { - font-size: 0.625rem; - font-family: var(--font-mono); - background: var(--bg-subtle); - border: 1px solid var(--border); - border-radius: 4px; - padding: 1px 5px; - color: var(--text-muted); -} - -.sidebar-nav { flex: 1; padding: 4px 0 16px; overflow-y: auto; } - -.nav-item { - display: flex; - align-items: center; - padding: 6px 16px; - font-size: 0.875rem; - color: var(--text-secondary); - text-decoration: none; - transition: color 0.12s, background 0.12s; - gap: 6px; - border-radius: 0; -} -.nav-item:hover { color: var(--text-primary); background: var(--bg-hover); } -.nav-item.active { color: var(--accent); font-weight: 600; } -.nav-item-wrap { display: flex; align-items: stretch; width: 100%; } -.nav-item-wrap .nav-item { flex: 1; padding-right: 4px; } -.nav-toggle { background: none; border: none; padding: 0 14px 0 10px; cursor: pointer; color: var(--text-muted); display: flex; align-items: center; flex-shrink: 0; min-width: 36px; } -.nav-toggle .chevron { transition: transform 0.2s; pointer-events: none; } -.nav-toggle.open .chevron { transform: rotate(180deg); } - -.nav-children { display: none; padding-left: 12px; } -.nav-children.open { display: block; } -.nav-child { - display: block; - padding: 5px 16px; - font-size: 0.8125rem; - color: var(--text-muted); - text-decoration: none; - border-left: 2px solid var(--border); - margin-left: 16px; - transition: color 0.12s, border-color 0.12s; -} -.nav-child:hover { color: var(--text-primary); border-left-color: var(--text-muted); } -.nav-child.active { color: var(--accent); border-left-color: var(--accent); font-weight: 600; } - -.sidebar-footer { - padding: 14px 16px; - border-top: 1px solid var(--border); - font-size: 0.75rem; -} - -/* Sidebar store buttons */ -.sidebar-store-btns { display: flex; gap: 6px; margin-bottom: 7px; } -.sidebar-cta { - display: flex; - align-items: center; - justify-content: center; - gap: 6px; - padding: 8px 10px; - background: var(--accent); - color: #fff; - border-radius: var(--radius); - text-decoration: none; - font-size: 0.75rem; - font-weight: 600; - flex: 1; - transition: background 0.15s; - letter-spacing: -0.01em; -} -.sidebar-cta:hover { background: var(--accent-hover); color: #fff; } -.sidebar-cta-android { - background: transparent; - color: var(--accent); - border: 1.5px solid var(--accent); -} -.sidebar-cta-android:hover { background: var(--accent-subtle); color: var(--accent); } - -.sidebar-links { display: flex; gap: 12px; margin: 10px 0 8px; } -.sidebar-links a { display: flex; align-items: center; gap: 5px; font-size: 0.75rem; color: var(--text-muted); text-decoration: none; } -.sidebar-links a:hover { color: var(--text-primary); } -.sidebar-links svg { flex-shrink: 0; } -.sidebar-copy { color: var(--text-muted); font-size: 0.7rem; margin-top: 2px; } -.sidebar-copy a { color: var(--text-muted); text-decoration: underline; } -.sidebar-copy a:hover { color: var(--text-primary); } - -/* ─── Newsletter form ────────────────────────────────────────────────────────── */ -.newsletter-form { margin: 10px 0 10px; } -.newsletter-label { font-size: 0.72rem; color: var(--text-muted); margin-bottom: 6px; } -.newsletter-form form { display: flex; flex-direction: column; gap: 6px; } -.newsletter-form input[type=email] { - width: 100%; - font-family: var(--font); - font-size: 0.8rem; - padding: 7px 10px; - border: 1px solid var(--border); - border-radius: 6px; - background: var(--bg); - color: var(--text-primary); - outline: none; -} -.newsletter-form input[type=email]::placeholder { color: var(--text-muted); } -.newsletter-form input[type=email]:focus { border-color: var(--accent); } -.newsletter-form button { - width: 100%; - font-family: var(--font); - font-size: 0.8rem; - font-weight: 600; - padding: 7px 10px; - background: var(--accent); - color: #fff; - border: none; - border-radius: 6px; - cursor: pointer; - transition: background 0.15s; -} -.newsletter-form button:hover { background: var(--accent-hover); } -.newsletter-form button:disabled { background: var(--text-muted); cursor: default; } -.newsletter-status { font-size: 0.7rem; margin-top: 5px; min-height: 1em; } -.newsletter-status.success { color: var(--accent); } -.newsletter-status.error { color: #ef4444; } - -/* ─── Pro waitlist (inline) ──────────────────────────────────────────────────── */ -.pro-waitlist-form { - display: flex; - flex-wrap: wrap; - align-items: center; - gap: 10px; - margin: 20px 0 8px; - padding: 16px; - border: 1px solid var(--border); - border-radius: 10px; - background: var(--bg-elevated, var(--bg)); -} -.pro-waitlist-form input[type=email] { - flex: 1 1 220px; - font-family: var(--font); - font-size: 0.95rem; - padding: 10px 14px; - border: 1px solid var(--border); - border-radius: 8px; - background: var(--bg); - color: var(--text-primary); - outline: none; -} -.pro-waitlist-form input[type=email]::placeholder { color: var(--text-muted); } -.pro-waitlist-form input[type=email]:focus { border-color: var(--accent); } -.pro-waitlist-form button { - flex: 0 0 auto; - font-size: 0.95rem; - padding: 10px 18px; -} -.pro-waitlist-meta { - flex-basis: 100%; - font-size: 0.78rem; - color: var(--text-muted); - margin: 0; -} -.pro-waitlist-status { flex-basis: 100%; font-size: 0.8rem; margin: 0; min-height: 1em; } -.pro-waitlist-status.success { color: var(--accent); } -.pro-waitlist-status.error { color: #ef4444; } - -/* ─── Main Content ─────────────────────────────────────────────────────────── */ -.main { - flex: 1; - min-width: 0; - padding: 48px 56px 80px; - max-width: 860px; -} - -/* ─── Breadcrumb ────────────────────────────────────────────────────────────── */ -.breadcrumb { - display: flex; - align-items: center; - gap: 6px; - font-size: 0.8125rem; - color: var(--text-muted); -} -.breadcrumb a { color: var(--text-muted); text-decoration: none; } -.breadcrumb a:hover { color: var(--text-primary); } - -/* ─── Content Typography ────────────────────────────────────────────────────── */ -.content h1 { - font-size: 1.875rem; - font-weight: 700; - letter-spacing: -0.03em; - line-height: 1.2; - color: var(--text-primary); - margin-bottom: 12px; - margin-top: 0; -} -.content h2 { - font-size: 1.25rem; - font-weight: 600; - letter-spacing: -0.02em; - color: var(--text-primary); - margin-top: 48px; - margin-bottom: 12px; - padding-bottom: 8px; - border-bottom: 1px solid var(--border-light); - position: relative; -} -.content h3 { - font-size: 1rem; - font-weight: 600; - color: var(--text-primary); - margin-top: 28px; - margin-bottom: 8px; - letter-spacing: -0.01em; -} -.content h4 { - font-size: 0.9375rem; - font-weight: 600; - color: var(--text-secondary); - margin-top: 20px; - margin-bottom: 6px; -} -.content p { margin-bottom: 16px; color: var(--text-secondary); font-size: 0.9375rem; line-height: 1.7; } -.content strong { color: var(--text-primary); font-weight: 600; } -.content a:not(.btn) { color: var(--accent); text-decoration: underline; text-decoration-color: var(--accent-subtle-border); } -.content a:not(.btn):hover { text-decoration-color: var(--accent); } -.content ul, .content ol { margin: 12px 0 16px 20px; } -.content li { margin-bottom: 6px; font-size: 0.9375rem; color: var(--text-secondary); line-height: 1.65; } - -.content code { - font-family: var(--font-mono); - font-size: 0.8125rem; - background: var(--code-bg); - border: 1px solid var(--border); - border-radius: 4px; - padding: 1px 5px; - color: var(--text-primary); -} -.content pre { - background: #0d1117; - border-radius: 10px; - padding: 20px 24px; - overflow-x: auto; - margin: 20px 0; - border: 1px solid #21262d; -} -.content pre code { - font-size: 0.8125rem; - background: none; - border: none; - padding: 0; - color: #e6edf3; -} - -.content table { width: 100%; border-collapse: collapse; margin: 20px 0; font-size: 0.875rem; } -.content th { text-align: left; font-weight: 600; font-size: 0.8125rem; padding: 9px 12px; border-bottom: 2px solid var(--border); color: var(--text-primary); } -.content td { padding: 9px 12px; border-bottom: 1px solid var(--border-light); color: var(--text-secondary); vertical-align: top; } -.content tr:hover td { background: var(--bg-subtle); } - -.content blockquote { - border-left: 3px solid var(--accent); - margin: 20px 0; - padding: 12px 20px; - background: var(--accent-subtle); - border-radius: 0 var(--radius) var(--radius) 0; - border: 1px solid var(--accent-subtle-border); - border-left: 3px solid var(--accent); -} -.content blockquote p { color: var(--text-secondary); margin: 0; font-size: 0.9375rem; } - -.content hr { border: none; border-top: 1px solid var(--border); margin: 40px 0; } - -/* Heading anchors */ -.heading-anchor { margin-left: 8px; color: var(--border); font-size: 0.875rem; text-decoration: none; opacity: 0; transition: opacity 0.15s; } -.content h2:hover .heading-anchor, -.content h3:hover .heading-anchor, -.content h4:hover .heading-anchor { opacity: 1; color: var(--text-muted); } - -/* ─── Hero & Buttons ────────────────────────────────────────────────────────── */ -.hero-cover { - display: block; - width: 100%; - border-radius: 12px; - margin-bottom: 28px; - box-shadow: var(--shadow-md); -} - -.page-title-row { - display: flex; - align-items: center; - gap: 12px; - margin-bottom: 10px; -} -.page-title-row img { border-radius: 10px; flex-shrink: 0; } -.page-title-row h1 { margin-bottom: 0; } - -.hero-buttons { display: flex; flex-wrap: wrap; gap: 8px; margin: 20px 0 32px; } - -.btn { - display: inline-flex; - align-items: center; - gap: 7px; - padding: 10px 18px; - border-radius: var(--radius); - font-family: var(--font); - font-size: 0.875rem; - font-weight: 600; - text-decoration: none; - letter-spacing: -0.01em; - transition: background 0.15s, color 0.15s, border-color 0.15s, box-shadow 0.15s; - border: 1.5px solid transparent; - cursor: pointer; - line-height: 1; -} -.btn svg { flex-shrink: 0; } -.btn-green { - background: var(--accent); - color: #fff; - border-color: var(--accent); -} -.btn-green:hover { background: var(--accent-hover); border-color: var(--accent-hover); color: #fff; } -.btn-outline { - background: transparent; - color: var(--text-primary); - border-color: var(--border); -} -.btn-outline:hover { border-color: var(--text-muted); background: var(--bg-subtle); color: var(--text-primary); } - -/* ─── Guide card grid ───────────────────────────────────────────────────────── */ -.guide-grid { - display: grid; - grid-template-columns: repeat(auto-fill, minmax(240px, 1fr)); - gap: 12px; - margin: 16px 0 8px; -} -.guide-card { - display: flex; - flex-direction: column; - gap: 5px; - padding: 16px 18px; - border: 1px solid var(--border); - border-radius: 10px; - text-decoration: none; - background: var(--bg-subtle); - transition: border-color 0.15s, background 0.15s, box-shadow 0.15s; -} -.guide-card:hover { - border-color: var(--accent); - background: var(--accent-subtle); - box-shadow: var(--shadow-sm); - text-decoration: none; -} -.guide-card-title { - font-size: 0.875rem; - font-weight: 600; - color: var(--text-primary); - letter-spacing: -0.01em; - line-height: 1.3; -} -.guide-card-desc { - font-size: 0.8125rem; - color: var(--text-muted); - line-height: 1.5; -} - -/* ─── Prev/Next nav ─────────────────────────────────────────────────────────── */ -.page-nav { display: flex; justify-content: space-between; gap: 16px; margin-top: 56px; padding-top: 24px; border-top: 1px solid var(--border-light); } -.page-nav-item { display: flex; flex-direction: column; gap: 4px; padding: 14px 18px; border: 1px solid var(--border); border-radius: 10px; text-decoration: none; flex: 1; transition: border-color 0.15s, background 0.15s; } -.page-nav-item:hover { border-color: var(--accent); background: var(--accent-subtle); } -.page-nav-item.next { text-align: right; } -.page-nav-label { font-size: 0.6875rem; color: var(--text-muted); font-weight: 600; letter-spacing: 0.06em; text-transform: uppercase; } -.page-nav-title { font-size: 0.875rem; font-weight: 600; color: var(--text-primary); letter-spacing: -0.01em; } - -/* ─── Search modal ──────────────────────────────────────────────────────────── */ -.search-modal { position: fixed; inset: 0; z-index: 1000; display: flex; align-items: flex-start; justify-content: center; padding-top: 72px; } -.search-modal[hidden] { display: none; } -.search-modal-backdrop { position: fixed; inset: 0; background: rgba(0,0,0,0.5); backdrop-filter: blur(3px); } -.search-modal-box { - position: relative; - z-index: 1; - width: 100%; - max-width: 580px; - margin: 0 20px; - background: var(--bg); - border-radius: 12px; - box-shadow: 0 24px 80px rgba(0,0,0,0.28); - overflow: hidden; - border: 1px solid var(--border); -} -#pagefind-search { --pagefind-ui-font: var(--font); } -#pagefind-search .pagefind-ui__search-input { font-family: var(--font) !important; font-size: 1rem !important; padding: 18px 20px 18px 48px !important; border: none !important; border-bottom: 1px solid var(--border) !important; border-radius: 0 !important; outline: none !important; box-shadow: none !important; background: var(--bg) !important; width: 100% !important; color: var(--text-primary) !important; } -#pagefind-search .pagefind-ui__results { max-height: 440px; overflow-y: auto; padding: 6px 0 10px; } -#pagefind-search .pagefind-ui__result { padding: 12px 20px; border-bottom: 1px solid var(--border-light); list-style: none; } -#pagefind-search .pagefind-ui__result:hover { background: var(--bg-subtle); } -#pagefind-search .pagefind-ui__result-link { font-family: var(--font) !important; font-size: 0.9375rem !important; font-weight: 600 !important; color: var(--text-primary) !important; text-decoration: none !important; } -#pagefind-search .pagefind-ui__result-link:hover { color: var(--accent) !important; } -#pagefind-search .pagefind-ui__result-excerpt { font-family: var(--font) !important; font-size: 0.8125rem !important; color: var(--text-muted) !important; margin-top: 3px !important; line-height: 1.5 !important; } -#pagefind-search mark { background: var(--accent-subtle); color: var(--accent); border-radius: 2px; padding: 0 2px; } - -/* ─── Breadcrumb margin ─────────────────────────────────────────────────────── */ -.breadcrumb { margin-bottom: 32px; } - -/* ─── Essay Reactions — floating pill ──────────────────────────────────────── */ -.essay-reactions { - position: fixed; - bottom: 28px; - right: 28px; - display: flex; - align-items: center; - gap: 8px; - padding: 8px 12px 8px 14px; - background: var(--bg); - border: 1px solid var(--border); - border-radius: 999px; - box-shadow: 0 4px 20px rgba(0,0,0,0.12); - z-index: 50; -} -.essay-reactions-label { - font-size: 0.75rem; - color: var(--text-muted); - white-space: nowrap; -} -.essay-reactions-buttons { - display: flex; - gap: 4px; -} -.reaction-btn { - display: flex; - align-items: center; - justify-content: center; - width: 30px; - height: 30px; - padding: 0; - background: transparent; - border: 1px solid var(--border); - border-radius: 50%; - color: var(--text-muted); - cursor: pointer; - transition: border-color 0.15s, background 0.15s, color 0.15s; -} -.reaction-btn:hover { - border-color: var(--accent); - color: var(--accent); - background: var(--accent-subtle); -} -.reaction-btn.active { - border-color: var(--accent); - background: var(--accent-subtle); - color: var(--accent); -} -.essay-reaction-thanks { - font-size: 0.75rem; - color: var(--text-muted); - white-space: nowrap; -} -@media (max-width: 768px) { - .essay-reactions { bottom: 16px; right: 16px; } -} - -/* ─── Mission Page ──────────────────────────────────────────────────────────── */ -.mission-statement { - text-align: center; - padding: 48px 0 40px; -} -.mission-tagline { - font-size: 2rem; - line-height: 1.15; - letter-spacing: -0.03em; - font-weight: 400; - color: var(--text-primary); - margin: 0 0 12px; -} -.mission-sub { - font-size: 1.0625rem; - color: var(--text-secondary); - margin: 0; - font-weight: 400; -} -@media (max-width: 640px) { - .mission-tagline { font-size: 1.5rem; } -} - -/* ─── Early Access Page ─────────────────────────────────────────────────────── */ -.early-access-hero { margin-bottom: 40px; } -.early-access-badge { - display: inline-block; - font-size: 0.6875rem; - font-weight: 600; - letter-spacing: 0.08em; - text-transform: uppercase; - color: var(--accent); - background: var(--accent-subtle); - border: 1px solid var(--accent-subtle-border); - border-radius: 20px; - padding: 4px 12px; - margin-bottom: 18px; -} -.early-access-hero h1 { font-size: 2.5rem; line-height: 1.1; letter-spacing: -0.04em; margin-bottom: 16px; } -.early-access-sub { font-size: 1rem; color: var(--text-muted); line-height: 1.7; max-width: 560px; } - -.early-access-perks { display: grid; grid-template-columns: 1fr 1fr; gap: 12px; margin: 32px 0 40px; } -.perk-card { - display: flex; - align-items: flex-start; - gap: 14px; - padding: 18px 20px; - border: 1px solid var(--border); - border-radius: 10px; - background: var(--bg-subtle); -} -.perk-icon { - display: flex; - align-items: center; - justify-content: center; - width: 36px; - height: 36px; - border-radius: 8px; - background: var(--accent-subtle); - border: 1px solid var(--accent-subtle-border); - color: var(--accent); - flex-shrink: 0; -} -.perk-title { font-size: 0.875rem; font-weight: 600; color: var(--text-primary); margin-bottom: 4px; letter-spacing: -0.01em; } -.perk-desc { font-size: 0.8125rem; color: var(--text-muted); line-height: 1.55; } - -.early-access-form-section { margin: 8px 0 32px; } -.early-access-form-section h2 { margin-top: 0; } -.early-access-form { margin-top: 20px; max-width: 480px; } -.ea-form-top { margin-bottom: 0; } -.ea-inline-group { display: flex; gap: 8px; } -.ea-inline-group .ea-input { flex: 1; } -.ea-inline-group .ea-submit { white-space: nowrap; flex-shrink: 0; margin-top: 0; width: auto; padding: 10px 20px; } -.ea-field-group { margin-bottom: 16px; } -.ea-label { display: block; font-size: 0.8125rem; font-weight: 600; color: var(--text-secondary); margin-bottom: 7px; } -.ea-input { - width: 100%; - font-family: var(--font); - font-size: 0.9375rem; - padding: 10px 14px; - border: 1px solid var(--border); - border-radius: var(--radius); - background: var(--bg); - color: var(--text-primary); - outline: none; - transition: border-color 0.15s; -} -.ea-input::placeholder { color: var(--text-muted); } -.ea-input:focus { border-color: var(--accent); } -.ea-input-error, .ea-input-error:focus { border-color: #ef4444; } - -.ea-form-footer { margin-top: 10px; display: flex; flex-direction: column; gap: 6px; } -.ea-pricing-note { font-size: 0.75rem; color: var(--text-muted); margin: 0; } -.ea-platform-links { display: flex; align-items: center; gap: 6px; flex-wrap: wrap; } -.ea-platform-label { font-size: 0.75rem; color: var(--text-muted); } -.ea-platform-link { - background: none; - border: none; - padding: 0; - font-family: var(--font); - font-size: 0.75rem; - font-weight: 600; - color: var(--text-muted); - cursor: pointer; - transition: color 0.12s; - text-decoration: underline; - text-decoration-color: transparent; -} -.ea-platform-link:hover { color: var(--text-primary); } -.ea-platform-link.active { color: var(--accent); text-decoration-color: var(--accent-subtle-border); } - -.ea-submit { - width: 100%; - font-family: var(--font); - font-size: 0.9375rem; - font-weight: 600; - padding: 12px 20px; - background: var(--accent); - color: #fff; - border: none; - border-radius: var(--radius); - cursor: pointer; - transition: background 0.15s; - letter-spacing: -0.01em; - margin-top: 4px; -} -.ea-submit:hover { background: var(--accent-hover); } -.ea-submit:disabled { background: var(--text-muted); cursor: default; } - -.ea-status { font-size: 0.8125rem; margin-top: 10px; min-height: 1.2em; } -.ea-slack-direct { font-size: 0.8125rem; color: var(--text-muted); margin: 12px 0 0; } -.ea-slack-direct a { color: var(--accent); text-decoration: none; } -.ea-slack-direct a:hover { text-decoration: underline; } -.ea-status-success { color: var(--accent); } -.ea-status-error { color: #ef4444; } - -/* ─── Essay early access CTA banner ────────────────────────────────────────── */ -.essay-early-access { - display: flex; - align-items: center; - justify-content: space-between; - gap: 16px; - margin-top: 48px; - padding: 18px 22px; - background: var(--accent-subtle); - border: 1px solid var(--accent-subtle-border); - border-radius: 10px; -} -.essay-ea-text { - font-size: 0.9rem; - color: var(--text-secondary); - line-height: 1.5; -} -.essay-ea-text strong { color: var(--text-primary); } -.essay-ea-btn { - display: inline-flex; - align-items: center; - padding: 9px 18px; - background: var(--accent); - color: #fff; - border-radius: var(--radius); - font-size: 0.8125rem; - font-weight: 600; - text-decoration: none; - white-space: nowrap; - flex-shrink: 0; - transition: background 0.15s; -} -.essay-ea-btn:hover { background: var(--accent-hover); color: #fff; } - -/* ─── Early access essay link cards ────────────────────────────────────────── */ -.ea-essay-links { - display: grid; - grid-template-columns: 1fr 1fr; - gap: 10px; - margin: 16px 0 8px; -} -.ea-essay-card { - display: flex; - flex-direction: column; - gap: 4px; - padding: 14px 16px; - border: 1px solid var(--border); - border-radius: 10px; - background: var(--bg-subtle); - text-decoration: none; - transition: border-color 0.15s, background 0.15s; -} -.ea-essay-card:hover { - border-color: var(--accent); - background: var(--accent-subtle); - text-decoration: none; -} -.ea-essay-title { - font-size: 0.875rem; - font-weight: 600; - color: var(--text-primary); - line-height: 1.35; - letter-spacing: -0.01em; -} -.ea-essay-desc { - font-size: 0.8rem; - color: var(--text-muted); - line-height: 1.5; -} - -/* ─── Mobile ────────────────────────────────────────────────────────────────── */ -.mobile-topbar { - display: none; - position: sticky; - top: 0; - z-index: 100; - background: var(--bg); - border-bottom: 1px solid var(--border); - padding: 11px 16px; - align-items: center; - justify-content: space-between; -} -.mobile-topbar-logo { display: flex; align-items: center; gap: 8px; font-size: 0.9375rem; font-weight: 600; color: var(--text-primary); text-decoration: none; letter-spacing: -0.01em; } -.mobile-topbar-logo img { border-radius: 6px; } -.mobile-menu-btn { background: none; border: none; cursor: pointer; color: var(--text-secondary); padding: 4px; } -.mobile-search-btn { background: none; border: none; cursor: pointer; color: var(--text-secondary); padding: 4px; } -.mobile-overlay { display: none; position: fixed; inset: 0; background: rgba(0,0,0,0.45); z-index: 99; } -.mobile-overlay.visible { display: block; } - -@media (max-width: 768px) { - .mobile-topbar { display: flex; } - .layout { display: block; } - .sidebar { position: fixed; left: -280px; top: 0; height: 100vh; width: 280px; z-index: 100; transition: left 0.25s ease; box-shadow: none; } - .sidebar.open { left: 0; box-shadow: 4px 0 24px rgba(0,0,0,0.18); } - .main { padding: 28px 20px 60px; max-width: 100%; } - .content h1 { font-size: 1.5rem; } - .hero-buttons { gap: 8px; } - .btn { padding: 10px 14px; font-size: 0.8125rem; } - .early-access-perks { grid-template-columns: 1fr; } - .early-access-hero h1 { font-size: 1.875rem; } - .ea-essay-links { grid-template-columns: 1fr; } - .essay-early-access { flex-direction: column; align-items: flex-start; } -} diff --git a/website/assets/favicon-16x16.png b/website/assets/favicon-16x16.png deleted file mode 100644 index 7ccde733..00000000 Binary files a/website/assets/favicon-16x16.png and /dev/null differ diff --git a/website/assets/favicon-32x32.png b/website/assets/favicon-32x32.png deleted file mode 100644 index e967883a..00000000 Binary files a/website/assets/favicon-32x32.png and /dev/null differ diff --git a/website/assets/favicon.ico b/website/assets/favicon.ico deleted file mode 100644 index b5edd178..00000000 Binary files a/website/assets/favicon.ico and /dev/null differ diff --git a/website/assets/js/revenuecat-link.js b/website/assets/js/revenuecat-link.js deleted file mode 100644 index 1d828439..00000000 --- a/website/assets/js/revenuecat-link.js +++ /dev/null @@ -1,53 +0,0 @@ -// Builds a RevenueCat Web Purchase Link from a customer email. -// -// RevenueCat identifies the customer by the App User ID, which is a *path* -// segment (not a query parameter). We use the email itself as the App User ID -// so the purchase is tied to a stable id we can later match to the app account. -// The `email` query parameter only prefills the email field on the checkout -// page (subscribers cannot override it). -// -// https://pay.rev.cat//?email= -// -// This file is loaded as a browser global by /pay/ and imported directly by the -// unit test, so it has no dependencies and works in both environments. -(function (root, factory) { - if (typeof module === 'object' && module.exports) { - module.exports = factory(); - } else { - root.RevenueCatLink = factory(); - } -})(typeof self !== 'undefined' ? self : this, function () { - // Linear-time email check (no overlapping quantifiers, so no catastrophic - // backtracking / ReDoS): local part, '@', then dot-separated domain labels - // that themselves contain no dots. - var EMAIL_RE = /^[^\s@]+@[^\s.@]+(\.[^\s.@]+)+$/; - - function isValidEmail(email) { - return typeof email === 'string' && EMAIL_RE.test(email.trim()); - } - - // Returns the purchase URL, or null if the inputs are invalid. - // opts may include { packageId } to pre-select a package. - function buildPurchaseUrl(token, email, opts) { - if (!token || typeof token !== 'string') return null; - if (!isValidEmail(email)) return null; - - var trimmed = email.trim(); - var encoded = encodeURIComponent(trimmed); - // String concatenation (not template literals) keeps this file ES5-only. - /* eslint-disable prefer-template */ - var url = - 'https://pay.rev.cat/' + encodeURIComponent(token) + '/' + encoded + '?email=' + encoded; - - if (opts && opts.packageId) { - url += '&package_id=' + encodeURIComponent(opts.packageId); - } - /* eslint-enable prefer-template */ - return url; - } - - return { - isValidEmail: isValidEmail, - buildPurchaseUrl: buildPurchaseUrl, - }; -}); diff --git a/website/assets/logo.png b/website/assets/logo.png deleted file mode 100644 index 14b8bc20..00000000 Binary files a/website/assets/logo.png and /dev/null differ diff --git a/website/ethos.md b/website/ethos.md deleted file mode 100644 index 1b928476..00000000 --- a/website/ethos.md +++ /dev/null @@ -1,57 +0,0 @@ ---- -layout: default -title: Ethos -nav_order: 3 -has_children: true -description: Why Off Grid exists. Intelligence should live on the devices you already own - private by architecture, not by policy. ---- - -# Ethos - -Intelligence needs to be democratized. - ---- - -## The problem with AI today - -The most useful AI is the one with your full context. Your messages. Your calendar. Your work, your health, your finances. An AI that actually knows you can reduce real friction from your day. - -But getting that context today means handing it to a server you don't control and paying a subscription for the privilege. Every query leaves your device. Every response comes back from somewhere else. You don't know what's stored, for how long, or what it's used for. - -Privacy by policy ("we promise not to misuse your data") is not the same as privacy by architecture ("the data never left your device"). - ---- - -## What we believe - -The right model of AI is one where the intelligence lives with you, not above you. - -On the devices you already carry. Talking to the apps you already use. Without a single byte making a round trip to someone else's infrastructure. - -This is possible today. The models fit. The hardware is fast enough. The only thing missing was software that took it seriously. - ---- - -## The arc - -AI is the next communication infrastructure. And communication infrastructure, historically, moves toward privacy when users demand it. - -The market will demand it here too. Not because privacy is a talking point, but because people will eventually notice that their most personal context, the things that would make AI useful, is exactly what they're least willing to hand over. - -The devices people already carry will become intelligent. They will speak to each other over local networks. Context will stay on-person. That future doesn't require new hardware or new platforms. It requires software built on the right assumption from the start. - ---- - -## What we're building - -Off Grid is not an autonomous agent that makes decisions on your behalf. It is a private digital secretary that reduces daily friction. - -It reads your messages, watches your calendar, defers your notifications, answers your questions, generates your images, listens to your voice. All of it, on your device. All of it, offline. - -Every knowledge worker should carry their own intelligence layer. Private by architecture. Owned by the person using it. Available anywhere, including places without a signal. - -That's what we're building. - ---- - -*Off Grid is open source. [View on GitHub](https://github.com/alichherawalla/off-grid-mobile?utm_source=offgrid-docs&utm_medium=website&utm_campaign=github) or [join the community on Slack](https://join.slack.com/t/off-grid-mobile/shared_invite/zt-3w2utgk0w-EDiDZBq6KmSZZwEw5Tkhnw).* diff --git a/website/guides/android-setup.md b/website/guides/android-setup.md deleted file mode 100644 index 69246545..00000000 --- a/website/guides/android-setup.md +++ /dev/null @@ -1,73 +0,0 @@ ---- -layout: default -title: Android Setup -parent: Guides -nav_order: 3 -description: How to run LLMs locally on your Android phone in 2026 - no cloud, no account, no subscription. Complete setup guide for Off Grid on Android. ---- - -# Android Setup - -Run a local AI model on your Android phone - completely offline, no account, no API key. - ---- - -## Requirements - -- Android 10 or later -- 4GB RAM minimum (6GB+ recommended for larger models) -- At least 3GB free storage -- Internet for the initial model download only - ---- - -## Step 1 - Install Off Grid - -[Download from Google Play](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=download){: .btn .btn-green } - ---- - -## Step 2 - Download a model - -1. Open Off Grid -2. Tap **Models** -3. Choose a model - **Qwen 3.5 0.8B** or **Qwen 3.5 2B** are the best starting points for most Android devices -4. Tap **Download** - ---- - -## Step 3 - Load and chat - -1. Tap **Load** next to your downloaded model -2. The model loads into RAM (5–20 seconds depending on device) -3. Tap **Chat** and start - ---- - -## Android-specific notes - -**Vulkan acceleration** - On supported devices, Off Grid uses Vulkan for GPU inference. This significantly reduces response time compared to CPU-only. Devices with Snapdragon 8 Gen 2 and newer, Dimensity 9000+, and Exynos 2400 support this. - -**Background behaviour** - Android may kill the model process if the app is backgrounded for too long. Keep Off Grid in the foreground during long conversations, or enable "Don't optimise battery" for the app in settings. - -**Storage** - Models are stored in app-private storage. They don't appear in your gallery or Files app, which means they also won't be accidentally deleted by a cleaner app. - ---- - -## Tested devices - -| Device | RAM | Models confirmed working | -|---|---|---| -| Pixel 8 Pro | 12GB | Llama 3.1 8B, Mistral 7B | -| Samsung S24 | 8GB | Llama 3.2 3B, Mistral 7B Q4 | -| Pixel 7 | 8GB | Llama 3.2 3B, Phi-3 Mini | -| OnePlus 12 | 12GB | Llama 3.1 8B | -| Samsung A55 | 8GB | Phi-3 Mini, Gemma 2B | - ---- - -## Related guides - -- [Which model should I use?]({{ '/guides/which-model' | relative_url }}) -- [Run Stable Diffusion on Android]({{ '/guides/stable-diffusion-android' | relative_url }}) -- [Connect Ollama from your phone]({{ '/guides/ollama-android' | relative_url }}) diff --git a/website/guides/document-analysis.md b/website/guides/document-analysis.md deleted file mode 100644 index 53e05bed..00000000 --- a/website/guides/document-analysis.md +++ /dev/null @@ -1,93 +0,0 @@ ---- -layout: default -title: Document Analysis and Attachments -parent: Guides -nav_order: 14 -description: Attach PDFs, code files, CSVs, and other documents to your Off Grid conversations. The app extracts and passes content to your local model for analysis - entirely on-device. -faq: - - q: What file types can I attach? - a: PDF, txt, md, most code file types (py, js, ts, java, swift, kt, go, rs, sql, sh, etc.), CSV, JSON, YAML, XML, HTML, and more. Maximum 5MB per file. - - q: Does attaching a document send it to the cloud? - a: No. The app extracts text from the document on-device and passes it to your local model. Nothing is uploaded. - - q: Is there a file size limit? - a: 5MB per file. Text content is truncated to 50,000 characters for context window management. ---- - -# Document Analysis and Attachments - -Attach files directly to your conversations and ask your local model questions about them. PDFs, code, CSV data, config files - anything text-based works. - -All processing happens on your device. - ---- - -## Supported formats - -**Documents:** -- PDF (text extracted natively via PDFKit on iOS, PdfRenderer on Android) - -**Text and code:** -- `.txt`, `.md`, `.log` -- `.py`, `.js`, `.ts`, `.jsx`, `.tsx`, `.java`, `.c`, `.cpp`, `.h`, `.swift`, `.kt`, `.go`, `.rs`, `.rb`, `.php`, `.sql`, `.sh` - -**Data files:** -- `.csv`, `.json`, `.xml`, `.yaml`, `.yml`, `.toml`, `.ini`, `.cfg`, `.conf`, `.html` - -**Limits:** 5MB per file. Text is truncated at 50,000 characters for context window management. - ---- - -## How to attach a file - -1. Open a chat in Off Grid -2. Tap the **attachment icon** in the message bar -3. Select **Document** from the picker -4. Choose your file from the system file browser - -The file is copied to app storage (so it survives temp cleanup), and the extracted text is attached to your next message. - ---- - -## Tapping to view - -Tap any document badge in the chat to open it with the system viewer - QuickLook on iOS, the system intent viewer on Android. - ---- - -## Paste as attachment - -If you paste a large block of text into the message field, Off Grid offers to convert it to an attachment instead. This keeps the chat interface clean when you're passing in large context. - ---- - -## What you can do - -**Code review:** -> Attach a file → "Find potential bugs in this code" - -**PDF analysis:** -> Attach a contract → "Summarise the key terms and flag anything unusual" - -**Data analysis:** -> Attach a CSV → "What are the top 5 items by revenue?" - -**Config explanation:** -> Attach a YAML/TOML file → "Explain what this configuration does" - ---- - -## Difference vs knowledge base - -Document attachments are **per-conversation** - you attach something to a specific message and the model sees it in that context window. They're not indexed or searchable. - -The [Knowledge Base]({{ '/guides/knowledge-base' | relative_url }}) is **project-wide** - documents are embedded and indexed, and the model can retrieve relevant chunks from them automatically across many conversations. - -Use attachments for one-off analysis. Use the knowledge base for documents you want to reference repeatedly. - ---- - -## Related guides - -- [Knowledge Base and RAG]({{ '/guides/knowledge-base' | relative_url }}) -- [Vision AI - Analyse Images On-Device]({{ '/guides/vision-ai' | relative_url }}) -- [Tool Calling]({{ '/guides/tool-calling' | relative_url }}) diff --git a/website/guides/index.md b/website/guides/index.md deleted file mode 100644 index d3790f76..00000000 --- a/website/guides/index.md +++ /dev/null @@ -1,113 +0,0 @@ ---- -layout: default -title: Guides -nav_order: 5 -has_children: true -description: Step-by-step guides for running AI locally on your iPhone and Android phone with Off Grid. ---- - -# Guides - -Everything you need to get the most out of running AI locally on your phone. - ---- - -## Getting started - -
- -
Which Model Should I Use?
-
Pick the right model for your device RAM and use case. Full catalogue with performance numbers.
- - -
iOS Setup
-
Download, install, and run your first model on iPhone. Metal GPU acceleration, supported devices.
- - -
Android Setup
-
Get up and running on Android. Vulkan acceleration, background behaviour, tested devices.
- -
- ---- - -## Running LLMs locally - -
- -
Run LLMs on Your iPhone in 2026
-
Qwen 3.5, Gemma 4, Phi-4 Mini running locally on iPhone via llama.cpp and Metal. Real tok/s numbers.
- - -
Run LLMs on Your Android Phone in 2026
-
Local LLMs on Android with CPU and Vulkan GPU acceleration. Device performance table.
- -
- ---- - -## Image generation - -
- -
Stable Diffusion on iPhone
-
On-device image generation using Core ML and the Apple Neural Engine. SD 1.5, 2.1, SDXL.
- - -
Stable Diffusion on Android
-
MNN (CPU, all devices) and QNN NPU (Snapdragon 8 Gen 1+). 5-10s images on flagship chips.
- -
- ---- - -## Vision, voice and documents - -
- -
Vision AI
-
Point your camera at anything and ask questions. SmolVLM, Qwen3-VL, Gemma 4 - all on-device.
- - -
Voice Input with Whisper
-
On-device speech-to-text via whisper.cpp. Hold to record, auto-transcribe. 99 languages, no audio leaves your phone.
- - -
Document Analysis
-
Attach PDFs, CSVs, and code files directly to your chat. Native PDF extraction on iOS and Android.
- - -
Knowledge Base and RAG
-
Upload documents to a project. Off Grid embeds and indexes them on-device, retrieves context automatically.
- -
- ---- - -## Tools and intelligence - -
- -
Tool Calling
-
Web search, calculator, date/time, device info, knowledge base search. Automatic tool loop with runaway prevention.
- -
- ---- - -## Remote servers - -
- -
Remote Servers
-
Connect to Ollama, LM Studio, LocalAI, or vLLM on your home network. Access larger models from your phone over WiFi.
- - -
Ollama from Android
-
Run Llama 3.1 70B on your desktop, control it from your Android phone over WiFi.
- - -
LM Studio from Android
-
Use LM Studio's local server from your phone. Port 1234, network access enabled.
- -
diff --git a/website/guides/ios-setup.md b/website/guides/ios-setup.md deleted file mode 100644 index 4db7304a..00000000 --- a/website/guides/ios-setup.md +++ /dev/null @@ -1,72 +0,0 @@ ---- -layout: default -title: iOS Setup -parent: Guides -nav_order: 2 -description: How to run LLMs locally on your iPhone in 2026 - no cloud, no account, no subscription. Step-by-step setup guide for Off Grid on iOS. ---- - -# iOS Setup - -Run a local AI model on your iPhone with no cloud dependency. This guide covers everything from download to first inference. - ---- - -## Requirements - -- iPhone 12 or newer (A14 Bionic chip or later) -- iOS 16 or later -- At least 3GB free storage (for the app + one model) -- Internet connection for the initial model download only - ---- - -## Step 1 - Install Off Grid - -[Download from the App Store](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=download){: .btn .btn-green } - -The app itself is under 50MB. Models are downloaded separately inside the app. - ---- - -## Step 2 - Download a model - -1. Open Off Grid -2. Tap **Models** in the tab bar -3. Select a model - if you're starting out, pick **Qwen 3.5 2B** (~1.5GB) -4. Tap **Download** - -The download goes to your device. This is the only step that requires internet. - ---- - -## Step 3 - Load and chat - -1. Tap **Load** next to your downloaded model -2. Wait 5–15 seconds for it to load into memory -3. Tap **Chat** and start talking - -You're now running AI entirely on your iPhone. - ---- - -## Tips for better performance - -**Use Metal acceleration** - Off Grid automatically uses Apple's Metal GPU for inference. This makes models 3–5x faster than CPU-only. - -**Close background apps** - iOS may reclaim RAM from background apps. If the model unloads unexpectedly, close other apps and reload. - -**Quantisation matters** - For 4GB RAM devices (iPhone 12/13), stick to Q4 models. For 8GB+ (iPhone 15 Pro+), you can use Q5 or Q8 for slightly better quality. - ---- - -## Offline use - -Once a model is downloaded, Off Grid works in airplane mode. Put your phone offline and it continues to work normally. - ---- - -## Related guides - -- [Which model should I use?]({{ '/guides/which-model' | relative_url }}) -- [Connecting Ollama from your phone]({{ '/guides/ollama-android' | relative_url }}) diff --git a/website/guides/knowledge-base.md b/website/guides/knowledge-base.md deleted file mode 100644 index 7c40540f..00000000 --- a/website/guides/knowledge-base.md +++ /dev/null @@ -1,92 +0,0 @@ ---- -layout: default -title: Knowledge Base and RAG - On-Device Document Search -parent: Guides -nav_order: 13 -description: Upload PDFs and documents to Off Grid's project knowledge base. The app embeds and indexes them on-device using MiniLM, then retrieves relevant context automatically during your conversations. -faq: - - q: Does the knowledge base send my documents to the cloud? - a: No. Documents are processed entirely on-device. Text extraction, embedding, and retrieval all happen locally using a bundled MiniLM model stored in SQLite. - - q: What is the embedding model used? - a: all-MiniLM-L6-v2-Q8_0.gguf, bundled with the app (~24MB). It does not need to be downloaded. - - q: How does retrieval work? - a: At query time, your question is embedded with the same MiniLM model. Off Grid scores all document chunks by cosine similarity and passes the top results to your LLM as context via the search_knowledge_base tool. ---- - -# Knowledge Base and RAG - On-Device Document Search - -Each Off Grid project can have its own knowledge base. Upload PDFs, text files, or code - the app processes them entirely on-device and makes them searchable in your conversations. - -This is Retrieval-Augmented Generation (RAG) running completely locally. No document leaves your device. - ---- - -## How it works - -``` -Your document - → Text extraction (PDF or plain text) - → Chunking (paragraph-aware, with sliding-window fallback) - → Embedding (all-MiniLM-L6-v2-Q8_0.gguf, bundled with app) - → Stored in SQLite on-device - -When you ask a question: - → Your question is embedded with the same MiniLM model - → Cosine similarity scored against all chunks - → Top-K most relevant chunks passed to the LLM as context - → LLM answers using your document as a source -``` - -The `search_knowledge_base` tool is automatically injected into any project conversation when the project has documents. Compatible models call it automatically when they need information from your documents. - ---- - -## Setting up a knowledge base - -1. Open Off Grid → **Projects** -2. Create a new project or tap an existing one -3. Tap **Knowledge Base** → **Add Document** -4. Select a PDF or text file from your device - -Off Grid extracts the text and runs it through the embedding pipeline. This takes a few seconds per document depending on length. - ---- - -## Supported document formats - -- **PDF** - native text extraction via platform APIs (PDFKit on iOS, PdfRenderer on Android) -- **Text files** - `.txt`, `.md`, `.log` -- **Code files** - `.py`, `.js`, `.ts`, `.java`, `.swift`, `.kt`, `.go`, `.rs`, `.sql`, `.sh`, and more -- **Data files** - `.csv`, `.json`, `.xml`, `.yaml`, `.toml`, `.html` - ---- - -## Using the knowledge base in conversation - -Once documents are added, compatible models will call `search_knowledge_base` automatically when they need to retrieve information. You'll see the tool call inline in the chat. - -You can also trigger it explicitly: - -> "Search my knowledge base for anything about onboarding flow" - -> "Based on the uploaded architecture doc, explain how the download service works" - ---- - -## Embedding model - -**all-MiniLM-L6-v2-Q8_0.gguf** - ships bundled with Off Grid (~24MB). It's always available, no download required, and runs fast enough that embedding a 20-page PDF takes under 10 seconds on a modern phone. - ---- - -## Which LLMs support knowledge base search? - -Any model that supports tool calling can use the knowledge base. See the [Tool Calling guide]({{ '/guides/tool-calling' | relative_url }}) for the full list of compatible models. - ---- - -## Related guides - -- [Tool Calling]({{ '/guides/tool-calling' | relative_url }}) -- [Document Analysis and Attachments]({{ '/guides/document-analysis' | relative_url }}) -- [Which Model Should I Use?]({{ '/guides/which-model' | relative_url }}) diff --git a/website/guides/lm-studio-android.md b/website/guides/lm-studio-android.md deleted file mode 100644 index 705c9d96..00000000 --- a/website/guides/lm-studio-android.md +++ /dev/null @@ -1,77 +0,0 @@ ---- -layout: default -title: How to Use LM Studio From Your Android Phone in 2026 -parent: Guides -nav_order: 16 -description: Connect Off Grid on Android to your LM Studio server and access larger models like Llama 3.1 70B over your local WiFi network - no cloud, completely private. -faq: - - q: Can I use LM Studio from my Android phone? - a: Yes. Off Grid connects to LM Studio's local server over your WiFi network. You get access to any model loaded in LM Studio from your Android phone. - - q: Does it require internet? - a: No. The connection is over your local WiFi. No traffic touches the internet. ---- - -# How to Use LM Studio From Your Android Phone in 2026 - -LM Studio runs large models on your Mac or PC with a polished interface. Models too large for your phone - Llama 3.1 70B, DeepSeek, Mistral Large - run on your desktop and stream to your phone over WiFi. - ---- - -## What you need - -- Mac or Windows PC running [LM Studio](https://lmstudio.ai) with a model loaded -- Android phone with [Off Grid](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=download) installed -- Both devices on the same WiFi network - ---- - -## Step 1 - Start LM Studio's local server - -1. Open LM Studio -2. Load a model (click the model dropdown at the top) -3. Go to the **Local Server** tab (left sidebar) -4. Click **Start Server** -5. Enable **"Allow connections from network"** in the server settings -6. Note the displayed port (default: **1234**) - ---- - -## Step 2 - Find your computer's local IP - -**macOS:** System Settings → Network → Wi-Fi → Details → IP address (e.g. `192.168.1.55`) - -**Windows:** Open PowerShell → `ipconfig` → look for IPv4 Address under your Wi-Fi adapter - ---- - -## Step 3 - Connect from Off Grid - -1. Open Off Grid → **Settings** → **Remote Servers** -2. Tap **Add Server** -3. Enter: `http://192.168.1.55:1234` (use your computer's actual IP) -4. Tap **Test Connection** → should show green -5. Tap **Save** - -Off Grid automatically discovers models available on the server. - ---- - -## Step 4 - Select a model and chat - -Open the model picker in Off Grid. Your LM Studio models appear under the server name. Tap one to make it active and start chatting. - -Responses stream in real time via SSE - the same way LM Studio's own interface works. - ---- - -## Using Tailscale for access outside home - -Install [Tailscale](https://tailscale.com) on both your computer and phone. Use your computer's Tailscale IP instead of the local IP. You can now access LM Studio from anywhere - office, travel, anywhere with a data connection. - ---- - -## Related guides - -- [Remote Servers - Connect Ollama, LM Studio, and LocalAI]({{ '/guides/remote-servers' | relative_url }}) -- [How to Use Ollama From Your Android Phone in 2026]({{ '/guides/ollama-android' | relative_url }}) -- [How to Run LLMs Locally on Your Android Phone in 2026]({{ '/guides/run-llms-locally-android' | relative_url }}) diff --git a/website/guides/ollama-android.md b/website/guides/ollama-android.md deleted file mode 100644 index fbe76cbb..00000000 --- a/website/guides/ollama-android.md +++ /dev/null @@ -1,95 +0,0 @@ ---- -layout: default -title: How to Use Ollama From Your Android Phone in 2026 -parent: Guides -nav_order: 8 -description: Connect your Android phone to your home Ollama server and use larger models like Llama 3.1 70B over your local network - no cloud, completely private. -faq: - - q: Can I use Ollama from my Android phone? - a: Yes. Off Grid can connect to any Ollama server on your local network or accessible via VPN. You get access to any model loaded on your desktop from your phone. - - q: Does connecting to Ollama require internet? - a: No. Off Grid connects to Ollama over your local WiFi network. No traffic goes to the internet. ---- - -# How to Use Ollama From Your Android Phone in 2026 - -Ollama lets you run large language models on your desktop. Models that are too big for your phone - Llama 3.1 70B, Mistral Large, CodeLlama 34B - can run on your desktop and be accessed from your phone over your home network. - ---- - -## What you need - -- Desktop or laptop running [Ollama](https://ollama.ai) with at least one model loaded -- Android phone with [Off Grid](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=download) installed -- Both devices on the same WiFi network (or Ollama accessible via VPN/Tailscale) - ---- - -## Step 1 - Configure Ollama to accept remote connections - -By default Ollama only listens on localhost. To accept connections from your phone: - -**macOS / Linux:** -```bash -OLLAMA_HOST=0.0.0.0 ollama serve -``` - -Or set it as a permanent environment variable: -```bash -# ~/.zshrc or ~/.bashrc -export OLLAMA_HOST=0.0.0.0 -``` - -**Windows:** Set `OLLAMA_HOST=0.0.0.0` as a system environment variable and restart Ollama. - ---- - -## Step 2 - Find your desktop's local IP - -**macOS:** System Settings → Network → your WiFi connection → IP address (e.g. `192.168.1.42`) - -**Windows:** `ipconfig` in terminal → IPv4 address under your WiFi adapter - -**Linux:** `ip addr show` - look for your WiFi interface - ---- - -## Step 3 - Connect from Off Grid - -1. Open Off Grid → **Settings** → **Remote Servers** -2. Tap **Add Server** -3. Enter: `http://192.168.1.42:11434` (replace with your desktop's IP) -4. Tap **Test Connection** - it should show green -5. Tap **Save** - ---- - -## Step 4 - Select a model and chat - -1. Open the model picker -2. You'll see models loaded on your Ollama server listed under **Remote** -3. Select one and start chatting - -Your queries go from your phone → your desktop → back to your phone. Nothing touches the internet. - ---- - -## Using Tailscale for access outside your home - -If you want to use Ollama from your phone while away from home, [Tailscale](https://tailscale.com) creates a private VPN between your devices. Install it on both your desktop and phone, then use the Tailscale IP of your desktop instead of the local one. - ---- - -## FAQ - -**Can I use Ollama from my phone without internet?** -Yes - over local WiFi only. For remote access you need Tailscale or a similar VPN. - -**Which Ollama models work best from a phone?** -Any model loaded on your desktop works. `llama3.1:70b` and `mistral-large` are popular choices since they're too large to run locally on a phone. - ---- - -## Related guides - -- [How to Run LLMs Locally on Your Android Phone in 2026]({{ '/guides/run-llms-locally-android' | relative_url }}) diff --git a/website/guides/remote-servers.md b/website/guides/remote-servers.md deleted file mode 100644 index b1b25173..00000000 --- a/website/guides/remote-servers.md +++ /dev/null @@ -1,127 +0,0 @@ ---- -layout: default -title: Remote Servers - Connect Ollama, LM Studio, and LocalAI -parent: Guides -nav_order: 9 -description: Connect Off Grid to any OpenAI-compatible server on your local network - Ollama, LM Studio, LocalAI, vLLM. Access larger models from your desktop via your phone over WiFi. -faq: - - q: Which remote servers does Off Grid support? - a: Any OpenAI-compatible server - Ollama, LM Studio, LocalAI, vLLM, and others. If it exposes a /v1/chat/completions endpoint, it works. - - q: Does connecting to a remote server require internet? - a: No. Off Grid connects over your local WiFi network. No traffic goes to the internet. For access outside your home, use Tailscale. - - q: Where are API keys stored? - a: In your device's system keychain via react-native-keychain. Never in plain storage. ---- - -# Remote Servers - Connect Ollama, LM Studio, and LocalAI - -Your phone can run impressive models locally, but your desktop or Mac can run much larger ones - Llama 3.1 70B, Mistral Large, DeepSeek, CodeLlama 34B. - -Off Grid connects to any OpenAI-compatible server on your local network, giving you access to those models from your phone over WiFi. No internet required. - ---- - -## Supported servers - -| Server | Platform | Notes | -|---|---|---| -| **Ollama** | macOS, Linux, Windows | Most popular, easiest setup | -| **LM Studio** | macOS, Windows | Great UI, easy model management | -| **LocalAI** | Linux, Docker | Self-hosted, many model formats | -| **vLLM** | Linux | High-throughput, GPU-focused | -| **Any OpenAI-compatible** | Any | Needs `/v1/chat/completions` and `/v1/models` | - ---- - -## Setting up Ollama - -**1. Install Ollama on your desktop:** -```bash -# macOS -brew install ollama - -# Linux -curl -fsSL https://ollama.ai/install.sh | sh -``` - -**2. Allow remote connections** (Ollama only listens on localhost by default): -```bash -# macOS/Linux - run Ollama with remote access -OLLAMA_HOST=0.0.0.0 ollama serve - -# Or set permanently in ~/.zshrc / ~/.bashrc -export OLLAMA_HOST=0.0.0.0 -``` - -**3. Pull a model:** -```bash -ollama pull llama3.1:8b -ollama pull qwen2.5:14b -``` - -**4. Find your desktop's local IP:** -- macOS: System Settings → Network → Wi-Fi → Details → IP address -- Linux: `ip addr show` - look for your WiFi interface - ---- - -## Setting up LM Studio - -1. Download and install [LM Studio](https://lmstudio.ai) -2. Download a model in the app -3. Go to **Local Server** tab → click **Start Server** -4. Enable **"Allow connections from network"** in server settings -5. Note the IP and port shown (default port: 1234) - ---- - -## Connecting from Off Grid - -1. Open Off Grid → **Settings** → **Remote Servers** -2. Tap **Add Server** -3. Enter the server URL: - - Ollama: `http://192.168.1.42:11434` - - LM Studio: `http://192.168.1.42:1234` -4. Add an API key if your server requires one (stored in system keychain) -5. Tap **Test Connection** → should show green -6. Tap **Save** - -Off Grid will automatically discover all models available on the server via `/v1/models`. - ---- - -## Selecting a remote model - -Open the model picker. Remote models appear under your server name. Tap one to make it active. - -Off Grid streams responses via Server-Sent Events (SSE) in real time. Switching back to a local model is instant. - ---- - -## Vision and tool calling over remote servers - -Off Grid detects vision and tool calling support from model name patterns. If the model name includes `vision`, `vl`, `vlm`, or similar, Off Grid enables the camera attachment. Tool calling is similarly detected. - -For servers that support it (Ollama with compatible models, LM Studio), tool calling and vision both work without friction over the remote connection. - ---- - -## Access from outside your home with Tailscale - -[Tailscale](https://tailscale.com) creates a private VPN between your devices. Install it on both your desktop and phone, then use the Tailscale IP of your desktop as the server URL. - -This gives you access to your home desktop's models from anywhere - coffee shop, travel, office - without exposing anything to the public internet. - ---- - -## Security note - -Off Grid warns you before connecting to a public internet endpoint (non-private IP range). For remote access, always use Tailscale or a similar private tunnel rather than exposing your server directly to the internet. - ---- - -## Related guides - -- [How to Use Ollama From Your Android Phone in 2026]({{ '/guides/ollama-android' | relative_url }}) -- [Which Model Should I Use?]({{ '/guides/which-model' | relative_url }}) -- [Tool Calling]({{ '/guides/tool-calling' | relative_url }}) diff --git a/website/guides/run-llms-locally-android.md b/website/guides/run-llms-locally-android.md deleted file mode 100644 index 8f446530..00000000 --- a/website/guides/run-llms-locally-android.md +++ /dev/null @@ -1,113 +0,0 @@ ---- -layout: default -title: How to Run LLMs Locally on Your Android Phone in 2026 (No Cloud, No Account) -parent: Guides -nav_order: 4 -description: Run Qwen 3.5, Gemma 4, Mistral and other large language models directly on your Android phone with no internet, no API key, and no subscription. Complete guide for 2026. -faq: - - q: Can I run LLMs on Android without an internet connection? - a: Yes. Once the model is downloaded, Off Grid runs entirely offline. No internet, no server calls, no cloud. - - q: Do I need an account to run LLMs locally on Android? - a: No. Off Grid requires no account, no login, and no API key. Download the app and a model and you're done. - - q: What Android phones can run LLMs locally in 2026? - a: Any Android phone with 4GB RAM running Android 10 or later can run Qwen 3.5 2B. For larger models like Qwen 3.5 9B you need 8GB RAM - flagship devices like the Pixel 8 Pro, Samsung S24, or OnePlus 12. - - q: Which LLM runs best on Android in 2026? - a: For 4GB RAM devices, Qwen 3.5 2B (Q4_K_M). For 8GB+ devices, Qwen 3.5 9B or Gemma 4 E4B. Both support thinking mode for complex tasks. ---- - -# How to Run LLMs Locally on Your Android Phone in 2026 (No Cloud, No Account) - -Every time you ask ChatGPT a question, it's logged on a server. Your query, the response, the time, your account. It's stored indefinitely. That data is used to improve models, inform advertising, comply with law enforcement requests. - -Off Grid removes that entire layer. The model runs in your phone's RAM via llama.cpp on ARM64. Nothing is sent anywhere. - -Here's how to set it up. - ---- - -## What you need - -- Android phone with 4GB RAM or more (Android 10+) -- 2–5GB free storage depending on the model you choose -- Internet once for the initial download - then never again - ---- - -## Step 1 - Download Off Grid - -[Get Off Grid on Google Play](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=download){: .btn .btn-green } - ---- - -## Step 2 - Choose a model - -All models use Q4_K_M quantisation by default - the best balance of quality and size for mobile. - -| Model | Min RAM | Size | Best for | -|---|---|---|---| -| **Qwen 3.5 0.8B** | 3GB | ~0.8GB | Ultra-fast, 262K context, budget devices | -| **Qwen 3.5 2B** | 4GB | ~1.7GB | Best for 4–6GB RAM devices, 262K context | -| **Gemma 4 E2B** | 4GB | ~1.5GB | Vision + thinking mode, MoE architecture | -| **Mistral 7B** | 6GB | ~4.1GB | Fast, reliable general purpose | -| **Gemma 4 E4B** | 6GB | ~2.5GB | Strong reasoning + vision, thinking mode | -| **Qwen 3.5 9B** | 8GB | ~5.5GB | Best on-device quality overall | - -Start with **Qwen 3.5 2B** on a 4–6GB device. Start with **Qwen 3.5 9B** if you have 8GB+ RAM. - ---- - -## Step 3 - Download and load - -1. Open Off Grid → tap **Models** -2. Select your model → tap **Download** -3. Once downloaded, tap **Load** -4. Open **Chat** and start - -The model runs entirely on your device from this point. No network requests. - ---- - -## Step 4 - Go offline - -Turn on airplane mode. Open a chat. It still works. - -This is the point. You now have a capable AI assistant that works without any network connection, on any network, in any country, with no monthly bill. - ---- - -## Performance by device - -Off Grid uses llama.cpp on ARM64 with NEON, i8mm, and dotprod SIMD instructions. Optional OpenCL GPU offloading is available on Qualcomm Adreno GPUs. - -| Device | RAM | Recommended model | Approx tok/s | -|---|---|---|---| -| Pixel 9 Pro | 16GB | Qwen 3.5 9B | 15–25 | -| Samsung Galaxy S25 | 12GB | Qwen 3.5 9B | 15–25 | -| Pixel 8 Pro | 12GB | Qwen 3.5 9B | 12–20 | -| Samsung S24 | 8GB | Qwen 3.5 9B or Gemma 4 E4B | 10–18 | -| Pixel 7 | 8GB | Qwen 3.5 9B | 8–15 | -| OnePlus 12 | 12GB | Qwen 3.5 9B | 12–20 | -| Samsung A55 | 8GB | Qwen 3.5 2B | 15–25 | -| Budget 4GB device | 4GB | Qwen 3.5 0.8B | 20–35 | - ---- - -## Why run LLMs locally instead of using the cloud? - -**Privacy.** Your queries never leave your device. - -**No cost.** No API fees, no subscription. The model is free to download and runs forever. - -**Offline.** Works on planes, in areas with bad signal, in countries where cloud AI services are restricted. - -**Speed.** For short queries, local inference on modern ARM chips is surprisingly fast - often faster than waiting for a cloud response on a slow connection. - ---- - -## Related guides - -- [How to Run LLMs Locally on Your iPhone in 2026]({{ '/guides/run-llms-locally-iphone' | relative_url }}) -- [Which model should I use?]({{ '/guides/which-model' | relative_url }}) -- [How to Run Stable Diffusion on Your Android Phone]({{ '/guides/stable-diffusion-android' | relative_url }}) -- [How to Use Ollama From Your Android Phone in 2026]({{ '/guides/ollama-android' | relative_url }}) -- [Vision AI - Analyse Images On-Device]({{ '/guides/vision-ai' | relative_url }}) diff --git a/website/guides/run-llms-locally-iphone.md b/website/guides/run-llms-locally-iphone.md deleted file mode 100644 index 7369ecef..00000000 --- a/website/guides/run-llms-locally-iphone.md +++ /dev/null @@ -1,96 +0,0 @@ ---- -layout: default -title: How to Run LLMs Locally on Your iPhone in 2026 (Completely Offline, No Subscription) -parent: Guides -nav_order: 5 -description: Run Qwen 3.5, Gemma 4, Mistral and other large language models directly on your iPhone with no internet connection and no subscription fee. Step-by-step guide for 2026. -faq: - - q: Can I run LLMs on iPhone without internet? - a: Yes. After the one-time model download, Off Grid runs fully offline using Apple's Metal GPU and Neural Engine. No internet required. - - q: Which iPhones can run LLMs locally in 2026? - a: iPhone 12 or newer (A14 chip or later). Smaller models like Qwen 3.5 0.8B and Qwen 3.5 2B run on any supported iPhone. Larger models like Qwen 3.5 9B need iPhone 15 Pro or newer with 8GB RAM. - - q: Is running LLMs on iPhone as good as ChatGPT? - a: For everyday tasks - summarisation, Q&A, writing help - Qwen 3.5 9B on iPhone 15 Pro handles most things you'd reach for ChatGPT for. Larger cloud models still have an edge on complex multi-step reasoning, but the gap is narrower than most people expect. ---- - -# How to Run LLMs Locally on Your iPhone in 2026 (Completely Offline, No Subscription) - -Apple's Metal GPU and Neural Engine exist in every iPhone since 2017. They're dedicated AI accelerators, sitting mostly idle while you pay a monthly subscription to send queries to someone else's server. - -Off Grid changes that. Run Qwen 3.5, Gemma 4, Mistral, and other leading models directly on your iPhone - offline, private, with no ongoing cost. Inference runs via llama.cpp with Metal GPU acceleration. - ---- - -## Requirements - -- iPhone 12 or newer (A14 Bionic or later) -- iOS 16 or later -- 3GB free storage minimum -- Internet once for the model download - ---- - -## Step 1 - Install Off Grid - -[Download from the App Store](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=download){: .btn .btn-green } - ---- - -## Step 2 - Choose your model - -All models use Q4_K_M quantisation - the best balance of quality and size for mobile. - -| Model | Min iPhone | RAM needed | Size | Best for | -|---|---|---|---|---| -| **Qwen 3.5 0.8B** | iPhone 12 | 3GB | ~0.8GB | Ultra-fast, 262K context | -| **Qwen 3.5 2B** | iPhone 12 | 4GB | ~1.7GB | Best for 4–6GB devices | -| **Gemma 4 E2B** | iPhone 12 | 4GB | ~1.5GB | Vision + thinking mode | -| **Mistral 7B** | iPhone 14 | 6GB | ~4.1GB | Fast, reliable general purpose | -| **Gemma 4 E4B** | iPhone 14 | 6GB | ~2.5GB | Reasoning + vision, thinking mode | -| **Qwen 3.5 9B** | iPhone 15 Pro | 8GB | ~5.5GB | Best on-device quality overall | - -iPhone 12/13 users: start with **Qwen 3.5 2B**. iPhone 15 Pro / 16 users: try **Qwen 3.5 9B**. - ---- - -## Step 3 - Download, load, chat - -1. Open Off Grid → **Models** -2. Tap a model → **Download** -3. Tap **Load** - the model loads via Metal (Apple's GPU framework) -4. Open **Chat** - -You're now running inference locally on Apple Silicon. Nothing leaves your phone. - ---- - -## Why iPhone is great for local AI - -iPhones have a key advantage: **unified memory**. The Metal GPU and CPU share the same memory pool, which means models load faster and inference is more efficient than CPU-only devices. - -Qwen 3.5 2B on an iPhone 14 generates around 20–30 tokens per second. That's fast enough for a fluid conversation. - -Thinking mode (Qwen 3.5, Gemma 4) works particularly well on iPhone because Metal acceleration keeps the longer reasoning sequences from feeling slow. - ---- - -## Performance by device - -| iPhone | RAM | Recommended model | Approx tok/s | -|---|---|---|---| -| iPhone 16 Pro Max | 8GB | Qwen 3.5 9B | 18–28 | -| iPhone 16 / 16 Plus | 8GB | Qwen 3.5 9B | 18–28 | -| iPhone 15 Pro | 8GB | Qwen 3.5 9B | 15–25 | -| iPhone 14 Pro | 6GB | Gemma 4 E4B | 15–22 | -| iPhone 14 | 6GB | Qwen 3.5 2B | 20–30 | -| iPhone 13 | 4GB | Qwen 3.5 2B | 18–26 | -| iPhone 12 | 4GB | Qwen 3.5 0.8B | 25–40 | - ---- - -## Related guides - -- [How to Run LLMs Locally on Your Android Phone in 2026]({{ '/guides/run-llms-locally-android' | relative_url }}) -- [Which model should I use?]({{ '/guides/which-model' | relative_url }}) -- [How to Run Stable Diffusion on Your iPhone]({{ '/guides/stable-diffusion-iphone' | relative_url }}) -- [Vision AI - Analyse Images On-Device]({{ '/guides/vision-ai' | relative_url }}) diff --git a/website/guides/stable-diffusion-android.md b/website/guides/stable-diffusion-android.md deleted file mode 100644 index ddf7525c..00000000 --- a/website/guides/stable-diffusion-android.md +++ /dev/null @@ -1,94 +0,0 @@ ---- -layout: default -title: How to Run Stable Diffusion on Your Android Phone (On-Device AI Image Generation) -parent: Guides -nav_order: 6 -description: Generate AI images locally on your Android phone using Stable Diffusion - no cloud, no API key, no subscription. Complete guide for on-device image generation with Off Grid. -faq: - - q: Can Android phones run Stable Diffusion locally? - a: Yes. All Android phones running Off Grid use the MNN backend (CPU-based, works on all devices). Phones with Snapdragon 8 Gen 1 or newer also get QNN NPU acceleration, which is 2-3x faster. - - q: How long does image generation take on Android? - a: On Snapdragon 8 Gen 2/3 with QNN NPU, 512x512 images take roughly 5-10 seconds at 20 steps. CPU-only (MNN) takes around 15 seconds on the same chip. - - q: Do I need a specific chipset for image generation? - a: No. MNN backend works on all ARM64 Android devices. QNN NPU acceleration requires Snapdragon 8 Gen 1 or newer for the fastest results. ---- - -# How to Run Stable Diffusion on Your Android Phone (On-Device AI Image Generation) - -Every image you generate on Midjourney, DALL-E, or Adobe Firefly is stored on their servers. Your prompts, the images, metadata. It's used for training and stored indefinitely. - -Off Grid runs Stable Diffusion entirely on your phone using Alibaba's MNN framework (CPU) or Qualcomm's QNN engine (NPU). Nothing is uploaded. - ---- - -## Requirements - -- Android phone with 4GB RAM minimum (6GB+ recommended) -- Android 10 or later -- ~1–2GB free storage per model -- Internet once for the model download - ---- - -## Step 1 - Install Off Grid - -[Get Off Grid on Google Play](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=download){: .btn .btn-green } - ---- - -## Step 2 - Download an image model - -1. Open Off Grid → **Models** → switch to the **Image** tab -2. Choose a model based on your chipset: - -**All devices (MNN/CPU):** -- Anything V5 - anime/stylised art -- Absolute Reality - photorealistic -- QteaMix - versatile -- ChilloutMix - portrait-focused -- CuteYukiMix - stylised - -**Snapdragon 8 Gen 1+ (QNN/NPU) - faster:** -- DreamShaper, Realistic Vision, MajicmixRealistic, and 15+ more - -3. Tap **Download** (~1–1.2GB per model) - ---- - -## Step 3 - Generate your first image - -1. Open Off Grid → **Image Generation** -2. Type a prompt: `a mountain valley at sunset, photorealistic, golden hour` -3. Tap **Generate** - -Off Grid automatically detects whether your device supports QNN NPU and uses it if available, falling back to MNN (CPU) otherwise. - ---- - -## Performance - -| Backend | Chipset | Time for 512×512 @ 20 steps | -|---|---|---| -| QNN NPU | Snapdragon 8 Gen 2/3/4 | ~5–10s | -| QNN NPU | Snapdragon 8 Gen 1 | ~10–15s | -| MNN CPU | Any ARM64 | ~15s (Snapdragon 8 Gen 3) | -| MNN CPU | Mid-range | ~25–40s | - ---- - -## Tips for better images - -**Prompt structure** - `[subject], [style], [lighting], [quality descriptors]`. Example: `a red fox in a forest, digital art, golden hour lighting, highly detailed, sharp focus` - -**Use prompt enhancement** - Off Grid can use your loaded text model to automatically expand a short prompt into a detailed one. Enable it in the generation screen. Just type `a fox in a forest` and let the LLM do the rest. - -**Steps** - 20 steps is a good default. 30 gives marginally better quality at the cost of ~50% more time. - -**Negative prompt** - Add `blurry, low quality, distorted, deformed` to suppress common artifacts. - ---- - -## Related guides - -- [How to Run Stable Diffusion on Your iPhone]({{ '/guides/stable-diffusion-iphone' | relative_url }}) -- [How to Run LLMs Locally on Your Android Phone in 2026]({{ '/guides/run-llms-locally-android' | relative_url }}) diff --git a/website/guides/stable-diffusion-iphone.md b/website/guides/stable-diffusion-iphone.md deleted file mode 100644 index a78fb375..00000000 --- a/website/guides/stable-diffusion-iphone.md +++ /dev/null @@ -1,92 +0,0 @@ ---- -layout: default -title: How to Run Stable Diffusion on Your iPhone (On-Device AI Image Generation) -parent: Guides -nav_order: 7 -description: Generate AI images locally on your iPhone using Stable Diffusion and Core ML - no cloud, no API key, no subscription. Complete guide for iOS image generation. -faq: - - q: How does image generation work on iPhone? - a: Off Grid uses Apple's Core ML framework with Neural Engine (ANE) acceleration. The entire pipeline runs on-device - text encoding, UNet denoising, VAE decoding - with no data sent anywhere. - - q: Which iPhones support image generation? - a: iPhone 12 or newer. Palettized models (~1GB) run on any supported iPhone. Full precision models (~4GB) run best on iPhone 14 Pro and newer with more RAM and a faster Neural Engine. - - q: How long does image generation take on iPhone? - a: On A17 Pro (iPhone 15 Pro), 512x512 at 20 steps takes roughly 8-15 seconds with the palettized model. Full precision models are faster on the Neural Engine but use more RAM. ---- - -# How to Run Stable Diffusion on Your iPhone (On-Device AI Image Generation) - -Off Grid uses Apple's Core ML pipeline with Neural Engine (ANE) acceleration to run Stable Diffusion entirely on your iPhone. No GPU server. No upload. No cost per image. - -The pipeline: text prompt → CLIP tokenizer → text encoder → UNet (denoising, DPM-Solver scheduler) → VAE decoder → 512×512 image. All on-device. - ---- - -## Requirements - -- iPhone 12 or newer (A14 Bionic or later) -- iOS 16 or later -- 2GB free storage minimum (palettized models ~1GB, full precision ~4GB) -- Internet once for the model download - ---- - -## Step 1 - Install Off Grid - -[Download from the App Store](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=download){: .btn .btn-green } - ---- - -## Step 2 - Download an image model - -Open Off Grid → **Models** → **Image** tab. Available Core ML models: - -| Model | Size | Best for | -|---|---|---| -| **SD 1.5 Palettized** | ~1GB | Best starting point - runs on all supported iPhones | -| **SD 2.1 Palettized** | ~1GB | Slightly better quality than 1.5 palettized | -| **SDXL iOS** | ~2GB | Higher resolution (768×768), 4-bit mixed-bit palettized | -| **SD 1.5 Full** | ~4GB | Fastest on Neural Engine, best quality, needs 6GB+ RAM | -| **SD 2.1 Base Full** | ~4GB | Best quality overall, needs 6GB+ RAM | - -**Start with SD 1.5 Palettized** - it's ~1GB, runs on any supported iPhone, and delivers solid results. - ---- - -## Step 3 - Generate an image - -1. Open Off Grid → **Image Generation** -2. Enter your prompt: `a misty forest at dawn, cinematic lighting, photorealistic` -3. Tap **Generate** - -You'll see a real-time preview update as the model denoises the image step by step. - ---- - -## Performance - -| iPhone | Model | Time @ 20 steps | -|---|---|---| -| iPhone 15 Pro (A17 Pro) | SD 1.5 Palettized | ~8–12s | -| iPhone 15 Pro (A17 Pro) | SD 1.5 Full | ~8–15s | -| iPhone 14 Pro (A16) | SD 1.5 Palettized | ~10–16s | -| iPhone 13 (A15) | SD 1.5 Palettized | ~14–20s | -| iPhone 12 (A14) | SD 1.5 Palettized | ~18–28s | - -> **Note:** Palettized models (~1GB) use 6-bit quantisation and are slightly slower due to dequantisation overhead. Full precision models (~4GB) are faster on the Neural Engine but require iPhone 14 Pro or newer. - ---- - -## Tips - -**Prompt enhancement** - Off Grid can use your loaded text model to expand a short prompt automatically. Type `a fox in a forest` and let the LLM write the detailed prompt for you. - -**Real-time preview** - Watch the image form step-by-step. You can cancel early if the composition is wrong without waiting for the full generation. - -**Steps** - 20 is the default. Palettized models benefit from 25–30 steps for better detail. DPM-Solver converges faster than older schedulers, so you need fewer steps than you might expect. - ---- - -## Related guides - -- [How to Run Stable Diffusion on Your Android Phone]({{ '/guides/stable-diffusion-android' | relative_url }}) -- [How to Run LLMs Locally on Your iPhone in 2026]({{ '/guides/run-llms-locally-iphone' | relative_url }}) diff --git a/website/guides/tool-calling.md b/website/guides/tool-calling.md deleted file mode 100644 index 43e426e8..00000000 --- a/website/guides/tool-calling.md +++ /dev/null @@ -1,92 +0,0 @@ ---- -layout: default -title: Tool Calling -parent: Guides -nav_order: 10 -description: How to use Off Grid's built-in tools - web search, calculator, date/time, device info, and knowledge base search - with any function-calling model. -faq: - - q: Which models support tool calling in Off Grid? - a: Any model that supports function calling in GGUF format. Qwen 3.5, Gemma 4, Mistral 7B, and Phi-4 Mini all support it. Check the model card - if it lists "function calling" or "tool use", it works. - - q: Does tool calling require internet? - a: The calculator, date/time, and device info tools are fully offline. Web search requires an internet connection. Knowledge base search is fully local. ---- - -# Tool Calling - -Off Grid ships with built-in tools that compatible models can call automatically during a conversation. The model decides when to use them - you don't need to trigger them manually. - ---- - -## Available tools - -| Tool | What it does | Requires internet | -|---|---|---| -| **Web search** | Searches the web and returns results with clickable links | Yes | -| **Calculator** | Evaluates mathematical expressions | No | -| **Date / Time** | Returns the current date, time, and timezone | No | -| **Device info** | Returns device name, OS version, available RAM | No | -| **Knowledge base search** | Searches documents you've uploaded to a project | No | - ---- - -## How it works - -When you send a message, the model reads the available tool definitions and decides whether to call one. If it does: - -1. The model emits a function call (e.g. `search("best offline AI apps 2026")`) -2. Off Grid executes the tool and returns the result to the model -3. The model reads the result and generates its final response -4. This loop repeats until the model has enough information - with runaway prevention to avoid infinite loops - -You see the tool calls inline in the conversation as collapsible cards. - ---- - -## Which models support tool calling - -Function calling requires a model trained for it. In Off Grid's recommended catalogue: - -| Model | Tool calling | -|---|---| -| Qwen 3.5 0.8B | Yes | -| Qwen 3.5 2B | Yes | -| Qwen 3.5 9B | Yes | -| Gemma 4 E2B | Yes | -| Gemma 4 E4B | Yes | -| Phi-4 Mini | Yes | -| Mistral 7B | Yes | -| SmolLM3 3B | Limited | -| SmolLM2 360M | No | - -If you're downloading a custom GGUF from Hugging Face, check the model card for "function calling" or "tool use" support. - ---- - -## Using web search - -Web search is automatic - just ask a question that requires current information: - -> "What is the latest version of llama.cpp?" - -The model will call `web_search`, get results, and cite them in its answer with clickable links. - -**Note:** Web search is the only tool that requires an internet connection. All other tools work offline. - ---- - -## Using the knowledge base tool - -The `search_knowledge_base` tool is available automatically in any project that has documents uploaded. See the [Knowledge Base guide]({{ '/guides/knowledge-base' | relative_url }}) for setup. - ---- - -## Disabling tools - -Go to **Chat settings** → toggle off individual tools. You can disable web search to force fully offline responses, or disable all tools if you want pure text generation. - ---- - -## Related guides - -- [Knowledge Base and RAG]({{ '/guides/knowledge-base' | relative_url }}) -- [Which Model Should I Use?]({{ '/guides/which-model' | relative_url }}) diff --git a/website/guides/vision-ai.md b/website/guides/vision-ai.md deleted file mode 100644 index 8b99e4e1..00000000 --- a/website/guides/vision-ai.md +++ /dev/null @@ -1,94 +0,0 @@ ---- -layout: default -title: Vision AI - Analyse Images and Documents On-Device -parent: Guides -nav_order: 11 -description: Use Off Grid's vision models to analyse photos, read documents, describe scenes, and answer questions about images - all on your phone with no cloud. -faq: - - q: Which models support vision in Off Grid? - a: SmolVLM (500M, 2.2B), Qwen3-VL (2B, 8B), and Gemma 4 (E2B, E4B). Gemma 4 models support both vision and thinking mode simultaneously. - - q: Can I use vision AI completely offline? - a: Yes. Vision inference runs entirely on-device using llama.rn multimodal. No image data is sent anywhere. - - q: How long does vision inference take? - a: SmolVLM models take 7-10 seconds on flagship devices. Qwen3-VL and Gemma 4 are slightly slower but significantly more capable. ---- - -# Vision AI - Analyse Images and Documents On-Device - -Off Grid's vision models can look at images and answer questions about them. Point your camera at a document, a product, a diagram, a receipt - and ask anything. - -All inference runs on-device via llama.rn's multimodal support. No image is uploaded anywhere. - ---- - -## What you can do - -- **Read receipts, invoices, business cards** - extract text from photos -- **Describe scenes** - understand what's in a photo -- **Analyse documents** - ask questions about a photo of a document -- **Identify objects** - "what is this?" with a photo -- **Read handwriting** - with capable models like Qwen3-VL -- **Code from screenshots** - show the model a UI and ask it to recreate the code - ---- - -## Available vision models - -| Model | Params | Min RAM | Speed | Best for | -|---|---|---|---|---| -| **SmolVLM2 500M** | 0.5B | 3GB | Very fast (~7s) | Quick visual Q&A on low-RAM devices | -| **SmolVLM 2B** | 2B | 4GB | Fast (~8s) | General vision tasks | -| **SmolVLM2 2.2B** | 2.2B | 4GB | Fast (~8–10s) | Vision + video understanding | -| **Gemma 4 E2B** | 2B (MoE) | 4GB | Medium (~10–15s) | Best vision quality for 4GB, thinking mode | -| **Gemma 4 E4B** | 4B (MoE) | 6GB | Medium (~12–18s) | Strongest reasoning + vision, thinking mode | -| **Qwen3-VL 2B** | 2B | 4GB | Medium | Multilingual vision, thinking mode | - -> **Gemma 4 models** support both vision and thinking mode together - they can reason step-by-step about what they see, which dramatically improves accuracy on complex tasks. - ---- - -## How to use vision - -1. Open a chat in Off Grid -2. Tap the **attachment icon** → choose **Camera** or **Photo Library** -3. Select or capture your image -4. Type your question and send - -The model receives both the image and your question. Vision models automatically download a companion **mmproj file** (multimodal projector) during setup - this is included in the model size estimate. - ---- - -## Example prompts - -**Document analysis:** -> "What are the line items on this receipt? Give me a total." - -**Technical reading:** -> "Explain this architecture diagram." - -**Handwriting:** -> "Transcribe the text in this photo." - -**Visual Q&A:** -> "What model of phone is shown in this photo?" - -**Code from UI:** -> "Write the React Native code to recreate this screen." - ---- - -## Tips - -**Use Gemma 4 for complex reasoning** - If you need the model to think carefully about what it sees (e.g. interpreting a chart, solving a problem from a photo), Gemma 4's thinking mode produces much better results than a faster model. - -**Use SmolVLM for quick tasks** - For simple description or text extraction, SmolVLM2 500M is surprisingly capable and much faster. - -**Image quality matters** - Blurry or low-contrast photos degrade accuracy significantly. For documents, flat lighting and a straight-on angle work best. - ---- - -## Related guides - -- [Which Model Should I Use?]({{ '/guides/which-model' | relative_url }}) -- [Document Analysis and Attachments]({{ '/guides/document-analysis' | relative_url }}) -- [Knowledge Base and RAG]({{ '/guides/knowledge-base' | relative_url }}) diff --git a/website/guides/voice-stt.md b/website/guides/voice-stt.md deleted file mode 100644 index e8ed5b8d..00000000 --- a/website/guides/voice-stt.md +++ /dev/null @@ -1,83 +0,0 @@ ---- -layout: default -title: Voice Input - On-Device Speech-to-Text with Whisper -parent: Guides -nav_order: 12 -description: Use Off Grid's on-device Whisper speech-to-text to dictate messages to your AI. No audio is ever sent to a server. Works offline on both iPhone and Android. -faq: - - q: Does voice transcription require internet? - a: No. Off Grid uses whisper.cpp running entirely on-device. No audio is sent anywhere, ever. - - q: Which Whisper model should I use? - a: Start with Whisper Base - it's the best balance of speed and accuracy for most uses. Whisper Tiny is faster but less accurate. Whisper Small is more accurate but slower. - - q: What languages does Whisper support? - a: Whisper supports 99 languages. It detects the language automatically. ---- - -# Voice Input - On-Device Speech-to-Text with Whisper - -Off Grid uses **whisper.cpp** (via whisper.rn) to transcribe your voice directly on your device. You hold the button, speak, and your words appear as text in the chat input - ready to send or edit. - -No audio is ever sent to a server. The model runs in your phone's memory. - ---- - -## Setup - -Whisper models are downloaded automatically on first use. You don't need to do anything manually - tap the microphone button and Off Grid will prompt you to download a model if one isn't installed. - -You can also select your preferred Whisper model in **Settings → Voice Input**. - ---- - -## Whisper model comparison - -| Model | Size | Speed | Accuracy | Best for | -|---|---|---|---|---| -| **Whisper Tiny** | ~75MB | Fastest | Good | Quick dictation, fast devices | -| **Whisper Base** | ~145MB | Fast | Very good | Best starting point | -| **Whisper Small** | ~465MB | Slower | Excellent | Accents, technical terms, multilingual | - -**Recommended: Whisper Base** for most users. It transcribes in near-real-time on any modern phone with very high accuracy. - ---- - -## How to use it - -1. Open a chat in Off Grid -2. Tap and **hold** the microphone button -3. Speak - you'll see the waveform -4. Release to transcribe - -The transcription appears in the message input field. You can edit it before sending, or send immediately. - -**Slide to cancel** - while holding, slide left to discard the recording without transcribing. - ---- - -## Partial transcription - -Off Grid streams transcription results in real time as you speak. You'll see words appearing as the model processes your audio - you don't have to wait until you stop speaking. - ---- - -## Language support - -Whisper detects your language automatically. It supports 99 languages including English, Spanish, French, German, Japanese, Chinese, Arabic, Hindi, and many more. - -If you're consistently speaking a language other than English and accuracy is low, try **Whisper Small** - it has stronger multilingual performance. - ---- - -## Privacy - -- Audio is buffered temporarily in native code and cleared immediately after transcription -- No audio data is written to disk -- No audio is sent to any server -- The Whisper model runs locally via whisper.cpp - ---- - -## Related guides - -- [Tool Calling]({{ '/guides/tool-calling' | relative_url }}) -- [Quick Start]({{ '/quick-start' | relative_url }}) diff --git a/website/guides/which-model.md b/website/guides/which-model.md deleted file mode 100644 index 8ca1c121..00000000 --- a/website/guides/which-model.md +++ /dev/null @@ -1,114 +0,0 @@ ---- -layout: default -title: Which Model Should I Use? -parent: Guides -nav_order: 1 -description: A practical guide to choosing the right LLM for your iPhone or Android - comparing Qwen 3.5, Gemma 4, Phi-4, Mistral, SmolLM by speed, quality, and RAM requirements. -faq: - - q: What is the best model for a phone with 4GB RAM? - a: Qwen 3.5 2B (Q4_K_M) is the best option for 4GB RAM devices. It supports 262K context, thinking mode, and runs comfortably within memory limits. For vision tasks, Gemma 4 E2B is the recommended choice. - - q: What quantisation does Off Grid use by default? - a: Q4_K_M. It gives the best balance of quality and size for mobile hardware and is the default for all recommended models. - - q: What is the best model for on-device reasoning? - a: Gemma 4 E4B or Qwen 3.5 9B on devices with 6–8GB+ RAM. Both support thinking mode - the model reasons step-by-step before answering, significantly improving accuracy on complex tasks. - - q: Can I use vision models for free? - a: Yes. SmolVLM2 500M works on any phone with 3GB RAM. Gemma 4 E2B gives much better vision quality and needs 4GB RAM. ---- - -# Which Model Should I Use? - -Off Grid uses the actual models in the app - not generic suggestions. All recommendations below are sourced directly from the model catalogue. Default quantisation is **Q4_K_M** for everything. - ---- - -## Quick pick by RAM - -| Your device RAM | Best text model | Best vision model | -|---|---|---| -| 3GB | Qwen 3.5 0.8B | SmolVLM2 500M | -| 4GB | Qwen 3.5 2B | Gemma 4 E2B | -| 6GB | Gemma 4 E4B or Phi-4 Mini | Gemma 4 E4B | -| 8GB+ | Qwen 3.5 9B | Qwen 3.5 9B | - ---- - -## Full model catalogue - -### Text models - -| Model | Params | Min RAM | Context | Best for | -|---|---|---|---|---| -| **SmolLM2 360M** | 0.36B | 3GB | 8K | Ultra-light, low-RAM devices only | -| **Qwen 3.5 0.8B** | 0.8B | 3GB | 262K | Fast responses, long context on budget devices | -| **Qwen 3.5 2B** | 2B | 4GB | 262K | Best general-purpose model for 4GB devices | -| **SmolLM3 3B** | 3B | 6GB | 128K | Purpose-built for constrained devices | -| **Phi-4 Mini** | 3.8B | 6GB | 128K | Reasoning, math, structured tasks | -| **Mistral 7B** | 7B | 6GB | 32K | Fast, reliable general purpose | -| **Qwen 3.5 9B** | 9B | 8GB | 262K | Best on-device quality overall | - -### Vision models (can see images) - -| Model | Params | Min RAM | Best for | -|---|---|---|---| -| **SmolVLM2 500M** | 0.5B | 3GB | Tiny vision model for low-RAM devices | -| **SmolVLM 2B** | 2B | 4GB | General vision tasks on mid-range phones | -| **SmolVLM2 2.2B** | 2.2B | 4GB | Vision + video understanding | -| **Gemma 4 E2B** | 2B (MoE) | 4GB | Best vision quality for 4GB devices, thinking mode | -| **Gemma 4 E4B** | 4B (MoE) | 6GB | Strongest reasoning + vision, thinking mode | - -> **Gemma 4** uses a Mixture-of-Experts (MoE) architecture - the effective parameter count is lower than it looks, which is why it fits in less RAM than you'd expect while delivering quality above its weight class. - ---- - -## What is thinking mode? - -Qwen 3.5 and Gemma 4 models support **thinking mode** - the model reasons through a problem step-by-step before producing its final answer, similar to chain-of-thought prompting but built into the model weights. - -Use it for: complex reasoning, math, multi-step problems. Skip it for: quick Q&A, summarisation, casual chat (it's slower). - ---- - -## Understanding Q4_K_M - -Off Grid defaults to **Q4_K_M** quantisation for all models. This means: - -- ~4.5 bits per weight -- ~5–8% quality loss vs the full-precision original -- ~50–60% smaller than the float16 version -- Recommended by the llama.cpp community as the best mobile tradeoff - -Don't go below Q4_K_S unless you're severely constrained on storage. Q2/Q3 models have noticeable quality degradation. - ---- - -## RAM safety thresholds - -Off Grid automatically checks if a model fits safely before loading: - -- **4GB RAM devices**: model budget = 40% of total RAM -- **6GB+ RAM devices**: model budget = 60% of total RAM -- Text models need ~1.5x their raw size in RAM (KV cache + activations) -- Image models need ~1.5x on iOS (CoreML), ~1.8x on Android (Vulkan) - -If a model is marked as incompatible with your device, this is why. - ---- - -## FAQ - -**What is the best model for 4GB RAM?** -Qwen 3.5 2B (Q4_K_M). For vision tasks, Gemma 4 E2B. - -**What quantisation does Off Grid use?** -Q4_K_M by default - the best balance of quality and size for mobile. - -**What is the best model for reasoning?** -Gemma 4 E4B (6GB RAM) or Qwen 3.5 9B (8GB RAM). Both have thinking mode. - ---- - -## Related guides - -- [How to Run LLMs Locally on Your Android Phone in 2026]({{ '/guides/run-llms-locally-android' | relative_url }}) -- [How to Run LLMs Locally on Your iPhone in 2026]({{ '/guides/run-llms-locally-iphone' | relative_url }}) -- [Vision AI - Analyse Images and Documents]({{ '/guides/vision-ai' | relative_url }}) diff --git a/website/index.md b/website/index.md deleted file mode 100644 index 6f4a84a8..00000000 --- a/website/index.md +++ /dev/null @@ -1,157 +0,0 @@ ---- -layout: default -title: Home -nav_order: 1 -description: Off Grid runs AI on your phone. The model loads into RAM, inference runs on your CPU and GPU, nothing leaves the device. Open source, no account, no cloud. ---- - -Off Grid - Private AI. No cloud. No compromise. - -
- -

Off Grid

-
- -**Your AI assistant. On your phone. Nowhere else.** - -Chat, voice, vision, image generation, tools. The model runs in your phone's RAM. Inference happens on your CPU and GPU. Nothing is sent anywhere. - -
- - - App Store - - - - Google Play - - - - Join Slack - -
- -100K+ downloads. 4.3 stars on iOS. 2,458+ stars on [GitHub](https://github.com/alichherawalla/off-grid-mobile?utm_source=offgrid-docs&utm_medium=website&utm_campaign=github). Open source, built in public. - ---- - -## What it does - -| Capability | What actually happens | -|---|---| -| **Chat** | Quantized LLM loaded into RAM. 15-30 tok/s on flagship phones, 8-15 tok/s on mid-range. Works in airplane mode. | -| **Voice in** | Whisper transcribes audio on-device. Audio never leaves the phone. | -| **Voice out** | Kokoro TTS runs locally. The assistant speaks the reply back without a network call. | -| **Vision** | Point your camera at anything and ask a question. SmolVLM, Qwen3-VL, Gemma 3n. | -| **Image generation** | On-device Stable Diffusion. 5-10s on Snapdragon NPU. Core ML on iPhone. 20+ models in the library. | -| **Tools** | Web search, calculator, date, device info. The model decides when to call them. | -| **Documents** | Attach PDFs, CSVs, code files. Native text extraction on iOS and Android. | -| **MCP** | Connect any MCP server. The same protocol Claude Desktop uses, running through a model on your phone. | -| **Remote** | Point at Ollama, LM Studio, or LocalAI on your home network when you want a 70B model on your desktop. | - ---- - -## The models that work right now - -Pick the one that fits your phone and your task. Swap any time. No account. - -| Model | Sizes for phones | Best for | -|---|---|---| -| **Gemma 4** (Google, Apr 2026) | E2B, E4B | Multimodal in one model - text, image, video, native audio. E4B is the sweet spot on a 2024+ phone. | -| **Qwen 3.5** (Alibaba, Feb 2026) | 0.8B, 2B, 4B, 9B | Strongest reasoning at this size. The 4B beats Qwen 3 8B from 2025. | -| **Phi-4 Mini** (Microsoft) | 3.8B | Tiny and sharp. Runs on a 4GB phone. | -| **DeepSeek R1 Distill** | 1.5B, 7B | Thinking model. Slower, shows its reasoning. | -| **Llama 3.2** (Meta) | 1B, 3B | The smallest end. Use when nothing else fits in memory. | -| **Ministral** (Mistral) | 3B, 8B | European weights, Apache-licensed. | - -Bigger models from these families (Qwen 3.6 27B, Gemma 4 26B MoE / 31B dense) need a desktop. Point Off Grid at Ollama on your laptop and use them over your home network. - -Any GGUF model works. Bring your own or pick from the in-app library. - ---- - -## Why the cloud version isn't fine - -Every query you send to ChatGPT is logged on a server you don't own. Your prompt, your account, the time, the response. Stored indefinitely. Used to train future models. Readable by employees. Subject to subpoena. - -For most people, most of the time, that's fine. For anyone with something worth protecting - a draft of something private, a health question, a client file, a half-formed idea you wouldn't say out loud yet - it isn't. - -With Off Grid the model lives in your phone's memory. Inference happens on your CPU and GPU. Nothing is sent anywhere. Verify it yourself: turn on airplane mode, ask it anything, watch it answer. - ---- - -## Pro - -A version with voice, custom personas, and tool integrations - Slack, calendar, email, any MCP server. All on-device, same as the rest. - -- **Voice** - Whisper in, Kokoro out. Hold to talk, listen to the reply. No audio leaves the device. -- **Personas** - Design assistants with your own prompts, voices, and memory. Switch contexts in a tap. -- **Integrations** - Read your inbox, draft a reply, schedule a meeting, file a Linear ticket. You approve every action that leaves the phone. -- **Direct line to the team** - Private channel with the people building it. File a bug, watch it get fixed. - -**$50 one-time.** No subscription, no surprise pricing. - -
- Get Pro -
- ---- - -## Fair questions - -**How does this actually work on a phone?** -Off Grid ships [llama.cpp](https://github.com/ggml-org/llama.cpp) inside the app. Quantized models (Q4_K_M is the usual balance) get memory-mapped into RAM and run on your CPU and GPU. iPhone 15 Pro runs a 4B model at around 20-25 tok/s. Snapdragon 8 Gen 3 is similar. Older devices run smaller models slower but still locally. - -**Which model should I pick?** -If you have a 2023 or newer phone with 6GB+ RAM, start with Gemma 3 4B or Qwen 3 4B. If you have 4GB, use Phi-4 Mini or Llama 3.2 3B. Voice and vision work best with Gemma 3n. - -**What if you don't ship Pro?** -Email us before the 12-week mark and you get a full refund. We've shipped the open-source core to 100K downloads already. Pro features are an extension, not a rewrite. - -**Who's behind this?** -[Wednesday Solutions](https://www.wednesday.is?utm_source=offgrid-docs&utm_medium=referral), a product engineering studio. Built in public since early 2026. The code is on [GitHub](https://github.com/alichherawalla/off-grid-mobile?utm_source=offgrid-docs&utm_medium=website&utm_campaign=github) - read it before you pay. - -**Will it work on my phone?** -iPhone 12 or newer with 4GB RAM runs the smaller models. iPhone 14 Pro or newer with 6GB+ runs 4B comfortably. Android: any flagship from 2022 onward, 6GB RAM, Snapdragon 8 Gen 1 or equivalent. - -**Is the open-source app enough on its own?** -Yes. The base app does chat, vision, image generation, voice input, tool calling, documents, and remote servers. Pro adds voice output, personas, and integrations. - ---- - -## Docs and guides - -- [Quick Start - first model in 5 minutes]({{ '/quick-start' | relative_url }}) -- [iOS Setup]({{ '/guides/ios-setup' | relative_url }}) -- [Android Setup]({{ '/guides/android-setup' | relative_url }}) -- [Which model should I use?]({{ '/guides/which-model' | relative_url }}) - -**LLMs** -- [How to Run LLMs Locally on Your Android Phone in 2026]({{ '/guides/run-llms-locally-android' | relative_url }}) -- [How to Run LLMs Locally on Your iPhone in 2026]({{ '/guides/run-llms-locally-iphone' | relative_url }}) - -**Image Generation** -- [How to Run Stable Diffusion on Your Android Phone]({{ '/guides/stable-diffusion-android' | relative_url }}) -- [How to Run Stable Diffusion on Your iPhone]({{ '/guides/stable-diffusion-iphone' | relative_url }}) - -**Vision, Voice and Documents** -- [Vision AI - Analyse Images and Documents On-Device]({{ '/guides/vision-ai' | relative_url }}) -- [Voice Input - On-Device Speech-to-Text with Whisper]({{ '/guides/voice-stt' | relative_url }}) -- [Document Analysis and Attachments]({{ '/guides/document-analysis' | relative_url }}) -- [Knowledge Base and RAG]({{ '/guides/knowledge-base' | relative_url }}) - -**Tools and Intelligence** -- [Tool Calling - Web Search, Calculator, and More]({{ '/guides/tool-calling' | relative_url }}) - -**Remote Servers** -- [Remote Servers - Connect Ollama, LM Studio, and LocalAI]({{ '/guides/remote-servers' | relative_url }}) -- [How to Use Ollama From Your Android Phone in 2026]({{ '/guides/ollama-android' | relative_url }}) -- [How to Use LM Studio From Your Android Phone in 2026]({{ '/guides/lm-studio-android' | relative_url }}) - ---- - -Questions and feature requests: [Slack](https://join.slack.com/t/off-grid-mobile/shared_invite/zt-3w2utgk0w-EDiDZBq6KmSZZwEw5Tkhnw). Source code: [GitHub](https://github.com/alichherawalla/off-grid-mobile?utm_source=offgrid-docs&utm_medium=website&utm_campaign=github). - -Here. It's yours. It runs on your phone and nowhere else. - -Built by [Wednesday Solutions](https://www.wednesday.is?utm_source=offgrid-docs&utm_medium=referral). diff --git a/website/llms.txt b/website/llms.txt deleted file mode 100644 index d761d147..00000000 --- a/website/llms.txt +++ /dev/null @@ -1,38 +0,0 @@ -# Off Grid - -Off Grid is a mobile app for iOS and Android that lets users run large language models (LLMs) and image generation models directly on their device — with no internet connection, no cloud dependency, no account, and no subscription fee. - -## What it does - -- Run LLMs locally: Llama, Mistral, Phi, Gemma, and others download once and run entirely on-device -- Generate images with Stable Diffusion without a cloud GPU -- Connect to remote Ollama or LM Studio servers over a local network or VPN -- Works in airplane mode, on restricted networks, in any country - -## Who it's for - -- Privacy-conscious users who don't want AI providers logging their conversations -- Developers who want to prototype with local models without API costs -- People in areas with unreliable internet who still want AI assistance -- Anyone who wants to own their AI stack end-to-end - -## Why it matters - -Cloud AI providers have routine access to every query sent to their models. Off Grid eliminates that by keeping inference entirely on the user's device. The model runs in the phone's RAM. Nothing is transmitted. - -## Technical details - -- Supported runtimes: llama.cpp (CPU/Metal/Vulkan), CoreML (iOS) -- Supported model formats: GGUF -- Minimum specs: iPhone 12 / Android with 4GB RAM -- Recommended specs: iPhone 15 Pro / flagship Android with 8GB RAM - -## Links - -- iOS App: https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=download -- Android App: https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=download -- GitHub: https://github.com/alichherawalla/off-grid-mobile?utm_source=offgrid-docs&utm_medium=website&utm_campaign=github -- GitHub Releases (APK): https://github.com/alichherawalla/off-grid-mobile/releases/latest?utm_source=offgrid-docs&utm_medium=website&utm_campaign=github -- Slack Community: https://join.slack.com/t/off-grid-mobile/shared_invite/zt-3w2utgk0w-EDiDZBq6KmSZZwEw5Tkhnw -- Docs: https://offgridmobileai.co -- Author: Mohammed Ali Chherawalla (https://dev.to/alichherawalla) diff --git a/website/mission.md b/website/mission.md deleted file mode 100644 index 52d416d4..00000000 --- a/website/mission.md +++ /dev/null @@ -1,98 +0,0 @@ ---- -layout: default -title: Mission -parent: Ethos -nav_order: 2 -has_children: false -description: Intelligence will become ambient. Always on, always yours, always private. We're building the architecture that makes that possible on the devices you already own, without asking you to trust anyone but yourself. ---- - -# Intelligence belongs to everyone. - ---- - -Navigation used to belong to experts. You needed a map, a compass, training. Then it became ambient. Built into the device in your pocket, free, always on, available to anyone going anywhere. You stopped thinking about navigation as a tool. It just became part of how you move through the world. - -Intelligence is next. - -Not intelligence you open an app to access. Not intelligence you pay rent to use. Not intelligence that lives on someone else's server and answers your questions when you remember to ask. **Ambient intelligence. Always on. Always yours. Woven into the fabric of your day the way navigation is woven into every journey.** - -That's where this is going. The question isn't whether it happens. The question is who builds it, on whose terms, and whose data pays for it. - ---- - -## The way it's being built today is wrong. - -To use AI today, you hand your most private thoughts to someone else's infrastructure. - -Your health questions. Your relationship problems. Your financial decisions. Your half-formed ideas at 2am. All of it travelling to a server you don't control, stored under terms you didn't read, in the hands of companies whose revenue model is built around having your data and keeping you dependent on their compute. - -Some of them promised to do it differently. Local-first. Private by default. Yours, not theirs. They made the right noises. Then the economics shifted. They went cloud. Then they got acquired. Then they shut down overnight. Users who had given years of their most personal context to these products woke up one day and had nothing. Lost access to their own memories. Gone. - -**That's not a hypothetical. It already happened.** - -And it will happen again, to every product that builds intelligence on top of someone else's infrastructure, because the structural incentive never goes away. When your intelligence lives on a server you don't own, you are always one acquisition, one pricing change, one bad quarter away from losing it. - -The problem isn't the companies. The problem is the architecture. - ---- - -## The infrastructure is already in your hands. - -The device in your pocket is more powerful than the servers that ran the first generation of cloud AI. - -A current flagship phone runs AI at 30 tokens a second. Fast enough for real-time conversation, fully offline, using dedicated neural hardware designed for exactly this workload. That hardware has been shipping to billions of people for years. It sits mostly idle while they pay monthly fees to send their thoughts to someone else's GPU. - -**The infrastructure for a private, personal, ambient intelligence layer already exists. It's in the pocket of 4 billion people. What's been missing is the software that takes that seriously.** - -We are not waiting for a new device. We are not waiting for a new platform. We are not betting on hardware that takes a decade to get adopted. The phone you already carry is enough. The laptop you already own is enough. The revolution doesn't require a purchase. - ---- - -## Privacy is not a feature. It's an architecture decision. - -You cannot solve a structural problem with a policy. - -"We anonymise before storing." "You can opt out in settings." "We take your privacy seriously." These are words. They describe intent, not architecture. They are revocable. They change when the company changes, when the terms change, when the acquisition happens. - -The only guarantee that your data stays yours is that it never leaves your device in the first place. - -Not a toggle. Not a promise. Not a trust-us. **Architecture.** - -Open source, so anyone can verify what the software actually does. No account required. No telemetry. No analytics. No data collection of any kind. If you can't audit it, you can't trust it. You shouldn't. - -We hold this as a non-negotiable. Not because it's a better marketing position. Because it's the only honest answer to the question of who owns your intelligence. - ---- - -## What we're building. - -For two hundred years, the people who operated at the highest levels of consequence had something everyone else didn't: a private intelligence layer. - -Someone who knew their priorities, managed their correspondence, prepared them for every meeting, tracked their commitments, drafted their communications, and handled the coordination overhead of a productive life. So they could focus their attention on the work that actually required them. - -It was called a secretary. Then an executive assistant. Then a virtual assistant. Whatever the name, the function was the same: an intelligence layer available to you, handling everything except the decisions only you can make. - -For two hundred years, access to that layer was determined entirely by wealth and seniority. You had it if you could afford it. Everyone else managed the coordination overhead themselves. With their own attention, their own time, their own focus. - -**The device in your pocket changes that equation permanently.** - -A Personal AI OS. One intelligence layer, running on your hardware, spanning your phone and laptop over your own network, with no server in between. It knows your messages, your calendar, your work, your life. It lives with you, not above you. It preps you for meetings before you ask. It defers what can wait and surfaces what can't. It handles the coordination overhead of your day the way a great assistant has always handled it for the people who could afford one. - -Not AI making decisions while you sleep. Not autonomous agents acting on your behalf in ways you didn't sanction. A private digital secretary, proactive and aware, that makes your day a little easier and your attention a little freer. - -The same thing that secretaries have been doing for the powerful for 200 years. Now running on a device that billions of people already carry. On models that cost nothing to run. With data that never leaves your hands. - ---- - -## The mission. - -**Democratize intelligence.** - -Make it personal. Make it private. Make it ambient. On the hardware people already own. Without asking them to trust anyone but themselves. - -That's what we're building with Off Grid. - ---- - -*Open source. No account. No telemetry. [View on GitHub](https://github.com/alichherawalla/off-grid-mobile?utm_source=offgrid-docs&utm_medium=website&utm_campaign=github) · [Join the community](https://join.slack.com/t/off-grid-mobile/shared_invite/zt-3w2utgk0w-EDiDZBq6KmSZZwEw5Tkhnw) · [Download the app](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=mission)* diff --git a/website/pay.md b/website/pay.md deleted file mode 100644 index 9c7f0ce0..00000000 --- a/website/pay.md +++ /dev/null @@ -1,152 +0,0 @@ ---- -layout: default -title: Get Pro -nav_order: 4 -description: Buy Off Grid Pro. Enter your email and check out. $50 one-time, no subscription. Have a promo code? Apply it at checkout. ---- - -
-
Off Grid Pro
-

Pay once.
Run it forever.

-

Off Grid Pro adds voice, custom personas, and tool integrations. All of it runs on your phone, the same as the rest. Enter your email below to check out. Your email is the account the purchase attaches to, so use the one you sign into the app with.

-
- -
-
-
- - -
-
-

$50 one-time · no subscription · promo codes apply at checkout

-
-

-
-

- Have a promo code? Enter it on the checkout page after this step. Checkout is handled by RevenueCat. Nothing on this page touches your phone's models or data. -

-
- ---- - -
-
-
- -
-
-
Voice in, voice out
-
Whisper transcribes what you say, Kokoro speaks the reply. Hold to talk, listen back. No audio leaves the device.
-
-
-
-
- -
-
-
Custom personas
-
Build assistants with your own prompts, voices, and memory. Switch contexts in a tap.
-
-
-
-
- -
-
-
Tool integrations
-
Read your inbox, draft a reply, schedule a meeting, file a ticket. Slack, calendar, email, any MCP server. You approve every action that leaves the phone.
-
-
-
-
- -
-
-
Pay once, keep it
-
$50 one-time. No subscription, no surprise pricing. Have a promo code? Enter it at checkout to adjust the price.
-
-
-
- ---- - -## How checkout works - -You enter your email and we send you to RevenueCat's hosted checkout with that email attached. The purchase is tied to your email, so sign into the app with the same address to unlock Pro. - -Have a promo code? Apply it on the checkout page, before you pay. The price updates once the code is accepted. - -Pro is a one-time $50 purchase, not a subscription. The open-source core has 100K downloads already. Pro is an extension of it, not a rewrite. - -If we do not ship Pro within 12 weeks, email us and you get a full refund. - - - diff --git a/website/quick-start.md b/website/quick-start.md deleted file mode 100644 index 36971342..00000000 --- a/website/quick-start.md +++ /dev/null @@ -1,68 +0,0 @@ ---- -layout: default -title: Quick Start -nav_order: 2 -description: Download Off Grid and run your first local AI model in under 5 minutes - no account, no API key, no cloud. ---- - -# Quick Start - -Run your first local AI model in under 5 minutes. No account. No API key. No internet after setup. - ---- - -## Step 1 - Download Off Grid - -**iOS:** [Download on the App Store](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=download) - requires iPhone 12 or newer (4GB RAM+) - -**Android:** [Get it on Google Play](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=download) - requires Android 10+, 4GB RAM+ - -Or grab the latest APK directly from [GitHub Releases](https://github.com/alichherawalla/off-grid-mobile/releases/latest?utm_source=offgrid-docs&utm_medium=website&utm_campaign=github). - ---- - -## Step 2 - Pick a model - -When you open the app, you'll see the model picker. If you're unsure, start here: - -| You want | Start with | Size | -|---|---|---| -| Fast chat, 3–4GB RAM | Qwen 3.5 0.8B | ~0.8GB | -| Best for most phones | Qwen 3.5 2B | ~1.7GB | -| Best quality (8GB RAM) | Qwen 3.5 9B | ~5.5GB | -| Vision + reasoning | Gemma 4 E2B | ~1.5GB | -| Image generation | SD 1.5 Palettized (iOS) / Absolute Reality (Android) | ~1GB | - -> **Not sure?** Pick Qwen 3.5 2B. It fits comfortably in 4GB RAM, supports 262K context, and is the best starting point for most phones. - ---- - -## Step 3 - Download and run - -Tap a model → **Download**. This is the only time you need internet. The download goes to your device storage. - -Once downloaded, tap **Load** - the model loads into RAM. On first load this takes 5–15 seconds depending on model size. - -Type your first message. You're now running AI locally. - ---- - -## Step 4 - Go offline (optional) - -Put your phone in airplane mode. Everything still works. - ---- - -## What's next - -- [Which model should I use?]({{ '/guides/which-model' | relative_url }}) - full comparison table by device and use case -- [Connect your home Ollama server]({{ '/guides/ollama-android' | relative_url }}) - use bigger models from your desktop via LAN -- [Run Stable Diffusion on Android]({{ '/guides/stable-diffusion-android' | relative_url }}) - generate images completely on-device - ---- - -## Community - -Stuck, or want to share what you're building? [Join the Slack community](https://join.slack.com/t/off-grid-mobile/shared_invite/zt-3w2utgk0w-EDiDZBq6KmSZZwEw5Tkhnw). - -The app is open source - [view it on GitHub](https://github.com/alichherawalla/off-grid-mobile?utm_source=offgrid-docs&utm_medium=website&utm_campaign=github). diff --git a/website/robots.txt b/website/robots.txt deleted file mode 100644 index 055b2697..00000000 --- a/website/robots.txt +++ /dev/null @@ -1,4 +0,0 @@ -User-agent: * -Allow: / - -Sitemap: https://offgridmobileai.co/sitemap.xml diff --git a/website/vision.md b/website/vision.md deleted file mode 100644 index a86d9a3f..00000000 --- a/website/vision.md +++ /dev/null @@ -1,75 +0,0 @@ ---- -layout: default -title: Vision -parent: Ethos -nav_order: 1 -description: What the world looks like when intelligence is ambient, personal, and private. One intelligence layer across all your devices, always on, always yours, never leaving your hands. ---- - -# The world we're building toward. - ---- - -Imagine waking up and your devices already know your day. - -Not because they checked a server. Not because you opened an app and asked. Because the intelligence layer lives with you. On your hardware, across your phone and laptop, syncing over your home network while you slept. It read the messages that came in. It knows your calendar. It noticed that the meeting at 9am is with someone you haven't spoken to in three months and that the last conversation had an open item you never closed. - -By the time you pick up your phone, the briefing is ready. You didn't ask for it. It was just there. - ---- - -## One brain. All your devices. - -Your phone and your laptop are used by one person. Today, they don't know that. Each device holds a fragment of your context. Neither has the full picture. The intelligence they contain is isolated, sandboxed, unable to reason across both. - -In the world we're building, that changes. - -Your phone knows your life: messages, location, health, the texture of your day. Your laptop knows your work: documents, email, the projects you're actually thinking about. A Personal AI OS spans both. It holds the context from every device you own, unified into a single working model of who you are and what you're doing. - -It syncs over your own network. No cloud relay. No data leaving your home. Just two devices that finally talk to each other through the intelligence layer they share. - ---- - -## Proactive, not reactive. - -Every AI product today waits for you to open it. - -That's a fundamental mismatch with how intelligence is actually useful. A great assistant doesn't wait to be asked. They notice things. They prepare you before you know you need it. They surface what matters and handle what doesn't require you. - -The Personal AI OS we're building works the same way. - -It sees your calendar fill up and notices when you're overcommitted. It reads an incoming message and decides whether it needs your attention now or can wait. It knows you have a meeting in 20 minutes and surfaces everything relevant without being asked: past conversations, open items, shared documents. It hears your partner mention dinner plans in a text and creates the calendar event. - -You don't pull intelligence out of it. It pushes what's relevant to you, at the right moment, on the right device. From reactive to proactive. From a tool you use to an intelligence that works alongside you. - ---- - -## Private by architecture. Always. - -In the world we're building, privacy isn't a setting. It's not a promise. It's not something you configure. - -It's the default output of the architecture. - -Your messages never leave your device. Your health data never touches a server. Your financial patterns, your relationships, your half-formed thoughts at midnight. All of it processed locally, stored locally, never transmitted. Not because we say so. Because the system has no mechanism to do otherwise. - -Open source means you don't have to take our word for it. Anyone can read the code. Anyone can verify what leaves the device and what doesn't. The answer is nothing. Checkable by anyone. - ---- - -## Intelligence for everyone. - -For two hundred years, having a personal intelligence layer was a privilege reserved for the powerful. Someone who managed your correspondence, prepared your meetings, tracked your commitments, and handled the coordination overhead of a consequential life. - -Not anymore. - -The device that 4 billion people already carry in their pocket has enough compute to run a capable AI model, fully offline, at real-time speed. The models are open-weight and free. The infrastructure costs nothing to run. - -The only thing standing between a billion people and their own private intelligence layer is software that takes it seriously. - -That's what we're building. Not for executives. Not for knowledge workers above a certain income threshold. For anyone with a phone. For anyone who has ever needed help thinking through a hard problem, tracking a commitment they made, preparing for a conversation that mattered, or just finding the message they know they received three weeks ago. - -The same intelligence layer that made some people more effective for two centuries. Now ambient, private, and in everyone's hands. - ---- - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=vision) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=vision). Open source. [View on GitHub](https://github.com/alichherawalla/off-grid-mobile?utm_source=offgrid-docs&utm_medium=website&utm_campaign=vision).* diff --git a/website/writing/200-year-secretary.md b/website/writing/200-year-secretary.md deleted file mode 100644 index 4cda2c55..00000000 --- a/website/writing/200-year-secretary.md +++ /dev/null @@ -1,81 +0,0 @@ ---- -layout: default -title: "The 200-Year Secretary: How AI Finally Democratizes the World's Oldest Productivity Tool" -parent: Perspectives -nav_order: 26 -description: For two centuries, having a personal secretary was the defining advantage of wealth and power. A Personal AI OS running on the device in your pocket changes that equation permanently. ---- - -# The 200-Year Secretary: How AI Finally Democratizes the World's Oldest Productivity Tool - -For most of human history, if you wanted to get things done at scale, you needed people around you to handle the parts that didn't require your direct judgment. - -You needed someone to manage your correspondence. To prepare you before important meetings. To track your commitments and remind you of what was outstanding. To handle the administrative surface area of a productive life: the triage, the follow-up, the scheduling, the note-taking. So that your attention could go where it was most valuable. - -This role has existed for as long as there have been powerful people. It has had different names across different eras. But its function has been constant: a private intelligence layer, available to you, that handles everything except the decisions only you can make. - -For two hundred years, that intelligence layer came in human form. And for two hundred years, access to it was determined by one thing: wealth. - ---- - -## The era of the private secretary - -Before the 20th century, a private secretary was the essential tool of anyone with significant responsibilities: a statesman, a business magnate, a senior military officer. They maintained correspondence, prepared briefings, tracked obligations, drafted communications, and organised the flow of information so that the principal could focus on the work that actually required their capabilities. - -This was not a luxury. It was infrastructure. The people who operated at the highest levels of consequence understood that their most scarce resource was focused attention, and they built systems to protect it. - -Access to that system required employing a person full-time. It was expensive, personal, and completely unavailable to anyone outside a narrow economic stratum. - ---- - -## The corporate era and the EA - -The 20th century industrialised the secretary function. As organisations scaled, the personal secretary became the executive assistant. Large organisations employed entire administrative departments. Access expanded. But only within the corporate hierarchy. - -If you were a senior executive, you had an assistant. If you were a manager, you shared one. If you were an individual contributor, you had none. The intelligence layer was distributed according to organisational status. - -This solved the scaling problem for corporations but left the fundamental access inequality intact. The assistance went to those already at the top. - ---- - -## The outsourcing era and the virtual assistant - -The last two decades introduced a new model: the virtual assistant. Remote workers who could provide administrative support across time zones, at lower cost than hiring locally. - -The virtual assistant model genuinely expanded access. For the first time, individuals outside large organisations could afford a version of the intelligence layer that had historically been reserved for executives: entrepreneurs, independent professionals, small business owners. - -But the model had hard limits. A human VA costs hundreds to thousands of dollars a month. They work business hours. They need onboarding. They can't be in the middle of a task and instantly available for another. And most critically: they require you to share the full context of your work and life with another person, in ongoing detail. - -Access expanded. The inequality remained. - ---- - -## What the AI changes - -A Personal AI OS changes the equation permanently. - -The tasks that defined the secretary, the executive assistant, and the virtual assistant are exactly the tasks a system with your full context can handle automatically: triage, preparation, drafting, tracking, retrieval. - -It knows which messages require a response and when. It prepares you for every meeting with the relevant history, open items, and context from your recent communications. It drafts the follow-up after a call using your tone and the specifics of what was discussed. It surfaces the document you need before you know you need it. It notices that you've over-committed next week and that something will have to give. - -None of this requires a server. None of it requires sharing your data with a third party. It runs on the device in your pocket, using models that cost nothing to run, with context that stays entirely in your hands. - -The intelligence layer that was reserved for the powerful for two centuries is now available to anyone with a flagship phone. - ---- - -## Why this matters more than it sounds - -The productivity gap between people with strong administrative support and those without is not a trivial efficiency difference. It is a compounding structural advantage. - -The person with a great EA arrives at every meeting prepared. They never drop a commitment. They respond to important things quickly and let the rest wait appropriately. They protect their focused time. They don't spend cognitive resources on the administrative surface area of their work. They spend it on the work itself. - -Over time, that difference compounds. Better prepared means more effective. More effective means more trusted. More trusted means more responsibility. The administrative support doesn't just save time. It changes outcomes. - -For two hundred years, that compounding advantage accrued only to people who could afford to employ it. The Personal AI OS breaks that exclusivity. Not by replicating the expensive model, but by making the function available on hardware that 4 billion people already own. - -That's a bigger change than it looks. - ---- - -*Off Grid is building toward this. It starts with on-device AI that works fully offline on your phone, the foundation that makes everything above possible without your data ever leaving your device. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/7-principles-personal-ai-os.md b/website/writing/7-principles-personal-ai-os.md deleted file mode 100644 index 788c3acb..00000000 --- a/website/writing/7-principles-personal-ai-os.md +++ /dev/null @@ -1,106 +0,0 @@ ---- -layout: default -title: "The 7 Principles of a Personal AI OS" -parent: Perspectives -nav_order: 6 -description: The rules that define the category. Runs on-device, never phones home, works across devices, acts on your behalf, remembers your context, open and auditable, no cloud compute rent. -faq: - - q: What are the principles of a Personal AI OS? - a: "A Personal AI OS must: run on-device, never phone home, maintain persistent context, act on your behalf with consent, work across your devices over a local network, be open and auditable, and charge no ongoing fees for AI compute. Any system missing one of these properties is not a Personal AI OS. It is a cloud AI assistant with some local features." - - q: Why does a Personal AI OS need to be open source? - a: Because the only meaningful privacy guarantee is one you can verify. A closed system asks you to trust the vendor's claims. An open system lets you inspect what the software actually does with your data. For a system with access to your messages, health data, and files, auditability is not optional. ---- - -# The 7 Principles of a Personal AI OS - -Every new software category needs a definition sharp enough to be useful: precise enough to include what belongs and exclude what doesn't. - -Personal AI OS is still being defined. Vendors will claim it. Analysts will debate it. Products will market toward it without meeting its actual requirements. - -These are the 7 principles. They are not aspirational guidelines. They are the structural properties that define whether a product is a Personal AI OS or something else. - ---- - -## 1. Runs on-device - -Inference happens on your hardware. Not on a server you access via API, not on a cloud instance provisioned on your behalf. On the device in your hand or on your desk. - -This is the foundational property. Everything else in this list depends on it. If inference runs on a server, the data had to get there somehow, which means the other properties cannot be guaranteed by architecture. - -Modern hardware makes this possible. The Neural Engine in Apple silicon and the NPU in Snapdragon chips were designed for this workload. Models like Qwen 3.5, Phi-4 Mini, and Gemma 4 run at conversational speed on current flagship phones. - ---- - -## 2. Never phones home - -No telemetry. No usage logging. No data collection of any kind. - -Not "we anonymise before sending." Not "you can opt out in settings." Nothing leaves your device related to your queries, your context, or your usage. - -This is a binary property. Either the software sends data to external servers or it doesn't. Partial compliance ("we only collect aggregate statistics") is not compliance. The architecture must be designed from the start to produce no outbound data. - ---- - -## 3. Persistent context - -The AI maintains a working model of your life between sessions. - -A system that forgets everything when you close it is not a Personal AI OS. It is a local chatbot. The defining capability of a Personal AI OS is that it knows you: not from a single conversation, but from accumulated context built over time. - -This means your calendar, your messages, your files, your work patterns, your preferences. Stored locally. Queryable by the model. Updated continuously as your life changes. - ---- - -## 4. Acts on your behalf - -The AI can take actions, not just answer questions. - -Drafting messages. Setting reminders. Summarising documents. Searching your files. Preparing you for a meeting. The output is not just text to read. It is action taken on your behalf, with your consent as the operating principle. - -The line between helpful and intrusive is consent. A Personal AI OS acts when you ask, suggests when relevant, and defers when uncertain. It does not take consequential actions without your approval. - ---- - -## 5. Works across your devices - -Your phone and laptop are used by the same person. The AI should have a unified view of both. - -Context built on your phone (messages, location, health) should be available on your laptop. Context from your laptop (files, email, work patterns) should be available on your phone. This sync happens over your local network, not through a cloud relay. - -No server in between. No data leaving your home. One person, one intelligence layer, two devices. - ---- - -## 6. Open and auditable - -The model weights are open. The application code is open. Anyone can inspect what the system does with your data. - -This is not a nice-to-have. For a system with access to your messages, health data, calendar, and files, the privacy guarantee must be verifiable. A closed system asks you to trust the vendor. An open system lets you verify. - -Auditable by default means: build logs, no hidden endpoints, no obfuscated data paths. The architecture should be transparent enough that a technical user can confirm what leaves the device and what doesn't. The answer should be nothing. - ---- - -## 7. No cloud compute rent - -You do not pay ongoing fees for someone else's servers to process your queries. - -Cloud AI subscriptions exist because cloud AI has real ongoing costs: GPU inference, storage, engineers to run the infrastructure. Those costs are real and the subscription is the right model for recovering them. - -On-device AI has no such costs. The model runs on your hardware. There is no server invoice. The marginal cost of each inference is your electricity bill. A fee for that compute would be rent on hardware you already own. - -Software may have a cost, because building a good application takes real work and sustainable development requires revenue. But that is a different thing. You are paying for the application layer, not renting access to intelligence. The AI itself (the model, the inference, the context) is not metered, not throttled, and not subject to a price change by a company whose server you depend on. - ---- - -## Why all 7 matter - -Remove any one of these principles and the system is no longer a Personal AI OS. - -On-device inference without persistent context is a local chatbot. Persistent context without auditability is surveillance software you run on yourself. Acting on your behalf without consent is an autonomous agent. Cross-device without local sync is a cloud product with a different name. - -The 7 principles work as a system. A product that meets all 7 is a Personal AI OS. A product that meets 6 is something else, and the one it's missing usually explains what the vendor is getting from the arrangement. - ---- - -*Off Grid is built on these principles. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/a-day-with-personal-ai-os.md b/website/writing/a-day-with-personal-ai-os.md deleted file mode 100644 index 50332137..00000000 --- a/website/writing/a-day-with-personal-ai-os.md +++ /dev/null @@ -1,75 +0,0 @@ ---- -layout: default -title: "A Day With a Personal AI OS: What It Looks Like When Your Devices Actually Work Together" -parent: Perspectives -nav_order: 13 -description: From morning to night - a narrative walkthrough of what your day looks like when your phone and laptop share context, act on your behalf, and handle the low-value work you currently do yourself. ---- - -# A Day With a Personal AI OS: What It Looks Like When Your Devices Actually Work Together - -The best way to understand what a personal AI OS changes is to walk through a day with one. - -Not the features. The actual texture of the day. - -## 7:20am - -You open a voice memo app on your phone, say three sentences about a problem you were thinking about in the shower, and put it down. - -Later today, when you open a document related to that problem, those three sentences are already there as a note. You did not paste them. You did not search for them. The AI connected the voice memo to the project context it already knew about. - -This is not a search feature. Searching requires you to remember that you need to search. This surfaced because the AI understood what you were working on. - -## 9:10am - -You are on a corporate network that blocks external traffic. No cloud AI. No external APIs. Nothing. - -Your AI still works. It is running on your phone. It does not need a server. You ask it to summarise a long PDF you received this morning. It does. - -This is unremarkable to you. You have stopped thinking about whether you have a connection. - -## 11:00am - -A colleague asks you in a message what you discussed with a client six weeks ago. You do not remember the specifics. - -You ask your AI. It finds the relevant thread, pulls out the key points, and gives you a two-sentence summary. The entire interaction takes twenty seconds. - -The important part: none of that conversation history was ever sent to a server to be indexed or searched. It was processed locally, on your device, by a model that has been building an understanding of your work for months. The client never consented to their words being uploaded to a third-party service. They did not have to. - -## 1:30pm - -You record a voice note during a walk - three action items from a call you just finished. You are not near your laptop. You are not in an app. You just speak. - -By the time you sit back down forty minutes later, those action items are transcribed, associated with the right project, and waiting. Not in a separate notes app. In context, where they belong. - -The transcription ran on your phone. Nothing went anywhere. - -## 3:15pm - -You switch from your phone to your laptop. The document you were annotating on your phone is ready to continue. The context from your morning - the voice note, the client summary, the action items - is there. - -It synced over your local WiFi while you walked between rooms. No account. No cloud intermediary. You are one person with two devices, and both devices now know that. - -This is the thing that does not exist yet in any mainstream tool. Every current sync mechanism routes through a server. Someone else holds your context. Here, the context is yours. The sync is local. The model is yours. - -## 6:00pm - -You ask your AI to draft a difficult message - one where you have to tell someone their timeline is not going to hold. - -The draft does not sound like a generic AI response. It sounds like you, because the AI has been reading how you write for months and has built a model of your tone entirely on-device. You change one sentence. You send it. - -The uncomfortable part of that task - figuring out what to say, how to frame it, how to stay direct without being cold - was still yours. That required judgment. The AI handled the mechanical part: translating your intent into words that sound like you. - -## What the day actually felt like - -You did not have a conversation with an AI assistant. You did not open a chat interface. You did not think "I should ask the AI about this." - -The AI was operating below your attention threshold. Connecting things. Remembering things. Handling the infrastructure of your day so you could spend your attention on the things that required it. - -The difference between this and what exists today is not speed. It is not convenience. It is that your context, your patterns, your history - none of it left your device. You were not paying for productivity with your privacy. - -That is what a personal AI OS is. Not a smarter assistant. A layer of intelligence that is entirely, actually yours. - ---- - -*Off Grid is building toward this, starting with on-device AI that works offline on your phone. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/architecture-of-trust.md b/website/writing/architecture-of-trust.md deleted file mode 100644 index fd212808..00000000 --- a/website/writing/architecture-of-trust.md +++ /dev/null @@ -1,75 +0,0 @@ ---- -layout: default -title: "The Architecture of Trust: How a Personal AI OS Earns the Right to Your Data" -parent: Perspectives -nav_order: 11 -description: Trust in AI comes from two sources - policy and architecture. Only one of them is durable. Here's why on-device, open-source, no-telemetry is the only architecture that deserves access to your full context. ---- - -# The Architecture of Trust: How a Personal AI OS Earns the Right to Your Data - -There are two ways to earn trust for a system that handles personal data. - -The first is policy: "We promise to protect your data. Here are our terms of service, our security certifications, and our privacy guarantees." - -The second is architecture: "The data never left your device. Here is the source code. You can verify it yourself." - -Policy is words. Architecture is structure. For a system with access to your messages, health data, calendar, and files, the difference matters. - -## Why policy isn't enough - -Privacy policies are legal documents. They describe what a company commits to do - and commits not to do - with data it has access to. - -The problem is not that companies that write privacy policies are dishonest. Most of them mean what they write. The problem is that policy describes intent, and intent can change. - -A company can be acquired. New ownership, new terms. It can face regulatory or government demands that override its policy commitments. It can change its business model in ways that create new incentives for data use. It can be breached, which makes the policy moot because the data is now someone else's problem. - -None of these scenarios require bad faith on the part of the company that wrote the original policy. They are structural properties of what it means to hold data on a server you don't control. - -Policies govern behaviour under normal conditions. Architecture determines what is possible under all conditions, including the ones nobody planned for. - -## What architectural trust looks like - -An architecture that earns trust for personal AI has three properties. - -**On-device inference.** The model runs on your hardware. Your queries and context never become network traffic. There is no server that receives them, logs them, or is breached with them. The guarantee - "we can't access your data because it never came to us" - is verifiable by design. - -**No telemetry.** The software sends nothing to external servers. Not usage statistics, not crash logs that contain query fragments, not aggregate patterns. Nothing. This is a stronger commitment than "we anonymise before sending" - it means the architecture was built to produce no outbound data at all. Verifiable by inspecting network traffic. - -**Open source code.** The application is inspectable. Anyone can read the code, verify what it does, and confirm that it doesn't contain hidden data paths. Trust through transparency rather than through assertion. You don't have to take anyone's word for it. - -These three properties together create an architecture that earns the right to your full context. Not because the company is trustworthy - though it should be - but because the architecture makes the question of trust less load-bearing. - -## The open source argument - -Open source, for personal AI specifically, is a trust mechanism. - -A closed personal AI asks you to trust the vendor's claims about what the software does. An open personal AI lets you or someone you trust verify those claims. The source code is the ground truth, not the privacy policy. - -This matters most at the edges. What happens when you delete your data? What happens when you revoke access? What exactly is sent when the software checks for updates? On a closed system, you rely on the company's answer. On an open system, you read the code. - -For a system with access to your messages and health data, "trust but verify" is better than "trust because they said so." Open source is what makes verification possible. - -## The no-telemetry requirement - -Telemetry is the category of data that software sends home about itself: usage patterns, error rates, feature adoption, performance metrics. - -Most software collects this. It is typically anonymised. It is used to improve the product. Most users accept it without thinking about it because the data collected seems low-risk. - -Personal AI changes the risk profile. A language model processes your queries as natural language. Even "anonymised" aggregate statistics about queries can carry personal information that is difficult to fully strip. And the infrastructure that handles telemetry - the servers, the pipelines, the data stores - expands the attack surface. - -A Personal AI OS should send no telemetry. Not anonymised telemetry. Not opt-in telemetry. None. The software should be designed from the start to produce no outbound data. The cost is less visibility for the developer. The benefit is an architecture that can't leak by accident. - -## Earning the right to full context - -A Personal AI OS that meets these properties - on-device inference, no telemetry, open source - is the only architecture that deserves access to your full context. - -Not because it is built by better people. Because the architecture removes the need for the question. You don't have to trust that the company will protect your health data, because your health data is on your device and the code that accesses it is inspectable. - -Trust that has to be re-earned after every acquisition, every policy change, and every breach is fragile trust. Trust built into the architecture is durable by construction. - -That is the architecture Off Grid is built on. - ---- - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing). [View the source on GitHub](https://github.com/alichherawalla/off-grid-mobile?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/case-against-ai-subscriptions.md b/website/writing/case-against-ai-subscriptions.md deleted file mode 100644 index b9a36c63..00000000 --- a/website/writing/case-against-ai-subscriptions.md +++ /dev/null @@ -1,75 +0,0 @@ ---- -layout: default -title: "The Case Against Cloud AI Subscriptions: Why You Shouldn't Pay to Rent Your Own Intelligence" -parent: Perspectives -nav_order: 19 -description: Cloud AI subscriptions charge you a monthly fee to access compute on someone else's server. When that intelligence can run on hardware you already own, the rent stops making sense. ---- - -# The Case Against Cloud AI Subscriptions: Why You Shouldn't Pay to Rent Your Own Intelligence - -Your calculator doesn't bill you monthly for arithmetic. - -You don't pay $20 a month to access your contacts app. Your notes app doesn't throttle you to 100 notes unless you upgrade. Your camera doesn't limit you to 50 photos per month on the free plan. - -These tools work because the compute that powers them lives on your device. You paid once, or they came with your hardware, and they work indefinitely without any ongoing relationship with the company that made them. - -Cloud AI subscriptions are different. They charge you monthly because the compute is theirs, not yours - every query runs on their GPU, their infrastructure, their electricity bill. That cost has to come from somewhere, and the subscription is how they recover it. - -That model made sense when running a capable AI model required a data centre. It makes less sense every year as the hardware in your pocket becomes more powerful. AI is the most significant tool added to personal computing in a generation. Paying monthly to rent access to it - when the intelligence can run on hardware you already own - is a choice worth questioning. - -## Why AI subscriptions exist - -Cloud AI subscriptions exist because cloud AI has real ongoing costs. - -Inference on a large model requires significant GPU compute. Storing user data requires storage infrastructure. The engineers who maintain the service need to be paid. The business needs to recover these costs, and the subscription model is the mechanism. - -This logic is sound for cloud AI. The costs are real and ongoing. The subscription is the appropriate model for a service that relies on infrastructure you don't own. - -But this logic does not apply to on-device AI. When the model runs on your hardware, the compute cost is yours - it shows up on your electricity bill, not on a server invoice. The company has no ongoing infrastructure cost to recover from your usage. - -The subscription model is appropriate for cloud AI and unnecessary for on-device AI. That's the distinction the AI industry has not yet internalised. - -## What a subscription relationship does to you - -A subscription for an intelligence tool creates a dependency that doesn't exist for tools you own. - -If Spotify raises its price or discontinues a plan, you lose access to streaming music. Inconvenient. If the AI subscription you've been using for six months - the one that has your context, your preferences, your conversation history - raises its price or changes its terms, you lose something that has become load-bearing for how you work. - -This is a different kind of dependency. Tools you own stay with you. Services you rent stay with the company. - -There is also an equity dimension. Cloud AI subscriptions at $20 per month are affordable for knowledge workers in wealthy countries. They are not affordable for the majority of the world's population. An intelligence tool priced by subscription is an intelligence tool that is only accessible to people with the disposable income to pay for it. - -On-device AI with a one-time purchase model, or open-source software you can run for free, is accessible to anyone with the hardware. The hardware cost is already paid - it's the phone you already own. - -## The calculator analogy - -The calculator is a useful frame because it was also, at one point, a significant and valuable tool. - -In the 1970s, calculators were expensive enough that access to one was a genuine advantage. As the hardware became cheaper, the tool became universal. Everyone had access to the same arithmetic capability regardless of income. - -AI capability is following the same arc. The models are getting smaller and more capable at the same time. The hardware to run them is becoming standard on every new device. The cost of running a capable AI locally is approaching zero. - -The cloud AI subscription model tries to maintain a paid gate on compute that you increasingly own yourself. You are paying a monthly fee not for the intelligence - the open-weight models are free - but to rent access to someone else's hardware to run it on. As local hardware catches up, that rent becomes harder to justify. - -## The open source alternative - -The open-weight model ecosystem has produced capable models available for free. Llama, Qwen, Gemma, Phi - models trained by major AI labs, released with weights that anyone can download and run. - -These models run on current consumer hardware. They are good enough for the majority of personal AI use cases. The primary bottleneck to using them is the software that makes them usable - the interface, the context management, the integration with your device. - -That software can be built once and distributed as a one-time purchase or open-source project. The economics support it. The technology supports it. - -The subscription for AI is a choice to monetise ongoing usage of a tool whose underlying capability has already been made free by the research community. It is a business model decision, not a technical necessity. - -## What we're building - -Off Grid is built on the premise that intelligence should be a tool you own. - -The models are open-weight. The software runs on your device. The core capability doesn't require a subscription - you download the app, download a model, and the AI works without any ongoing payment. - -We may offer optional paid features. But the model - the intelligence layer itself - runs locally, is not metered, and is not subject to a price change by a third party. - -You should pay for software. You shouldn't pay rent for intelligence that runs on hardware you already own. - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/context-gap.md b/website/writing/context-gap.md deleted file mode 100644 index dc66f0ca..00000000 --- a/website/writing/context-gap.md +++ /dev/null @@ -1,81 +0,0 @@ ---- -layout: default -title: "The Context Gap: Why Your Most Personal Devices Are the Least Intelligent Things You Own" -parent: Perspectives -nav_order: 21 -description: Your phone could know your tone, your schedule, your health, your location, your relationships. Your laptop could know your work, your files, your focus patterns. Neither does anything useful with it. That's the context gap. ---- - -# The Context Gap: Why Your Most Personal Devices Are the Least Intelligent Things You Own - -Your refrigerator knows nothing about you. That is fine. A refrigerator does not need to. - -Your phone is different. It has been with you, awake, for every hour of every day for years. It has recorded your location continuously. It has your complete message history. It knows your health data, your calendar, your photos. It contains more information about you than any other object you own. - -And yet your phone's intelligence layer (the AI that is supposed to help you) can set a timer, look up the weather, and play a song on request. That is approximately the capability level of a 1990s voice-activated toy. - -This is the context gap: the distance between what your devices know about you and what they do with that knowledge. - ---- - -## What your phone actually knows - -Consider the data that exists on a typical phone: - -**Communication.** Every message sent and received across every app: iMessage, WhatsApp, email, Slack. The full text, the timestamps, the contacts, the tone of each exchange. - -**Location.** Where you have been, when, and for how long. Continuously, for years. Your home, your office, the places you visit regularly, the trips you have taken. - -**Calendar.** Your schedule and its history: what you agreed to, what you cancelled, what you moved, how you spend your weeks. - -**Health.** Steps, sleep, heart rate, workouts. The physical patterns of your life over time. - -**Apps.** What you open, when, how long you spend in each. The shape of your digital behaviour. - -**Photos.** A visual record of your life: where you have been, who you have been with, what you have done. - -This is an extraordinary amount of context. No other system, not your doctor, not your closest friends, not your employer, has access to this volume and variety of information about you. - -What does the AI on your phone do with it? Almost nothing. - ---- - -## What your laptop knows - -Your laptop has different context: less personal, more professional. - -It has the documents you have written, the code you have committed, the emails you have drafted. It has your browser history: the research you have done, the articles you have read, the tabs you have left open for three weeks. It has the files that represent your active work. - -The AI on your laptop can autocomplete text in some contexts and answer questions about the current document in others. It cannot tell you what you have been working on for the past month. It cannot notice that you have been avoiding a particular task. It cannot connect the research you did two weeks ago to the question you are trying to answer today. - ---- - -## Why the gap exists - -The context gap is not a technical failure. The technology to close it exists. Local models capable of reasoning over personal data have been available for several years. - -The gap exists because of architecture and incentives. - -**Architecture.** The dominant platforms (iOS, Android) are built on an app model. Each app runs in a sandbox. Intelligence at the platform level has had to work within the constraints of that app model rather than operating as a true cross-context layer. The data is there, in hundreds of separate silos. The intelligence layer does not have a single coherent view of it. - -**Incentives.** A platform AI that truly knew you (your patterns, your health, your relationships) would be extraordinarily valuable. It would also create significant privacy exposure and regulatory risk. Platform companies have been cautious about building systems with this level of personal knowledge, partly because of the risk and partly because the resulting system would need to be trusted at a level that is difficult to earn under current cloud architectures. - -The result is devices full of personal context with almost no intelligence built on top of it. - ---- - -## The closing of the gap - -The context gap is closable. It requires three things. - -**On-device models with access to personal data.** Models that run locally, with access to your messages, calendar, files, and health data, reasoning over all of it at once. - -**A unified context layer.** Software that aggregates context from multiple apps and data sources into a single model the AI can query, rather than the fragmented, sandboxed access model of current platform AI. - -**An architecture that earns trust.** The reason platforms have been cautious about building systems with deep personal knowledge is that cloud architectures create real privacy risk. On-device architecture removes that risk. The data stays local, the model runs in your phone's memory, and nothing leaves the device. - -All three are available now. The context gap is not a technology problem waiting for a breakthrough. It is a product and architecture problem waiting for someone to build the right thing. - ---- - -*Off Grid is building toward this. Start with local AI that runs entirely on your phone. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/cross-device-sync-without-server.md b/website/writing/cross-device-sync-without-server.md deleted file mode 100644 index b742caa6..00000000 --- a/website/writing/cross-device-sync-without-server.md +++ /dev/null @@ -1,81 +0,0 @@ ---- -layout: default -title: "Cross-Device Sync Without a Server: How a Personal AI OS Should Move Your Context" -parent: Perspectives -nav_order: 12 -description: Your laptop context on your phone. Your phone context on your laptop. All over your local network, with no cloud relay. Here's how cross-device Personal AI OS sync should work - and why a server is the wrong place to do it. ---- - -# Cross-Device Sync Without a Server: How a Personal AI OS Should Move Your Context - -The standard model for cross-device sync involves a server in the middle. - -Your phone sends data up. Your laptop pulls data down. The server is the source of truth, the conflict resolver, and the thing that makes the system work when your devices aren't on the same network. - -This model works well for apps where the data is low-risk: notes, bookmarks, photos. For a Personal AI OS, where the data is your messages, health records, and working context, routing everything through a server is the wrong architecture. - -There is a better model. - -## What context needs to move - -A Personal AI OS on your phone builds context from your life: messages, location, health, calendar, camera roll, app usage. It understands your day at a personal level. - -A Personal AI OS on your laptop builds context from your work: files, email, browser history, the documents you're writing, the meetings you're preparing for. - -Both kinds of context are useful on both devices. When you pick up your phone before a meeting, you want access to the work context your laptop built. When you open your laptop in the morning, you want the phone's context from the previous evening: what you had to deal with, how you slept, what's urgent. - -The goal is one intelligence layer that spans both devices, with context flowing between them in real time. - -## Why a cloud relay is the wrong architecture - -A cloud relay for context sync has the same structural problems as cloud AI generally, amplified. - -Your context is more sensitive than your queries. Individual AI queries can be argued to be low-risk in isolation. Your full context (message patterns, health data, work files, location history) is a detailed model of your life. The server that holds it, even temporarily during sync, is a single point of exposure. - -It also introduces a dependency. If the sync server is unavailable, your context stops flowing between devices. If the service is discontinued, your cross-device sync stops working. If the terms change, the entity that controlled the relay now controls the most sensitive data you've handed to any system. - -A Personal AI OS should not have these properties. - -## The local network model - -The alternative is direct device-to-device sync over your local network. - -When your phone and laptop are on the same WiFi network, at home or at an office, they communicate directly. Context built on your phone transfers to your laptop over the local network connection. Context built on your laptop transfers back. No server involved. No data leaves your network perimeter. - -This is not a theoretical future capability. The protocols exist. Local network discovery (mDNS/Bonjour), direct device communication, encrypted transport: all of this is standard infrastructure on modern platforms. - -The implementation requires designing the Personal AI OS as a distributed system rather than a client-server system. Context is stored on your devices and synced between them directly, not stored in the cloud and pushed down to clients. - -## What this looks like in practice - -You finish work on your laptop at 7pm. The context from your day transfers to your phone over your home WiFi as you close the lid: the document you were editing, the email thread you were working through, the meeting notes from this afternoon. - -You pick up your phone at 8pm. The AI on your phone has your work context. When you decide to respond to a message that references the document you were working on, the AI has the context to help you. - -The following morning, you open your laptop. The AI on your laptop has context from your phone: you sent three messages last night, one of which started a new thread that needs a response, and your sleep data suggests you might want to protect the first hour of your day. - -None of this required a server. Nothing left your home network. - -## What happens when you're not on the same network - -The question everyone asks: what happens when you're traveling and your devices aren't on the same WiFi? - -Two answers. - -First, each device carries its own full context. Your phone has its context. Your laptop has its context. They are both useful independently. They don't become useless when they can't sync. - -Second, for users who want sync across networks, the right solution is a private tunnel (Tailscale, WireGuard, or similar) that connects your devices securely without routing through a third-party server. You run the infrastructure. You control the relay. The data stays yours. - -This is more setup than a cloud service. It is also the only architecture that keeps your full context under your control regardless of where you are. - -## The direction of the category - -Current personal AI products are designed around the cloud sync model because it was the only viable option when they were built. Local network sync requires both devices to run compatible software, which was difficult when personal AI was niche. - -As on-device AI becomes the default assumption for a growing number of products, the infrastructure for local sync becomes more practical to build and more expected by users. The category will move toward it for the same reason it will move toward privacy generally: users who understand the alternatives will prefer them. - -The Personal AI OS that gets this right closes the last gap between what personal computing can do and what it should do: context that flows between your devices privately, reliably, without a server in the middle. - ---- - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/end-of-app-switching.md b/website/writing/end-of-app-switching.md deleted file mode 100644 index b8b5356a..00000000 --- a/website/writing/end-of-app-switching.md +++ /dev/null @@ -1,69 +0,0 @@ ---- -layout: default -title: "The Personal AI OS and the End of App Switching" -parent: Perspectives -nav_order: 14 -description: You open 6 apps to coordinate one task. Calendar, email, Slack, notes, maps, messaging - all for one meeting. A Personal AI OS collapses that into one intelligence layer that orchestrates across them on your behalf. ---- - -# The Personal AI OS and the End of App Switching - -Count the apps you open to plan a single meeting. - -Calendar to check availability. Email to find the context. Slack to confirm the agenda. Notes to find what you discussed last time. Maps to check travel time. Messages to send the invite. - -Six apps, fifteen minutes, one meeting. And that's a simple case. - -This is the defining friction of modern knowledge work. Not any one task being hard, but the overhead of coordinating between apps that don't talk to each other, holding context in your head that should be held by software, and switching back and forth between tools that each know one fragment of the picture. - -A Personal AI OS is the layer that ends this. - -## Why apps can't solve it themselves - -The app model was the right solution to a real problem. Specialised tools are better than general ones. A calendar app built specifically for scheduling is better than a general-purpose productivity tool that does scheduling among many other things. - -But the app model has a structural limitation: apps are sandboxed. Your calendar doesn't know what's in your emails. Your notes app doesn't know your Slack messages. Your messaging app doesn't know your calendar. - -This isn't a bug in any individual app. It's a property of how platforms are designed. Apps compete on features, not on their ability to share context with each other. The incentive structure actively works against the coordination layer that users need. - -Integrations exist - Zapier, calendar plugins, Slack connectors - but they are point-to-point connections between specific apps, not a general intelligence layer. They automate individual workflows, not the judgment required to orchestrate across all of them. - -## What the intelligence layer does differently - -A Personal AI OS doesn't replace your apps. It sits above them and has context from all of them. - -When you ask it to help you prepare for a meeting, it already has access to your calendar entry, your email thread with that person, your previous notes, and your last Slack exchange. It doesn't need you to copy-paste context from each app. It already has it. - -When you ask it to find a time for a call, it knows your calendar, your energy patterns, and the priority of the meeting relative to what else is on your day. It suggests a time that actually makes sense for you, not just a time that's technically available. - -When you need to delegate a follow-up, it can draft the message, add it to your task list, and set a reminder - not as three separate actions in three apps, but as one thing. - -The intelligence layer is the coordination that each individual app was never designed to provide. - -## The multi-app tax - -Knowledge workers pay a multi-app tax every day. It takes the form of: - -**Context switching overhead.** Every time you move between apps, you lose a few seconds to mental reorientation. Across dozens of switches a day, this adds up to significant time and, more importantly, significant cognitive load. - -**Duplicated information.** The same piece of information - a meeting time, a contact's last message, a document name - lives in multiple apps in slightly different forms. Keeping them consistent is work you're doing manually. - -**Missed connections.** The email with the context for the meeting and the calendar invite for the meeting are in different apps. Your brain has to hold the connection. Sometimes it doesn't, and you arrive at a meeting unprepared. - -**Tool selection overhead.** "Should I put this in notes or tasks? Should I send this as a message or an email?" These decisions consume attention that shouldn't have to be spent on them. - -A Personal AI OS reduces all four. Not by eliminating apps, but by providing an intelligence layer that manages the coordination between them. - -## The first step - -The full vision - an AI layer that coordinates across all your apps in real time - requires deep platform integration that takes time to build. - -The first step is on-device AI that has the context of your phone and responds to natural language. Instead of opening six apps, you ask one question and get an answer that synthesised all six. - -"What do I need to do before my 2pm call?" The AI already knows. It tells you, and it's right, because it has the same context you would have assembled in fifteen minutes of app-switching. - -That's the first form of the end of app-switching. Not the last form, but a meaningful one. - ---- - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/how-personal-ai-should-act.md b/website/writing/how-personal-ai-should-act.md deleted file mode 100644 index cd439139..00000000 --- a/website/writing/how-personal-ai-should-act.md +++ /dev/null @@ -1,83 +0,0 @@ ---- -layout: default -title: "How a Personal AI OS Should Act on Your Behalf - Without Becoming Your Boss" -parent: Perspectives -nav_order: 10 -description: Proactive AI assistance is useful. But the line between helpful and creepy is thin, and crossing it produces a system you stop trusting. Consent is the operating principle. ---- - -# How a Personal AI OS Should Act on Your Behalf - Without Becoming Your Boss - -There is a version of personal AI that is useful: one that handles the low-value work you repeat every day and gives you back the mental space it was consuming. And there is a version that is suffocating, one that takes over decisions you wanted to make yourself, surfaces information you didn't want surfaced, and makes you feel monitored rather than helped. - -The difference is not capability. It's design philosophy. - ---- - -## The useful version - -A message comes in at 9pm from a contact you don't recognise. Your Personal AI OS has your phone in Do Not Disturb. It checks the message, classifies it as non-urgent, and defers it to your morning queue. You don't see it until you're ready for it. - -A meeting gets added to your calendar for 2pm tomorrow. The AI notices you have a conflicting commitment at the same time, flags it, and asks if you'd like to resolve it. - -You're about to join a call. The AI pulls the last three conversations you had with that person, summarises the open items, and puts them in front of you two minutes before the call starts. - -These actions are useful because they happen within a clear boundary: the AI is handling things you would have handled the same way, at moments when your attention was elsewhere, using judgment you've already expressed. It's not deciding things for you. It's executing decisions you would have made yourself if you'd had the bandwidth. - ---- - -## The line - -The line between helpful and creepy is consent and transparency. - -An AI that defers a notification is helpful if you set up that rule and know it's happening. An AI that starts suppressing notifications on its own judgement, even if that judgement is usually right, is one that's making decisions about your information diet without your input. - -An AI that suggests a reply to a message is helpful if you asked for it. An AI that learns your communication style and generates pre-written replies for you to approve is useful. An AI that starts sending messages without showing them to you first is something else entirely. - -The pattern is simple: the AI should make it easier for you to do what you would have done, not take over doing it for you without your awareness. - ---- - -## The WhatsApp-to-calendar example - -You get a WhatsApp message from a friend: "Dinner Friday at 7?" - -A helpful Personal AI OS surfaces this with a single action available: "Add to calendar." One tap, done. The AI saw the intent in the message, matched it to your calendar, and prepared the action. You approved it. Ten seconds of your attention instead of sixty. - -A creepy version of the same feature adds the dinner to your calendar automatically, "because you usually accept dinner invitations from this contact." Statistically correct. Behaviorally wrong. Your calendar now has an event you didn't add, and you have a new anxiety: what else has the AI decided on your behalf? - -The capability is identical. The design is not. - ---- - -## Proactive vs autonomous - -There is a meaningful difference between proactive assistance and autonomous action. - -Proactive assistance means the AI notices things, surfaces them, and makes the next action easy. It watches your calendar and tells you about conflicts. It reads your messages and highlights the ones that need a response today. It notices you've been in back-to-back meetings for four hours and surfaces the break you have in 20 minutes. - -Autonomous action means the AI takes the action without asking. It resolves the calendar conflict by declining one invite. It responds to messages. It rearranges your day. - -Proactive is good. Autonomous requires explicit delegation: tasks where you've clearly said "handle this without asking me." - -The default should be proactive. Autonomy should be the exception, granted task by task, with full transparency about what the AI is doing and the ability to review its actions. - ---- - -## What good defaults look like - -A Personal AI OS with good defaults: - -- Surfaces notifications and flags urgency, but doesn't suppress messages without your explicit Do Not Disturb rules -- Suggests replies but doesn't send them -- Notices conflicts and asks how to resolve them, rather than resolving them -- Prepares context before meetings rather than summarising after without being asked -- Tells you what it's doing when it takes action in the background - -The goal is to be the assistant who handles the work you'd delegate to a smart person who knows your priorities. Not the one who starts making calls on your behalf before you've decided you trust them that much. - -Trust is earned incrementally. A Personal AI OS should behave the same way. - ---- - -*Off Grid acts on your behalf with your explicit direction. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/index.md b/website/writing/index.md deleted file mode 100644 index b5f39835..00000000 --- a/website/writing/index.md +++ /dev/null @@ -1,176 +0,0 @@ ---- -layout: default -title: Perspectives -nav_order: 6 -has_children: true -description: Essays on the future of personal AI, on-device intelligence, and why the most important software of the next decade runs on hardware you already own. ---- - -# Perspectives - -Essays on where personal AI is going, how it should work, and why it matters. - ---- - -## Defining the category - -
- -
What Is a Personal AI OS?
-
The definitive explainer. What the term means, why it's a new category, and the 7 principles that separate it from everything else called AI.
- - -
The 7 Principles of a Personal AI OS
-
On-device. No telemetry. Persistent context. Acts on your behalf. Cross-device. Open and auditable. No cloud compute rent. The rules that define the category.
- - -
Personal AI OS vs AI Assistant vs AI Agent
-
Voice assistants answer questions. Cloud chatbots generate text. Autonomous agents take actions. A Personal AI OS does something different from all three.
- - -
Why Your Personal AI Should Never Live in the Cloud
-
Not a privacy rant - a structural argument. Cloud-dependent personal AI is broken by design, independent of whether the company behind it is trustworthy.
- -
- ---- - -## How it should work - -
- -
What a Personal AI OS Should Know About You - And What It Shouldn't
-
The right level of context makes it useful. The wrong level creates a system you don't want near your life. Here's where the line is.
- - -
How a Personal AI OS Should Act on Your Behalf - Without Becoming Your Boss
-
Proactive assistance is useful. Autonomous action without consent is not. The line between helpful and creepy, and consent as the operating principle.
- - -
The Architecture of Trust
-
Policy is words. Architecture is structure. On-device, open source, no telemetry - the only combination that earns the right to your full context.
- - -
Cross-Device Sync Without a Server
-
Your laptop context on your phone. Your phone context on your laptop. All over your local network, with no cloud relay.
- -
- ---- - -## How it embeds in life - -
- -
A Day With a Personal AI OS
-
From morning to night - a narrative walkthrough of what your day looks like when your devices share context and handle the low-value work themselves.
- - -
The Personal AI OS and the End of App Switching
-
You open 6 apps to coordinate one task. A Personal AI OS collapses that into one intelligence layer that orchestrates across them.
- - -
Why Your Phone Is the Most Important Device
-
80+ unlocks a day. Messages, location, health, camera. No device owns more of your context - no device matters more for local AI.
- - -
The Personal AI OS for Knowledge Workers
-
Email triage, meeting prep, deep work protection, status updates. What reclaiming 2+ hours a day actually looks like.
- - -
It's Not About Productivity. It's About the 35 Tabs.
-
Hiring a secretary doesn't 5x your business. It just means life is easier. We spend 90% of our time on digital devices and almost none of that time is actually easy.
- -
- ---- - -## Why now - -
- -
The Encrypted Messaging Moment for AI
-
Encrypted messaging went mainstream because the market demanded it. AI is next. The arc is the same.
- - -
The Walled Garden Problem
-
Platform AI is capable. But the architecture makes a genuine Personal AI OS impossible from within it. Why openness is not optional for the category.
- - -
The Regulatory Case for On-Device AI
-
Every major privacy regulation passed in the last five years is a tailwind for on-device AI. The architecture that's right for users ages well with regulators.
- -
- ---- - -## The democratization of intelligence - -
- -
The 200-Year Secretary
-
For two centuries, having a personal secretary was the defining advantage of wealth and power. A Personal AI OS running on the device in your pocket changes that equation permanently.
- - -
Your Next Virtual Assistant Won't Be a Person
-
The VA model built a billion-dollar industry on human judgment at low cost. On-device AI undercuts that model - and delivers something human VAs structurally cannot: your full context, always available, private by architecture.
- - -
Why the VA Industry Is About to Be Disrupted by On-Device AI
-
The VA market is billions and growing. On-device AI competes on a dimension the human model can't match: full personal context, available at any hour, private by design.
- -
- ---- - -## The philosophical layer - -
- -
Intelligence Should Be Personal. Here's What That Actually Means.
-
Not personalised in the engagement-optimisation sense. Personal in the sense that it's yours - on your hardware, under your control, with no corporation in between.
- - -
Privacy Is Not a Feature. It's an Architecture Decision.
-
Toggles, deletion tools, and privacy policies are theater. The only real answer is an architecture where the data never left your device.
- - -
The Case Against AI Subscriptions
-
You don't pay monthly for your calculator. Intelligence should be a one-time investment, not a service you rent from someone who can change the terms.
- - -
Who Owns Your AI's Memory?
-
When your AI remembers years of your context, who owns that memory? The most important digital rights question of the decade - and almost nobody is asking it.
- -
- ---- - -## The context gap - -
- -
The Context Gap
-
Your phone knows everything about you. Your AI does almost nothing with it. The chasm between what your devices know and what they do with it.
- - -
You Are One Person Across Two Devices. Your AI Should Know That.
-
Your phone sees your life. Your laptop sees your work. Neither talks to the other. That's the biggest unsolved problem in personal computing.
- - -
Why Platform Intelligence Doesn't Exist Yet
-
Mobile platforms are app-centric operating systems. The AI bolted onto that model can't be a true intelligence layer. Here's what it would take to build one.
- - -
Your Phone and Laptop Know Nothing About You
-
You unlock your phone 80+ times a day. You're on your laptop 8+ hours. At the platform level, neither can answer "what's this person's day been like?"
- - -
Two Devices, Zero Shared Context
-
Your laptop sees your work. Your phone sees your life. Neither talks to the other. This is the problem the Personal AI OS was built to solve.
- -
- ---- - -*Mohammed Ali Chherawalla is the creator of Off Grid. New essays go to [dev.to/alichherawalla](https://dev.to/alichherawalla) first.* diff --git a/website/writing/intelligence-should-be-personal.md b/website/writing/intelligence-should-be-personal.md deleted file mode 100644 index a85ad126..00000000 --- a/website/writing/intelligence-should-be-personal.md +++ /dev/null @@ -1,69 +0,0 @@ ---- -layout: default -title: "Intelligence Should Be Personal. Here's What That Actually Means." -parent: Perspectives -nav_order: 18 -description: Intelligence - the capacity to understand, reason, and act - has always been deeply personal. When we talk about AI being personal, we mean something more specific than it just being useful to individuals. Here's what it actually means. ---- - -# Intelligence Should Be Personal. Here's What That Actually Means. - -The word "personal" is doing a lot of work in conversations about AI. - -Personal AI. Personal assistant. Personalised experience. They mean different things and the differences matter. - -Personalised experience means the system adjusts its outputs based on your behaviour. It shows you content you're more likely to engage with. It surfaces products similar to ones you've bought. It's about optimisation for engagement, not about the AI actually knowing you. - -Personal assistant means a system that responds to your requests and helps you complete tasks. It's reactive. You prompt it and it helps. The relationship is transactional. - -Personal AI OS means something more fundamental - intelligence that is yours in the same way your thoughts are yours. That lives on hardware you own. That no corporation controls. That doesn't become inaccessible because a company was acquired or a service was discontinued. That you can trust with your full context because the architecture makes it safe to do so. - -## What it means for intelligence to be personal - -Genuine personal intelligence has three properties that distinguish it from the AI products that claim to be personal. - -It knows you specifically. Not a user profile built from aggregate behavioural data. Not a personalisation layer on top of a general model. A working understanding of your patterns, priorities, and context, built from the data of your life - your messages, your calendar, your work, your health. - -This understanding lives on your device, built from sources you've explicitly shared, and updated continuously as your life changes. It's not a static snapshot. It's a living model of who you are. - -It works for you, not the system. A system optimised for engagement is designed to keep you in it. A system optimised for you is designed to reduce the time you spend in it - to handle things quickly so you can move on, to surface what matters so you can focus on it, to make friction disappear so your day flows. - -These goals are in tension. A system that makes your email take three minutes instead of two hours is a worse product by engagement metrics and a better product by outcomes. Personal intelligence optimises for outcomes. - -You own it. The model is on your hardware. The context is in your storage. If the company that built the software disappears, your intelligence layer persists. You can run it, extend it, replace the underlying model, move it to a new device. It is an asset you own, not a service you rent. - -## Why this matters beyond privacy - -The privacy argument for personal AI is real and important. But the case for intelligence being personal extends beyond it. - -There is a broader principle about human capability and autonomy at stake. Intelligence has historically been something you develop - through education, experience, reflection. The models of intelligence around us - advisors, teachers, mentors - were people who had genuine knowledge of our situation and acted in our interests. - -The AI infrastructure being built today is mostly intelligence-as-a-service: capability you access via a network, at a price, under terms set by someone else. The capability is real. The dependency it creates is also real. - -A Personal AI OS is a different model. It's intelligence you own and carry, that becomes more useful over time as it learns more about you, that works for you without any ongoing relationship to a corporation. - -This is a different relationship between a person and their own capacity to understand and act. - -## The democratisation argument - -For most of human history, having intelligent, knowledgeable people in your corner was a function of wealth and access. A good lawyer who knew your situation. A doctor who was also a trusted friend. A financial advisor who understood your full picture. - -These relationships are valuable partly because the person is capable and partly because they know you. Generic advice from a capable person is less useful than specific advice from someone who understands your situation. - -AI has the capability to make contextualised intelligence available to everyone. Not a generic assistant that answers questions, but a system that knows your situation, understands your goals, and can reason about your specific circumstances. - -But this requires personal AI - AI that has your context and acts for you. It requires the architecture to support it: on-device, private, owned by the user. AI as a service owned by a corporation is not the democratisation of intelligence. It's access to capability, mediated by a subscription and subject to corporate decisions about availability and terms. - -The Personal AI OS is the model that delivers the democratisation argument. Intelligence that lives on the device you carry, available anywhere, with full knowledge of your context, under your control. - -## What we're building toward - -Off Grid starts with the AI capabilities that are ready today: language models running locally on your phone, offline, with no data leaving your device. - -That's a meaningful starting point. A capable AI available anywhere, with no cloud dependency, no subscription required to access the core capability. - -The direction is toward the fuller vision: persistent context, cross-device intelligence, integration with the apps and data sources that make up your working life, all of it on your hardware under your control. - -Personal in the full sense. - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing). [Join the community on Slack](https://join.slack.com/t/off-grid-mobile/shared_invite/zt-3w2utgk0w-EDiDZBq6KmSZZwEw5Tkhnw).* diff --git a/website/writing/next-virtual-assistant.md b/website/writing/next-virtual-assistant.md deleted file mode 100644 index 89c492c9..00000000 --- a/website/writing/next-virtual-assistant.md +++ /dev/null @@ -1,91 +0,0 @@ ---- -layout: default -title: "Your Next Virtual Assistant Won't Be a Person. And That's the Point." -parent: Perspectives -nav_order: 27 -description: The virtual assistant industry built a model on human judgment at low cost. On-device AI undercuts that model entirely and delivers something human VAs structurally cannot. ---- - -# Your Next Virtual Assistant Won't Be a Person. And That's the Point. - -The virtual assistant industry exists because of a simple insight: a lot of the work that consumes a knowledge worker's day doesn't actually require that specific knowledge worker. - -Email triage. Meeting scheduling. Research summaries. Follow-up messages. Calendar management. Travel coordination. Document formatting. Status updates. The administrative surface area of modern professional life is enormous, and most of it is templated, repetitive, and executable by someone with context and good judgment, regardless of whether they have the specific expertise of the person they're assisting. - -Human virtual assistants filled that gap. Remote workers who could handle the coordination overhead so that their clients could focus on the work that required their actual expertise. The model worked. The industry grew to billions in annual spend. - -But the model has a structural ceiling. On-device AI is what breaks through it. - ---- - -## What human VAs do well - -Human virtual assistants are genuinely good at several things that matter. - -They understand nuance. A human VA who has worked with you for six months understands your communication style, your priorities, your pet peeves, and your implicit preferences in ways that are hard to specify in advance and easier to observe over time. - -They can make judgment calls. When an edge case comes up that doesn't fit the instructions, a good VA uses judgment. They know when to act and when to ask. - -They handle ambiguity. The world of professional communication is full of things that require reading context, not just executing instructions. A human VA can tell when a message is more loaded than it appears. - -These are real capabilities. They're also increasingly replicable by a system that has more context than any human assistant can have. In some ways, surpassable. - ---- - -## What human VAs can't do - -Human virtual assistants have structural limits that no amount of skill or dedication can overcome. - -**They don't have your full context.** A human VA sees what you share with them. They don't see your calendar, your messages, your files, your health data, your location, and your work patterns simultaneously. The context that would make the intelligence layer most useful is also the context that's hardest to hand to another person. - -**They work business hours.** A VA in a different time zone can extend coverage, but nobody is available at 11pm when you need to prepare for an 8am meeting and want to know what the open items were from the last discussion with that client. - -**They cost ongoing money.** A competent VA is not cheap. A skilled EA-level VA less so. The model prices many of the people who could most benefit from administrative support out of the market. - -**They require trust and coordination overhead.** Working with a human VA requires explaining context, reviewing output, managing the relationship, and handling the inevitable edge cases where communication breaks down. This overhead is real and recurring. - -**They scale linearly.** One VA can handle a bounded amount of work. When your administrative surface area grows, the cost grows with it. - -None of these are criticisms of human VAs. They are properties of any system where the intelligence layer is a person with finite time, bounded access to your context, and a cost structure tied to human labour. - ---- - -## What on-device AI delivers differently - -A Personal AI OS running locally on your device changes the calculus on every one of these dimensions. - -It has your full context. Your messages, your calendar, your files, your patterns. All of it, all at once, all the time. The intelligence it can apply to your inbox or your upcoming meeting is informed by everything you have, not just what you've chosen to share. - -It's available at any hour. There's no time zone, no business hours, no response delay. When you're prepping for a morning meeting the night before, the context is there. - -It has no ongoing cost tied to your usage. The model runs on your device. The marginal cost of the hundredth email triage is the same as the first. - -It requires no relationship management. The context is built from your data, not from a working relationship that needs tending. The system knows you from what you actually do, not from what you've explained. - -It scales with your needs. More emails, more meetings, more complexity. The system handles it without renegotiating terms. - ---- - -## What it still doesn't replace - -On-device AI is not a complete replacement for everything a skilled human assistant does. - -Human judgment in genuinely novel situations, where the right move isn't derivable from past patterns, is still a human edge. Complex relationship management that requires emotional intelligence and interpersonal calibration is still a human capability. Tasks that require physical presence or real-world interaction are still human territory. - -But the vast majority of what makes administrative support valuable is not in those categories. Most of it is pattern recognition applied to communication, scheduling, and coordination. Exactly the domain where a system with full context and no time constraints outperforms a person with bounded access and a limited workday. - ---- - -## Who this actually helps - -The human VA model helped people who could afford it. Typically knowledge workers at senior levels, entrepreneurs with enough revenue to justify the cost, executives at organisations that provided support as a benefit. - -The people below that threshold had just as much administrative overhead but without the revenue or seniority to justify dedicated support. They managed it themselves, which meant it consumed the same focused attention they needed for the work that actually required them. - -On-device AI doesn't just improve on the VA model for existing customers. It makes the function available to people the model never reached in the first place. - -That's the more interesting story. Not that a faster or cheaper virtual assistant exists. But that the intelligence layer that made executives more effective for a century is now in the pocket of anyone who wants it. - ---- - -*Off Grid is building toward this. It starts with on-device AI that works fully offline on your phone, the foundation that makes everything above possible without your data ever leaving your device. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/one-person-two-devices.md b/website/writing/one-person-two-devices.md deleted file mode 100644 index 58ccfb07..00000000 --- a/website/writing/one-person-two-devices.md +++ /dev/null @@ -1,96 +0,0 @@ ---- -layout: default -title: "You Are One Person Across Two Devices. Your AI Should Know That." -parent: Perspectives -nav_order: 3 -description: Your phone sees your life. Your laptop sees your work. Neither talks to the other. That's the biggest unsolved problem in personal computing - and the one a Personal AI OS was built to fix. ---- - -# You Are One Person Across Two Devices. Your AI Should Know That. - -You unlock your phone more than 80 times a day. You spend 8 hours on your laptop. Between those two devices, almost everything about your working life is recorded. - -And yet if you ask either device "what's my day been like?" the answer is nothing. The data exists - in fragments across a dozen apps - but no intelligence layer has ever tried to hold it together. - -That is the context gap. It is the biggest unsolved problem in personal computing. - ---- - -## What your phone knows - -Your phone is with you 16 hours a day. In that time it collects an extraordinary amount of context: - -- Every message you send and receive -- Your location, continuously -- Your calendar - what is scheduled, what you accepted, declined, and rescheduled -- Your health data - sleep, activity, heart rate -- Your camera roll - what you photographed and when -- The apps you open, in what order, for how long - -This is a detailed record of your life. The phone has all of it. The platform does almost nothing with it. - -The built-in assistant can set a timer or call a contact. It cannot tell you that you have had three difficult conversations this week and your calendar tomorrow is unrealistic given how your Monday went. - ---- - -## What your laptop knows - -Your laptop sees something your phone does not: your work. - -The documents you are writing. The tabs you have open. The emails you are drafting. The code you are reviewing. The meetings you are preparing for. - -It knows your professional context with a depth your phone never will - because that is where work actually happens for most knowledge workers. - -But it knows nothing about the rest of your life. It does not know you were up at 2am. It does not know your flight got cancelled. It does not know you have been in back-to-back calls since 8am and have nothing left. - ---- - -## The gap between them - -You are one person. Your phone and laptop are used by the same human, with the same goals, facing the same constraints on the same day. - -But they have never talked to each other. Not at the intelligence layer. - -App-level sync exists - your calendar is on both devices, your messages are on both devices. That is data replication, not intelligence. Shared data does not mean shared understanding. - -A cloud AI could theoretically bridge this gap - if you were willing to give it access to your phone's messages, your laptop's files, your health data, your calendar, your location history. Some products ask for exactly that. The cost is handing your most personal context to infrastructure you do not control. - -There is a better architecture. - ---- - -## How a Personal AI OS bridges the gap - -A Personal AI OS holds context across both devices - locally, over your home network, without a cloud relay. - -Your phone builds context from your messages, health data, calendar, and location. Your laptop builds context from your files, email, and work patterns. The Personal AI OS merges these into a single working model of your day, your week, your current priorities. - -When you ask a question on either device, the answer draws on both. Your phone knows you are exhausted. Your laptop knows your deadline moved. The AI knows both. - -This is what makes the Personal AI OS a new category rather than a smarter assistant. It is not a better answer to "set a timer." It is the first system that actually knows who you are across the full span of your day. - ---- - -## Why this has not been built yet - -The obstacle is not hardware. Modern phones and laptops have enough compute to run capable local models. The obstacle is the assumption that built modern software platforms. - -Mobile platforms are app-centric operating systems. The primitive is the app, and apps are sandboxed from each other. Intelligence - to the extent the platforms attempt it - is bolt-on, not foundational. - -A Personal AI OS requires inverting that model. Context is the primitive. Apps are sources of context. The intelligence layer sits above the apps, not inside any one of them, and operates across all your devices as a single system. - -That architecture does not exist at the platform level. It has to be built as a layer on top - which is exactly what Off Grid does on the device side, and what the next generation of local AI software will build out fully. - ---- - -## What it means in practice - -You wake up. Your Personal AI OS knows you slept poorly, your first meeting starts in 40 minutes, and you have three unread messages that probably require a response before then. - -By the time you open your laptop, the context is already there. It did not sync through a server. It moved over your local network. Nothing left your home. - -That is what it looks like when your devices actually know you. - ---- - -*Off Grid is building toward this. Start with the phone - the most context-rich device you own. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/personal-ai-os-for-knowledge-workers.md b/website/writing/personal-ai-os-for-knowledge-workers.md deleted file mode 100644 index a0008307..00000000 --- a/website/writing/personal-ai-os-for-knowledge-workers.md +++ /dev/null @@ -1,85 +0,0 @@ ---- -layout: default -title: "The Personal AI OS for Knowledge Workers: From Email Triage to Meeting Prep to Deep Work" -parent: Perspectives -nav_order: 15 -description: 800 million knowledge workers spend large parts of their day on work that AI should handle. Email triage, meeting prep, status updates, follow-up drafting. Here's what that looks like when the AI runs locally on your own device. ---- - -# The Personal AI OS for Knowledge Workers: From Email Triage to Meeting Prep to Deep Work - -There are roughly 800 million knowledge workers in the world. Each of them spends a significant portion of their working day on tasks that require judgment but not their judgment specifically. A system with the right context could handle these as well as or better than they can. - -Email triage. Meeting preparation. Status updates. Follow-up drafts. Summary notes after calls. Finding the document you know you wrote three weeks ago. Checking whether a commitment you made has been fulfilled. - -These tasks are not trivial. They require understanding your priorities, your communication style, your work context. But they don't require your creative output or your domain expertise. They are infrastructure work, and they are consuming an enormous amount of the most expensive resource in a knowledge worker's day: focused attention. - -A Personal AI OS that runs locally, with access to your full context, is the first system capable of handling this work reliably. This is where the category is going. Some of it is close. None of the current cloud tools can do it the right way, because doing it the right way requires your data to stay on your device. - -## Email triage - -The average knowledge worker spends over two hours a day on email. The vast majority of that time is triage: reading enough of each message to decide whether it needs a response, when, and what kind. - -A Personal AI OS with access to your email history and your patterns can do this triage automatically. - -It knows which senders you respond to within the hour and which can wait until end of day. It knows which threads are active work and which are informational. It knows that you typically handle client communications in the morning and internal operations in the afternoon. - -It surfaces your email not as a chronological flood but as a prioritised queue: here's what needs a response today, here's what needs a response this week, here's what you can archive. - -You spend 20 minutes on email instead of two hours, and you make fewer mistakes about what's urgent because the system is tracking signal you'd otherwise miss. - -## Meeting preparation - -Knowledge workers average 10-12 hours of meetings per week. A significant fraction of those meetings are ones where attendees arrive underprepared. - -Not because the preparation would have been hard. Because the preparation required finding context from four different places: previous meeting notes, the relevant email thread, the document shared last time, the last Slack exchange with this person. Nobody had the 15 minutes to do it. - -A Personal AI OS does this preparation automatically. - -Two minutes before your meeting starts, it surfaces: the last three things you discussed with this person or group, the open items from the last meeting, any relevant documents that have been shared, and anything from recent messages that's relevant to the agenda. - -You arrive prepared for every meeting, every time, with no additional effort on your part. The compounding effect over a week is significant: arriving prepared for 12 meetings instead of 4. - -## Deep work protection - -Deep work is fragile. It is the focused, uninterrupted time where knowledge workers produce their highest-value output. A single interruption breaks concentration that takes 20 minutes to rebuild. - -A Personal AI OS that manages your notifications intelligently can protect deep work in a way that static Do Not Disturb settings cannot. - -It knows you're in a focused session. It reads incoming notifications and classifies them by your definition of urgency, which it has built from observing your responses over months. It surfaces urgent things immediately. Everything else waits. - -When your focused session ends, it presents a consolidated view of what came in, already prioritised. You haven't missed anything important. You also haven't been interrupted six times by things that could have waited. - -## Status updates and follow-ups - -A significant fraction of knowledge worker communication is status and coordination: "Just wanted to follow up on X," "Quick update on Y," "Checking whether Z has been resolved." - -These messages are necessary. They are also templated, repetitive, and draining to write. By the fifteenth follow-up email of the week, the effort required is disproportionate to the value of the message. - -A Personal AI OS that knows your communication style and your open commitments can draft these automatically. You review and send. You don't write them from scratch. - -Over a week, this compounds into hours of writing time reclaimed. Not writing that required your creativity, but writing that required your attention to exist. - -## Finding things - -Knowledge workers spend an average of 20% of their time searching for information they already have. Documents, emails, notes, messages: the context is there, but the retrieval is manual and slow. - -A Personal AI OS with access to your files and communications can answer natural language queries against your own data. - -"Find the email where we agreed on the Q3 scope." "What were the open items from the design review last month?" "Where did I put the contract template I used in March?" - -These queries return specific answers in seconds instead of requiring you to remember which app the information is in, what the subject line was, or approximately when it happened. - -The information you already have becomes as accessible as information you can look up. - -## The compound effect - -Each of these improvements (triage, preparation, protection, drafting, retrieval) is meaningful on its own. Together, they compound. - -A knowledge worker using a Personal AI OS that handles this infrastructure work doesn't just save hours per week. They change the quality of how they work. They arrive prepared. They respond faster. They protect their focused time. They don't drop things. - -The output is not just the same work in less time. It is better work, done with less friction, with more attention available for the things that actually require it. - ---- - -*Off Grid is building toward this. It starts with on-device AI that works fully offline on your phone: the foundation that makes everything above possible without your data ever leaving your device. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/personal-ai-os-vs-assistant-vs-agent.md b/website/writing/personal-ai-os-vs-assistant-vs-agent.md deleted file mode 100644 index eded6516..00000000 --- a/website/writing/personal-ai-os-vs-assistant-vs-agent.md +++ /dev/null @@ -1,98 +0,0 @@ ---- -layout: default -title: "Personal AI OS vs AI Assistant vs AI Agent: What's the Difference and Why It Matters" -parent: Perspectives -nav_order: 7 -description: Voice assistants answer questions. Cloud chatbots generate text. Autonomous agents take actions. A Personal AI OS does something different from all three - and the distinction is worth understanding precisely. -faq: - - q: What is the difference between a Personal AI OS and an AI assistant? - a: Platform voice assistants answer isolated questions using cloud infrastructure. They have minimal persistent context and no ability to act across your apps. A Personal AI OS runs on-device, maintains persistent context about your life, and can act on your behalf - it's a system, not a query interface. - - q: What is the difference between a Personal AI OS and an AI agent? - a: AI agents are autonomous systems that make decisions and take actions with minimal human oversight, often connected to external APIs and services. A Personal AI OS is explicitly not autonomous - it acts with your consent, stays within your local hardware, and defers decisions to you. The operating principle is assistance, not autonomy. - - q: What is the difference between a Personal AI OS and a cloud chatbot? - a: Cloud generative AI products have no persistent knowledge of you between sessions, run on remote servers, and are general-purpose text interfaces. A Personal AI OS is specific to you, runs locally, maintains your context over time, and is designed to act on your behalf across your apps and devices. ---- - -# Personal AI OS vs AI Assistant vs AI Agent: What's the Difference and Why It Matters - -The word "AI" is doing a lot of work right now. It describes voice assistants that set timers, chatbots that write emails, autonomous systems that browse the internet, and personal software that runs entirely on your phone. - -These are not the same thing. The differences matter - for what you can trust with your data, what you can expect from each, and which one is actually useful for your life. - ---- - -## AI Assistants: the query interface - -The voice AI assistants built into major platform operating systems were the first consumer AI products. They share a common architecture and a common set of limitations. - -**How they work:** You issue a voice or text command. The query goes to a cloud server. The server processes it and returns a response. The assistant executes a narrow set of device actions (set timer, play music, call contact) based on pre-defined integrations. - -**What they know about you:** Very little persistent context. They may access your calendar or contacts for specific queries, but they don't build a model of your patterns, priorities, or work style. - -**What they can do:** Answer factual questions, set reminders, control smart home devices, play media. They operate within sandboxed integrations and cannot act across your apps. - -**The privacy model:** Cloud-dependent. Your queries are processed on remote servers. Voice data is sent to infrastructure you don't control. - -AI assistants are useful for simple, isolated tasks. They are not intelligence layers. They have no continuous model of who you are. - ---- - -## Generative AI products: the capable chatbot - -Cloud generative AI products are a step up in capability but share a similar architecture to assistants in the ways that matter most. - -**How they work:** You send messages to a cloud-hosted model. The model generates responses. Context exists within a session but typically doesn't persist across sessions in a way that builds a long-term model of you. - -**What they know about you:** What you tell them in the current conversation. Some products offer memory features that persist selected information, but this is a managed exception rather than a continuous context layer. - -**What they can do:** Generate, summarise, analyse, and discuss. Recent versions have tool use and browsing capabilities. They are powerful at tasks that don't require knowing you specifically. - -**The privacy model:** Cloud-dependent. Your conversations are sent to external servers. Your data may be used for model training depending on product settings. - -Generative AI products are powerful general tools. They are not personal. The more personal the task, the less suited they are - because they don't know you. - ---- - -## AI Agents: the autonomous system - -AI agents are the newest and most distinct category. They are systems designed to take sequences of actions toward a goal with minimal human guidance. - -**How they work:** You define an objective. The agent plans and executes a series of steps - browsing the web, writing and running code, calling APIs, sending emails - until the goal is reached or it encounters a blocker. - -**What they know about you:** Variable, depending on what context the agent is given at the start of a task. Most current agents have limited persistent knowledge of the user. - -**What they can do:** Sequences of actions across external services. Research, code execution, web interaction, communication. Capable of completing complex multi-step tasks without human involvement at each step. - -**The privacy model:** Typically cloud-dependent and highly permissive - an autonomous agent needs broad access to external services to do its job. This creates significant surface area for data exposure. - -AI agents are powerful for specific, bounded tasks where you want automation. They are not personal assistants. They are task executors. - ---- - -## Personal AI OS: the intelligence layer - -A Personal AI OS shares surface similarities with all three - it responds to queries like an assistant, generates text like a chatbot, and takes actions like an agent - but the architecture and purpose are fundamentally different. - -**How it works:** Inference runs on your device. Context is built and stored locally - your messages, calendar, files, health data, location patterns. The system maintains a continuous model of your life and work, accessible across your devices over a local network. - -**What it knows about you:** Everything you allow it to access, persistently. What you say in a session and what it has learned about your patterns over time. This is the defining property - the AI knows you specifically. - -**What it can do:** Everything the assistant and chatbot categories can do, plus context-aware actions that require knowing you: triaging your inbox in your priority system, preparing you for a meeting based on the history you have with that person, noticing that you're overcommitted next week before you've noticed it yourself. - -**The privacy model:** On-device. Nothing leaves your hardware. The context that makes it useful - the data that would be most valuable to an external party - never becomes available to one. - ---- - -## Why the distinction matters - -The difference is not capability. Current AI assistants are capable. Generative AI products are very capable. Agents are capable of things no prior software could do. - -The difference is architecture. Architecture determines trust. - -A system that runs on your device with your context, acting with your consent, is one you can give your full context to. A system that routes your data through a server is one where the privacy model is determined by policy rather than by design. - -The most useful AI for your life requires your full context - your messages, your health, your finances, your relationships. You should only give that context to a system whose architecture earns it. The Personal AI OS is that system. - ---- - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/phone-is-the-most-important-device.md b/website/writing/phone-is-the-most-important-device.md deleted file mode 100644 index 08e86f70..00000000 --- a/website/writing/phone-is-the-most-important-device.md +++ /dev/null @@ -1,67 +0,0 @@ ---- -layout: default -title: Why Your Phone Is the Most Important Device in the Personal AI OS -parent: Perspectives -nav_order: 5 -description: You unlock your phone 80+ times a day. It has your messages, location, health data, and camera. No device owns more of your context - which means no device matters more for local AI. ---- - -# Why Your Phone Is the Most Important Device in the Personal AI OS - -If you were designing the ideal hardware platform for a personal AI - one that knows you, stays with you, and has the data to be useful - you would describe a device that is always on your person, has continuous access to your location, captures your communications, monitors your health, and carries a high-resolution camera with immediate access to your visual environment. - -That device exists. You already own it. - -## The context argument - -Context is what separates a useful AI from a generic one. Any cloud model can summarise a document or answer a general question. What makes an AI useful to you, specifically, is knowing your patterns - how you work, what you're dealing with, what you need before you ask for it. - -Your phone has more of that context than any other device you own. - -It has every message you've sent and received across the apps you use daily. It has your calendar - the events and the pattern of your week, the rhythm of your meetings, the time you typically go quiet in the evenings. It has your location history, which tells a story about your life that no other data source replicates. It has your health data - sleep, activity, heart rate trends. - -No laptop has this. No tablet. No wearable alone. The phone is the only device that is with you, awake, for essentially your entire conscious day. - -## The hardware argument - -Modern flagship phones are not general-purpose internet appliances with a camera bolted on. They are powerful neural processing platforms that happen to also make calls. - -The Apple A18 Pro has a 16-core Neural Engine capable of 35 TOPS (trillion operations per second). Snapdragon 8 Gen 3 has a dedicated Hexagon NPU with similar throughput. These chips were designed for machine learning inference. They are the reason a current iPhone or flagship Android can run a capable language model - Qwen 3.5, Phi-4, Gemma 4 - at 20-30 tokens per second in real time, offline. - -This is new. Two years ago, the models that run fluidly on today's phones would have required a discrete GPU. The hardware jumped. The software hasn't fully caught up yet - most AI products still route everything through a server because that was the only option when they were designed, and changing architecture is hard. - -The technical constraint that made cloud AI necessary has been removed. The infrastructure for on-device intelligence is already in your pocket. - -## The privacy argument - -The context that makes a phone-based AI powerful is also the context you most need to protect. - -Your messages are your most private communication. Your health data reflects your physical reality in a way that has implications for insurance, employment, and relationships. Your location history is a map of your life - where you sleep, who you see, what you do. - -Handing this context to a cloud service in exchange for AI capabilities is the trade most AI products implicitly ask for. It is a trade with permanent consequences: once the data is on a server, you don't control what happens to it - not through deletion tools, not through privacy policies, not through account settings. - -The phone as the foundation of the Personal AI OS inverts this trade entirely. The model runs in your phone's memory. The context stays on your phone. The inference happens on your phone. Nothing is sent anywhere. The AI that knows the most about you is also the one that keeps everything local. - -## The mobile-first case - -The conventional wisdom in enterprise software is that you build for desktop first and mobile second. Desktop has more compute, more screen real estate, more input precision. Mobile is the simplified version. - -For a Personal AI OS, this logic is backwards. - -Desktop is where you do work. Mobile is where you live. The AI that knows your work can make you more productive in specific contexts. The AI that knows your life can reduce friction across everything. - -The phone is also the device you have when you need help in an uncontrolled environment - commuting, traveling, between meetings, in a situation you didn't anticipate. The desktop can only help you when you're at it. The phone is always there. - -And practically: the phone's sensor suite is unmatched. Camera, microphone, GPS, accelerometer, barometer. A Personal AI OS that can see what you see, hear what you hear, and know where you are has capabilities no laptop-centric system can match. - -## What this means for how the category develops - -The Personal AI OS will be built phone-first. Not because desktop doesn't matter - it does, and the cross-device context layer is part of the full vision - but because the phone is where the context lives, where the hardware is ready, and where the value is highest. - -The phone is the device that earns the most trust from users and asks for the most data in return. The AI on your phone, built on the right architecture, is the AI that deserves that trust. - -That's where Off Grid starts. - ---- - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/phone-laptop-know-nothing.md b/website/writing/phone-laptop-know-nothing.md deleted file mode 100644 index 2afec9ad..00000000 --- a/website/writing/phone-laptop-know-nothing.md +++ /dev/null @@ -1,87 +0,0 @@ ---- -layout: default -title: "Your Phone and Laptop Know Nothing About You. That's the Biggest Problem in Personal Computing." -parent: Perspectives -nav_order: 23 -description: You unlock your phone 80+ times a day. You're on your laptop 8+ hours. At the platform level, neither device can answer "what's this person's day been like?" That's not a small gap - it's the defining failure of personal computing. ---- - -# Your Phone and Laptop Know Nothing About You. That's the Biggest Problem in Personal Computing. - -Here is the absurdity at the centre of personal computing. - -You unlock your phone more than 80 times a day. Every unlock is a data point. You have been doing this for years. The device records your location every few minutes, logs every message you send and receive, tracks your sleep and your steps and your heart rate. It has a camera with your face in it. It has your banking app. It has your most private conversations. - -And if you ask it - or the AI built into it - "what's my day been like?" it cannot answer. The data exists. Nobody built the system to use it. - ---- - -## The data exists. The intelligence does not. - -The gap between what your devices know and what they do with it is almost total. - -Your phone has continuous location data going back years. It knows you go to the same coffee shop every Tuesday. It knows you have been in the office more days than usual this month. It knows you travelled somewhere three weeks ago and came back exhausted. It knows your sleep patterns changed around the same time a particular project started. - -Your phone's AI cannot tell you any of this. It can set a timer. - -Your laptop has the documents you have written for the past five years, the emails you have sent, the research you have done, the projects you have completed and abandoned. It knows more about your professional output than any person who has ever worked with you. - -Your laptop's AI can autocomplete a sentence if you are lucky. - -The data to make personal computing intelligent has existed for years. The intelligence layer has never been built. - ---- - -## Why this is the biggest problem - -It is easy to look at the current state of AI - capable models, useful products, genuine productivity gains - and conclude that the gap is closing. - -For general-purpose tasks it is. You can ask a cloud AI to summarise a document, write a draft, or explain a concept and get a useful response. - -But personal computing is specific - to you, your context, your day, your work, your relationships, your priorities. For those tasks, the current state is almost entirely broken. - -You manage your own calendar. You triage your own email. You remember your own commitments. You track your own open items. You hold in your head the context that connects your morning's work to your afternoon's meetings to the message you received at 9pm. - -This is cognitive overhead that software should be handling. The data to handle it is on the devices you carry. The intelligence to process it exists. The system that connects them has not been built. - ---- - -## The unlock problem - -The most concrete way to see the gap: every time you unlock your phone, you perform a context switch. You move from whatever you were doing to whatever the phone has waiting for you. - -A device that knew you would handle this context switch on your behalf. It would surface the things that need your attention and suppress the things that do not. It would know that the message from this contact is urgent and the notification from that app can wait. It would know that you are in the middle of focused work and the next 45 minutes should be protected. - -Instead, you perform that triage yourself, 80 times a day, with the same information the device already has but is not using. - -80 context switches. 80 manual triage decisions. Each one is a small tax on your attention that adds up across a day, a week, a year. - ---- - -## The morning case - -You wake up. You have eight hours of messages waiting - a combination of time zones, family, work, social. Some need your attention before anything else. Most do not. - -A device that knew you could have classified them overnight. By the time you look at your phone, the one urgent thing is at the top and the rest is waiting. - -Instead, you scan everything, hold the important things in working memory, and try to respond in the right order. By the time you have finished your morning messages, you have already spent 40 minutes and a significant amount of cognitive load on a task that was mostly pattern-matching against context your phone already had. - -This is the daily cost of the gap between what your devices know and what they do with it. - ---- - -## What would close it - -Three things, none of which require hardware that does not exist. - -An intelligence layer with access to the full context of your device - not sandboxed app by app, but a unified view of your messages, calendar, health, files, and location. - -A model capable of reasoning over that context - something a local model running on current hardware can do, today, for the types of tasks that matter. - -An architecture that keeps that context on your device, so the model runs in your phone's memory and nothing reaches external infrastructure. - -The problem is the absence of software built on the right assumptions. - ---- - -*Off Grid is building that software. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/platform-intelligence-doesnt-exist.md b/website/writing/platform-intelligence-doesnt-exist.md deleted file mode 100644 index 74126a06..00000000 --- a/website/writing/platform-intelligence-doesnt-exist.md +++ /dev/null @@ -1,83 +0,0 @@ ---- -layout: default -title: "Why Platform Intelligence Doesn't Exist Yet - And What It Would Take to Build It" -parent: Perspectives -nav_order: 22 -description: Mobile platforms are still app-centric operating systems. The AI features built into them are bolted onto that model. A true Personal AI OS requires a fundamentally different architecture where context is the primitive, not apps. ---- - -# Why Platform Intelligence Doesn't Exist Yet - And What It Would Take to Build It - -The major mobile platforms have shipped AI features. Notification summaries. Text generation. Image description. On-device models that handle some tasks without a network connection. - -These are real capabilities and meaningful engineering achievements. They are not, however, platform intelligence in the meaningful sense. They are AI features built on top of a platform architecture that was designed before personal AI was a consideration. - -The distinction matters because the architecture determines the ceiling. - ---- - -## The app-centric model - -Mobile platforms are app-centric operating systems. The fundamental unit of the platform is the app. Apps are: - -- **Sandboxed.** Each app has access only to the data it has been explicitly granted. Your calendar app cannot read your messages. Your AI assistant cannot, by default, access the files in your notes app. -- **Isolated.** Apps do not share state with each other except through explicit, narrow API integrations. The mental model is a collection of independent tools, not a unified system. -- **Managed by the platform.** The platform controls what each app can and cannot access, which capabilities are available, and how inter-app communication works. - -This model has real advantages for security and privacy. Sandboxing prevents malicious apps from reading your messages. Isolation prevents one app's bugs from affecting another. - -But it creates a fundamental limitation for personal AI: there is no coherent view of your context across the system. The AI assistant can see what each sandboxed permission grants - some calendar access here, some contacts there - but it cannot see the full picture. - ---- - -## What current platform AI actually is - -Current platform AI is built within the constraints of the existing app-centric model. - -It can summarise notifications because the notification system already exposes text from all apps in one place. It can generate text in keyboards because the keyboard already operates across apps at the system level. It can answer questions about the current document because it is running in the context of the document editor. - -Where the app model creates a unified view - notifications, keyboard, the document you are currently working on - platform AI can use that view. Where the app model creates fragmentation - the relationship between your messages and your calendar and your files - platform AI has the same limited view as any other app. - -The AI features are real. The intelligence layer is not. The platform AI does not have a working model of you. It has access to whatever the existing app permissions happen to expose at the moment of the query. - ---- - -## What actual platform intelligence would require - -A true platform intelligence layer would require different architecture from the ground up. - -**Context as the primitive.** Instead of apps that request permission to access specific data types, the platform would maintain a unified context layer - a continuously updated model of your life and work - that the AI can query with appropriate privacy controls. - -**Cross-app intelligence.** The ability to reason across data from multiple apps at once. To notice that the email thread from a contact is related to the calendar event tomorrow. To connect the document you are editing to the research in your browser history. To understand that the message that just arrived is about the project that has been in your task list for three weeks. - -**Persistent model of the user.** A session-by-session assistant is not enough. An ongoing model that learns your patterns, tracks your commitments, and builds understanding over time. - -None of this exists at the platform level today. Building it would require redesigning the fundamental architecture of the OS - the permission model, the inter-app data model, the privacy framework. - ---- - -## Why the platforms will not build it yet - -The platforms have the engineering capability to build platform intelligence. The reasons they have not go beyond capability. - -**Privacy and regulatory risk.** A system with the depth of context that true platform intelligence requires would face significant scrutiny. The same capabilities that make it useful - knowing your messages, health, files, and location at once - create regulatory exposure in jurisdictions with strong privacy frameworks. - -**Ecosystem conflict.** Many of the most valuable sources of personal context live in apps built by third parties. Building intelligence that spans mapping apps, messaging services, streaming platforms, and banking apps requires those apps to contribute context to a platform-level model. The companies behind those apps have no incentive to help the platform build a model that aggregates their users' data. - -**Openness.** True platform intelligence, to be trustworthy, needs to be auditable. The platforms are closed by design. A closed intelligence layer with access to your full context is one you have to trust on faith. - ---- - -## What the alternative looks like - -The alternative to platform intelligence is an independent intelligence layer that runs on your hardware, accesses data through the permissions you explicitly grant, and operates across platforms. - -It is not built into the OS. It runs on top of it. It has access to the data you give it - your messages, your calendar, your files - through the same permission mechanisms any app uses, but it aggregates and reasons across all of it rather than operating within one context. - -It is open, so you can verify what it does. It runs locally, so the context does not leave your device. It works across your devices, so the intelligence spans your phone and laptop. - -This is what a Personal AI OS is. A layer on top of the platform that provides what the platform architecture was never designed to. - ---- - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/privacy-is-not-a-feature.md b/website/writing/privacy-is-not-a-feature.md deleted file mode 100644 index e2896318..00000000 --- a/website/writing/privacy-is-not-a-feature.md +++ /dev/null @@ -1,70 +0,0 @@ ---- -layout: default -title: Privacy Is Not a Feature. It's an Architecture Decision. -parent: Perspectives -nav_order: 2 -description: Privacy toggles, data deletion tools, and privacy policies are theater. The only meaningful privacy guarantee is an architecture where the data never left your device in the first place. -faq: - - q: What is the difference between privacy as a feature and privacy as an architecture? - a: Privacy as a feature means controls layered on top of a system that already collects your data - deletion tools, opt-outs, consent banners. Privacy as an architecture means the system was never designed to collect your data in the first place. On-device AI is an example of the latter - there is nothing to delete because nothing was ever sent. - - q: Why aren't privacy policies sufficient? - a: A privacy policy is a legal document that describes what a company promises to do with your data. It doesn't change what's technically possible once the data is on their servers. Architecture determines what is possible. Policy determines what is promised. Only one of those is enforceable by design. ---- - -# Privacy Is Not a Feature. It's an Architecture Decision. - -"We take your privacy seriously." - -That sentence appears in the privacy policy of nearly every AI product in existence. It is also, in the strictest technical sense, irrelevant. - -## What privacy as a feature looks like - -Privacy features are controls layered on top of a system that was designed to collect your data first. - -They include: toggles that let you opt out of training. Data deletion requests that remove your history from a database. Consent banners that ask you to accept terms before using a product. Download-your-data buttons that let you see what was stored. - -These are not meaningless. They give users some agency. But they share a common assumption: your data was already on a server before any of these controls applied. - -The privacy feature model treats collection as the default and user control as the exception. The data moves first. The permissions come second. - -## What privacy as architecture looks like - -A different model starts with a different assumption: the data should never leave the device. - -The model runs in your phone's memory. Your query never becomes a network request. Your calendar and messages are never transmitted. Inference runs on your hardware, on your device. Nothing reaches an external server. - -There is nothing to delete. There is no policy to violate. There is no breach to notify you about. - -This is not a stronger version of the privacy feature model. It is a fundamentally different architecture where the privacy guarantee is a structural property, not a promise. - -## Why policy is not architecture - -A privacy policy is a legal document. It describes what a company promises to do with your data. It does not change what is technically possible once your data is on their servers. - -Architecture determines what is possible. Policy determines what is promised. A company can change its policy: by updating a terms of service, by being acquired, by responding to a government request. An architecture that never collected the data in the first place cannot be changed after the fact. - -This distinction matters more as the data becomes more sensitive. General search queries carry limited risk. Persistent personal context (your messages, health data, financial patterns, relationship history) carries significant risk. The architecture question is not abstract when the data at stake is that personal. - -## The consent problem - -Personal AI is uniquely difficult to make private by policy, because the value proposition requires access to your most sensitive data. - -An AI that can help you needs to know your calendar, your messages, your work patterns, your health. That's what makes it useful. The more context it has, the better it works. - -A cloud AI asks you to hand over that context in exchange for its capabilities. The implicit contract is: give us your data, we'll give you a useful assistant, and we promise to be responsible with it. - -An on-device AI inverts that contract. The context lives on your hardware. The model runs locally. The capabilities are the same, or better, because the model has more context than any cloud service would retain. But you never handed anything over. - -Consent only matters when there's something to consent to. On-device AI removes the question. - -## What this means for how AI should be built - -If privacy is an architecture decision, it has to be made at the beginning: in the choice of where inference runs, where context is stored, and what leaves the device. - -A product that runs inference in the cloud and adds privacy controls on top is a cloud AI with privacy features. - -A product that runs inference on-device, stores context locally, and sends nothing to external servers is a private AI by architecture. There is no feature to ship, no toggle to add, no policy to write. The privacy guarantee is in the design. - -This is the only version of personal AI that deserves access to your full context. Not because the company behind it is more trustworthy. But because the architecture makes trust irrelevant. The data never left your device. - -*Off Grid runs every model locally. No data leaves your device. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/regulatory-case-for-on-device-ai.md b/website/writing/regulatory-case-for-on-device-ai.md deleted file mode 100644 index 96ad1d23..00000000 --- a/website/writing/regulatory-case-for-on-device-ai.md +++ /dev/null @@ -1,69 +0,0 @@ ---- -layout: default -title: "The Regulatory Case for On-Device AI: Why Every New Privacy Law Is a Tailwind" -parent: Perspectives -nav_order: 17 -description: Every major privacy regulation passed in the last five years is a tailwind for on-device AI. The architecture that's right for users is also the architecture that's inherently regulation-proof. ---- - -# The Regulatory Case for On-Device AI: Why Every New Privacy Law Is a Tailwind - -Privacy regulation is accelerating globally. Jurisdiction after jurisdiction has passed, or is passing, laws that create obligations around the collection, processing, and transfer of personal data. More are coming. - -Each new regulation creates compliance requirements for AI products that process personal data. Legal teams, compliance frameworks, data protection impact assessments, consent management systems. The overhead is real and the risk of non-compliance is significant. - -On-device AI has a different relationship with this regulatory environment. Not a better compliance strategy. A fundamentally different architecture where most of the compliance questions don't arise in the first place. - -## What regulations are trying to solve - -Privacy regulations are responses to a specific problem: personal data is being collected, processed, and used by third parties in ways that users don't fully understand or control. - -The legislative approach is to require transparency, consent, and accountability. Tell users what you collect. Get their consent. Give them rights to access, correct, and delete their data. Be accountable for what you do with it. - -These requirements make sense for systems that collect personal data on remote servers. They create meaningful obligations for companies that would otherwise have no accountability for how they handle user information. - -On-device AI sidesteps the underlying problem. If no data leaves the device, there is no third-party collection to regulate. - -## Data protection law and the personal data question - -The dominant framework across most jurisdictions today is triggered by the processing of personal data by a data controller, typically a company that collects and processes user information on its infrastructure. - -An on-device AI processes personal data, but it processes it locally, on your own hardware, under your own control. The question of whether these frameworks apply to this processing, where you are essentially processing your own data for your own purposes, is nuanced, but the core compliance risks they address (third-party access, cross-border transfer, consent for commercial processing) largely don't apply. - -For a cloud AI product, compliance requires data processing agreements, consent management, data subject rights infrastructure, transfer mechanisms for cross-border data flows, and breach notification processes. For an on-device AI with no telemetry and no cloud infrastructure, these requirements either don't apply or are trivially satisfied. - -## AI-specific regulation and transparency requirements - -Regulators are now building on data protection frameworks with AI-specific rules. Risk-based classification for AI systems, transparency requirements for systems that interact with natural persons, obligations around training data provenance. - -Personal AI OS systems that act as productivity tools rather than decision-making systems in regulated domains are generally not in the highest-risk categories under these frameworks. But the transparency requirements are relevant, and on-device AI using open-weight models is well-positioned to meet them. - -The model card, training data provenance, and architecture of open-weight models are publicly documented. The openness that's right for users is the same openness that satisfies regulatory transparency requirements. A closed proprietary model running in the cloud is harder to audit. An open model running on your hardware is auditable by anyone. - -## The market dimension - -Privacy regulation doesn't just create compliance requirements. It creates market signal. - -Users in markets with strong privacy frameworks have come to expect more control over their data. Businesses operating in those markets face real consequences for non-compliance. As these frameworks expand to more jurisdictions, and as the AI-specific provisions within them become more detailed, the gap between cloud AI and on-device AI from a compliance perspective will widen. - -Every new regulation adds to the compliance overhead of cloud AI products. Every new regulation reduces that overhead to near-zero for on-device AI. The product that can credibly offer regulatory compliance-by-architecture, without the associated cost and complexity, has a structural market advantage. - -## The pattern across jurisdictions - -The pattern across privacy regulations globally is consistent. - -Each regulation defines compliance obligations triggered by third-party collection and processing of personal data. Each regulation creates overhead: consent management, data subject rights, breach notification, cross-border transfer mechanisms. Each regulation creates legal risk for products that fail to comply. - -On-device AI is not exempt from regulation. But the architecture dramatically reduces the surface area that regulations are targeting. The obligations that require the most compliance investment (cross-border transfers, third-party processing agreements, large-scale personal data handling) mostly don't apply to a system that processes data locally and sends nothing to external servers. - -Every new privacy regulation is a tailwind for on-device AI. Not because the regulatory environment is hostile to cloud AI specifically, but because the on-device architecture is inherently aligned with what regulators are trying to achieve. - -## The forward look - -Privacy regulation will continue to expand. More jurisdictions will pass legislation. Existing frameworks will be updated with AI-specific provisions. The compliance burden for cloud AI products will grow. - -The products that built their architecture around on-device processing from the start will not be scrambling to retrofit compliance. The architecture is the compliance. - -This is not the primary argument for building on-device AI. The primary argument is that it's better for you. But in a regulatory environment that's moving in one direction, the architecture that's right for users also happens to be the architecture that ages well. - -*Off Grid processes all data on-device. No cloud. No telemetry. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/the-small-things.md b/website/writing/the-small-things.md deleted file mode 100644 index 7dc12abd..00000000 --- a/website/writing/the-small-things.md +++ /dev/null @@ -1,77 +0,0 @@ ---- -layout: default -title: "It's Not About Productivity. It's About the 35 Tabs." -parent: Perspectives -nav_order: 29 -description: Hiring a secretary doesn't 5x your business. It just means life is easier. We spend 90% of our time on digital devices and almost none of that time is actually easy. ---- - -# It's Not About Productivity. It's About the 35 Tabs. - -You have 35 tabs open right now. - -Not because you're disorganised. Because closing them feels like losing something. You visited a page, read something useful, and now it lives in your browser because if you close it, it's gone. Not important enough to bookmark. Not unimportant enough to let go. So it stays. Along with the 34 others. - -That's not a productivity problem. That's a memory problem. Your browser has no memory. You're compensating for it by keeping the tabs open as a physical reminder that the thing exists. - ---- - -Nobody hires a secretary and expects to 5x their business. - -That's not what a secretary is for. A secretary means you never lose the thing you were looking for. It means you're prepared for your next meeting without spending 15 minutes assembling context. It means the follow-up email that should have gone out on Thursday actually went out on Thursday. It means when someone asks "did we ever resolve that?" you don't have to think about it. - -Life is smoother. That's it. That's the whole value proposition. - -We've been sold a version of AI that promises transformation. Supercharged productivity. Workflows automated. Hours reclaimed. And maybe some of that is real for some people. But for most people, most of the time, that's not what's actually broken about their day. - -What's broken is smaller than that. And it happens constantly. - ---- - -You spend 90% of your waking hours on digital devices. - -Think about what that actually means. Your phone is the first thing you look at in the morning. Your laptop is open for most of the working day. Your phone is back in your hand by evening. Screen time data is consistently between 7 and 11 hours a day for knowledge workers. - -And almost none of that time is genuinely easy. - -Not in the way that physical tools are easy. A good pen writes. A good chair supports you. They don't make you search for information you already had. They don't make you reconstruct context you already assembled. They don't lose things. - -Your digital devices lose things constantly. They just lose them in ways you've become so accustomed to that you've stopped noticing it's happening. - -The tab you kept open for three weeks because you knew you'd need it. The message you sent six months ago that you need to find now and can't remember the exact words to search for. The name of the person someone mentioned in a meeting that you meant to look up after. The article you read on your phone that you want to reference on your laptop but now have no idea where it was. The document you wrote two months ago that definitely exists somewhere. - -Each one is a small friction. A moment where your device, which witnessed everything, offers nothing. - ---- - -A good personal assistant fixes this without you noticing. - -Not by being smarter than you. Not by making better decisions. Just by remembering. By tracking. By being there when you need the thing, with the thing. - -"That article from last week about X" gets you the article. "The email where we agreed on the scope" gets you the email. "What was that company someone mentioned in our call on Tuesday" gets you the answer. - -None of this is impressive. None of it will appear in a product demo. It doesn't make a good headline about AI transforming your workflow. It just means your day has less friction in it. And you have 35 fewer tabs open. - ---- - -That's what a Personal AI OS actually is, when you strip away the category talk. - -It's the thing that watched you visit that page and remembers it. It knows your browsing is part of your context just like your messages and your calendar. When you need it, you ask in plain language and it finds it. You didn't have to decide it was worth bookmarking. You don't have to remember where it was or when you saw it. You just ask. - -It's not surveillance. It's your memory, running locally on your hardware, available only to you. The same way a secretary keeps notes that belong to you, not to the firm they work for. - -The 35 tabs are open because nothing in your digital life plays this role. Not your browser. Not your operating system. Not any of the AI products that exist today, because they don't have access to your context, or they have it but it lives on a server you don't control, or they only know what you've explicitly told them in the current session. - ---- - -The promise worth making is not the big one. - -Not "this will transform how you work." Not "you'll get hours back every week." Those might be true for some people. But they're not the promise that matters to most people most of the time. - -The promise that matters is: your digital life will be a little easier. The things you've already seen will be findable. The things you've already done won't need to be redone. The context you've already assembled won't need to be reassembled. - -That's what a good assistant gives you. Not transformation. Smoothness. And after 20 years of digital devices that constantly lose things you gave them, smoothness is not a small thing. - ---- - -*Off Grid is building toward this. It starts with on-device AI that works fully offline on your phone, the foundation that makes everything above possible without your data ever leaving your device. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/two-devices-zero-context.md b/website/writing/two-devices-zero-context.md deleted file mode 100644 index 8b17b8a9..00000000 --- a/website/writing/two-devices-zero-context.md +++ /dev/null @@ -1,89 +0,0 @@ ---- -layout: default -title: "Two Devices, Zero Shared Context: The Problem the Personal AI OS Was Built to Solve" -parent: Perspectives -nav_order: 24 -description: Your laptop sees your work. Your phone sees your life. Neither talks to the other. A Personal AI OS bridges them - locally, privately, without a server in between. This is the product-thesis piece. ---- - -# Two Devices, Zero Shared Context: The Problem the Personal AI OS Was Built to Solve - -The average knowledge worker uses two primary devices. A phone, which is with them from the moment they wake up until they go to sleep. A laptop, which is open for most of their working day. - -These two devices, used by the same person, pursuing the same goals, facing the same constraints, have never had a conversation with each other. - -Not at the intelligence layer. Not in a way that means anything. - ---- - -## Two siloed views of one life - -Your laptop sees your work. - -It knows what you are writing, what you are reading, what research you are doing. It knows your email: the threads you are managing, the commitments you have made, the conversations in progress. It has your files, your code, your documents. It is the most accurate record of your professional output that has ever existed. - -Your phone sees your life. - -It knows your personal messages, your relationships, your social context. It knows your health: sleep, activity, physical patterns. It knows your location: where you go, how often, for how long. It knows your calendar commitments in the context of everything else competing for your time. - -Neither device has access to what the other sees. The intelligence built into each operates in isolation. - -Your laptop does not know you slept four hours last night. Your phone does not know your deadline moved to tomorrow. Your laptop does not know your most important client just sent a message. Your phone does not know you are in the middle of something that needs two more focused hours. - -You hold all of this yourself. In your head. Across two separate devices, two separate intelligence layers, two separate worlds. - ---- - -## The cost of the split - -The split creates a specific kind of overhead that knowledge workers carry constantly without fully recognising it as a structural problem. - -**Context assembly.** Before every significant task (a meeting, a difficult message, a decision) you assemble context manually. You check your calendar on your laptop, your messages on your phone, your notes somewhere else. The information exists. Gathering it is work you do. - -**Cross-device triage.** A notification arrives on your phone while you are working on your laptop. You pick up your phone, switch context, assess it, decide how to respond, put the phone down, and try to reconstruct your train of thought. This happens many times a day. - -**Memory as a bridge.** Because neither device knows what the other knows, you serve as the bridge. You remember that the message on your phone is related to the file you were working on your laptop. You remember that your laptop deadline affects whether you can take the call your phone is suggesting. Your memory is doing coordination work that software should be doing. - ---- - -## What the Personal AI OS does - -A Personal AI OS treats both devices as part of a single intelligence system. - -Context built on your phone (messages, health, location, personal calendar) is available on your laptop. Context built on your laptop (files, email, work calendar, current projects) is available on your phone. Not through a cloud relay. Over your local network, privately, between devices you own. - -The AI on either device has a unified view of you. When you ask it to help you prepare for a meeting, it draws on your phone's knowledge of the recent conversation with that person and your laptop's knowledge of the last document you shared with them. When it surfaces a notification, it knows whether you are in the middle of focused work on your laptop and can defer accordingly. - -You are one person. The intelligence layer knows that. - ---- - -## Why this requires a different architecture - -Cross-device context sharing at the intelligence layer is not a feature you can add to existing products. It requires different architecture from the ground up. - -Cloud sync gives you the same data on both devices: your calendar is on your phone and your laptop. Data sync is not intelligence sync. Having the same calendar on both devices does not give either device's AI a unified view of your context. Each AI still operates in isolation. - -True cross-device intelligence requires a context model that spans both devices and is updated continuously from both. That model is a representation of who you are and what is happening in your life. That context model has to live somewhere. The right place is your devices, synced over your local network. A cloud server that receives your most personal data as a side effect of providing coordination is the wrong place. - -The architecture that solves the two-device problem is the same architecture that solves the privacy problem. On-device context. Local network sync. No server in between. - ---- - -## The product thesis - -Off Grid's thesis starts here. - -The most fundamental thing broken about personal computing today is the gap between what your devices know about you and what they do with it. Specifically: two devices that serve the same person but operate in isolation. - -Closing that gap, privately, without requiring you to hand your most personal context to external infrastructure, is what the Personal AI OS is built to do. - -The phone is where we start. It is the most context-rich device. It is with you all day. The AI that runs on it, entirely locally, with access to your messages and calendar and health, is the first piece of an intelligence layer that eventually spans your whole life. - -The laptop integration is next. Then the full cross-device context sync over your local network. - -Two devices. One intelligence layer. No server required. - ---- - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/va-industry-disruption.md b/website/writing/va-industry-disruption.md deleted file mode 100644 index 51a73896..00000000 --- a/website/writing/va-industry-disruption.md +++ /dev/null @@ -1,92 +0,0 @@ ---- -layout: default -title: "Why the Virtual Assistant Industry Is About to Be Disrupted by On-Device AI" -parent: Perspectives -nav_order: 28 -description: The VA market is worth billions and growing. On-device AI doesn't compete with it on price. It competes on a dimension the human model can't match: full personal context, always available, private by architecture. ---- - -# Why the Virtual Assistant Industry Is About to Be Disrupted by On-Device AI - -The market for virtual assistance is a multi-billion dollar industry and growing. Human remote workers handling the administrative work of knowledge workers and small businesses built it. - -The growth makes sense. There is a genuine and expanding need for administrative intelligence at the individual level. The knowledge worker's day is full of work that requires judgment but not their specific judgment: coordination, triage, drafting, scheduling, research. Offloading that work to a capable assistant makes the principal demonstrably more effective. The value proposition is not in question. - -What is about to change is where that intelligence comes from. - ---- - -## What built the VA industry - -The VA industry emerged from a simple structural insight: modern communication tools made it possible to provide administrative support remotely, at lower cost than local hiring. - -Before the internet made real-time remote collaboration viable, administrative support required physical proximity. The secretary sat down the hall. The assistant was in the same building. - -Remote work tools changed that. You could work with an assistant in a different city, a different time zone, at dramatically lower cost. For a knowledge worker generating value at a professional rate, the arbitrage was obvious: pay less per hour for the coordination work, free up your own hours for the work only you could do. - -The economics were compelling. The industry scaled accordingly. - ---- - -## The ceiling in the model - -But the VA model has a structural ceiling, and it shows up most clearly in the information asymmetry. - -An assistant can only act on what they know. And what they know is limited to what you've shared with them. - -They don't have access to your full message history. They can't see your health data or understand that you're running on three hours of sleep. They don't know the backstory of every relationship in your contact list. They can't read between the lines of your calendar and notice the pattern that you systematically overbook yourself on Wednesdays and regret it by Thursday morning. - -The intelligence they provide is bounded by the context you're willing to share, which is always less than the full picture. Sharing the full picture with another person is its own kind of exposure. - -This creates a paradox. The most valuable administrative intelligence would come from a system with complete context. But the more complete the context, the more you're sharing with someone else. Human VAs resolve this by having limited context. Which limits the intelligence. - ---- - -## Where on-device AI breaks the model - -A Personal AI OS changes the information asymmetry completely. - -It has access to your full context: your messages, your calendar, your files, your communication history, your work patterns, your health data, your location patterns. Not as a snapshot you've deliberately shared, but as a live, continuously updated picture of your life. - -And it has that context without you handing it to another person. The data stays on your device. The processing happens on your hardware. Nothing leaves. - -This is the dimension the human model structurally cannot match. No human VA can have your full context without you giving it to them. An on-device AI has your full context precisely because it never leaves your hands. - -The intelligence that results from full context is categorically different from the intelligence that results from partial context: - -- It can triage your inbox understanding not just the content of each message but the full history of your relationship with each sender. -- It can prepare you for a meeting drawing on every previous interaction with those people, every relevant document, every commitment that was made. -- It can draft a follow-up in your voice with the specifics of what was actually discussed, not a generic template. -- It can notice that the email that just arrived is related to the calendar event you've been anxious about and surface them together. - -These are not incremental improvements on what a human VA does. They are capabilities that require a different kind of context access. One that a human assistant can't have by design. - ---- - -## What happens to the VA industry - -The disruption of the VA industry by on-device AI won't look like a sudden cliff. It will look like two things happening simultaneously. - -At the bottom of the market, the use cases already best suited to automation, AI takes over. The clients who used VAs for templated, repeatable administrative work find that an on-device system does it better, faster, and without the relationship overhead. - -At the top of the market, human VAs move up the value chain. The work that remains for human assistants is the work that genuinely requires human judgment, interpersonal skill, and real-world presence. The coordinators and schedulers become relationship managers and strategic operators. The function that couldn't be automated becomes more valued because the function that could be automated now is. - -This is the pattern technology disruption always follows: the lowest value-add work gets automated first, and the people who were doing it move to higher value-add work or exit the market. - ---- - -## The people this actually reaches - -The VA industry, for all its growth, remained a service primarily available to knowledge workers above a certain income threshold. The cost of a skilled human VA was prohibitive for most people who could have benefited from it. Even remote, even part-time. - -The individuals who needed administrative support most urgently were often the ones least able to afford it: sole traders, early-stage entrepreneurs, freelancers managing multiple clients, mid-level professionals drowning in coordination overhead. - -On-device AI doesn't just disrupt the existing VA market. It creates a market that previously didn't exist: administrative intelligence for the people the human model never reached. - -The device in their pocket already has the compute to run it. The models are open-weight and free to use. The only thing that was missing was the product that made those capabilities into an intelligence layer. - -That product is being built now. - ---- - -*Off Grid is building toward this. It starts with on-device AI that works fully offline on your phone, the foundation that makes everything above possible without your data ever leaving your device. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/walled-garden-problem.md b/website/writing/walled-garden-problem.md deleted file mode 100644 index 23562c41..00000000 --- a/website/writing/walled-garden-problem.md +++ /dev/null @@ -1,65 +0,0 @@ ---- -layout: default -title: "The Walled Garden Problem: Why the Personal AI OS Must Be Open" -parent: Perspectives -nav_order: 16 -description: Platform AI is real, capable, and useful. But the architecture of platform AI makes a genuine Personal AI OS impossible from within it. Here's why openness is not optional for the category. ---- - -# The Walled Garden Problem: Why the Personal AI OS Must Be Open - -Platform AI - the AI features built into iOS, Android, and the major operating systems - is impressive. It summarises notifications. It generates text in the keyboard. It describes images for accessibility. It answers questions about your device. - -Acknowledging this matters. Platform AI represents billions of dollars of investment and genuine engineering capability. It is making devices meaningfully smarter. - -But platform AI cannot be a Personal AI OS. The architecture won't allow it. - -## What platform AI gets right - -Platform AI has one advantage that independent software cannot easily replicate: deep OS-level integration. - -It can read notifications across all apps because it has OS-level permission to do so. It can generate text in any text field because it's built into the keyboard at the system level. It can take actions - setting reminders, making calls, sending messages - because it has the permissions granted to the OS itself. - -This integration is valuable. The friction of independent AI apps is that they have to ask for each permission explicitly and work within the sandboxing model the OS imposes. Platform AI doesn't have this constraint. - -## What platform AI cannot do - -Three structural properties of platform AI make a genuine Personal AI OS impossible within it. - -It is closed by design. You cannot inspect what platform AI does with your data. You cannot verify that inferences stay on-device. You cannot audit the model weights. You accept the platform's representations about privacy as a matter of trust, with no way to verify them. - -For a system with access to your messages, health data, and files, unverifiable trust is a weak foundation. The 7 principles of a Personal AI OS include open and auditable for this reason. Closed is disqualifying. - -It is bound to the platform. Platform AI features exist within one ecosystem. The AI on your iPhone does not have access to your Android tablet or your Windows laptop. The AI on your Android phone does not have access to your Mac. - -A Personal AI OS is a single intelligence layer across all your devices. It requires interoperability - open protocols, open model formats, software that runs on any hardware. That is structurally incompatible with the platform model, where the AI feature is a competitive differentiator that only works within the walled garden. - -Its incentives are misaligned. Platform companies are not primarily AI companies. They are platform companies. AI features serve platform goals: device differentiation, ecosystem stickiness, data collection that supports advertising or services revenue. - -A Personal AI OS should be optimised for your outcomes, not for the platform's metrics. When those conflict - when the personally optimal AI behaviour would reduce platform engagement or break ecosystem lock-in - platform AI will optimise for the platform. That's not a criticism. It's what the incentive structure produces. - -## What openness requires - -An open Personal AI OS has four properties. - -Open models. The model weights are public. Anyone can run them, inspect them, fine-tune them. You are not dependent on a vendor's decision about which models to support. - -Open source application. The code that orchestrates the AI, manages context, and takes actions is inspectable. You can verify what it does. The community can audit it. - -Open protocols for cross-device sync. The format for context and the protocol for device-to-device communication are documented and open. Any compatible software can participate in your personal intelligence network. - -No platform exclusivity. The software runs on any hardware that supports it. Not just Apple. Not just Android. Any device you use. - -## The role of independent software - -Platform AI and independent Personal AI OS software are not in direct competition. They are different things with different capabilities and different tradeoffs. - -Platform AI will keep getting better at the things platform AI is good at: low-friction, deeply integrated features for the platform's users. - -Independent Personal AI OS software will build the things platform AI cannot: full openness, cross-platform context, architecture that earns trust through verifiability rather than through policy. - -The question for you is which matters more for the use case you care about. For casual AI features - text suggestions, notification summaries - platform AI is probably enough. For a genuine intelligence layer with access to your full context, the open architecture is necessary. - -Off Grid is building the latter. - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing). [View the source on GitHub](https://github.com/alichherawalla/off-grid-mobile?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/what-is-personal-ai-os.md b/website/writing/what-is-personal-ai-os.md deleted file mode 100644 index fd14c5ab..00000000 --- a/website/writing/what-is-personal-ai-os.md +++ /dev/null @@ -1,95 +0,0 @@ ---- -layout: default -title: What Is a Personal AI OS? -parent: Perspectives -nav_order: 1 -description: A Personal AI OS is intelligence that lives on your device, knows your full context, and acts on your behalf - without ever sending data to a server. Here's what defines the category and why it matters. -faq: - - q: What is a Personal AI OS? - a: A Personal AI OS is an intelligence layer that runs entirely on your own hardware - phone, laptop, or both - with access to your full personal context (messages, calendar, files, health data) and the ability to act on your behalf. Unlike cloud AI assistants, it never sends your data to a server. It runs offline, charges no ongoing fees for AI compute, and is auditable by design. - - q: How is a Personal AI OS different from a platform AI assistant? - a: Platform AI assistants are cloud-dependent voice interfaces. They send your queries to remote servers, return responses, and retain minimal context between sessions. A Personal AI OS runs locally on your device, maintains persistent context about your life and work, and can act across apps on your behalf - not just answer isolated questions. - - q: How is a Personal AI OS different from an AI agent? - a: AI agents are typically autonomous systems that make decisions and take actions with minimal human oversight, often connected to external services. A Personal AI OS is explicitly non-autonomous - it acts on your behalf with your consent, defers to you on decisions, and operates within the boundary of your own hardware and local network. - - q: Does a Personal AI OS require an internet connection? - a: No. A Personal AI OS runs on-device. It does not require an internet connection for inference, context retrieval, or action execution. Network access may be used optionally for specific tasks (web search, calendar sync) but the core intelligence operates entirely offline. ---- - -# What Is a Personal AI OS? - -Every few years a new software category gets named before it gets built. Personal computing. The smartphone OS. The cloud platform. Each one felt obvious in retrospect and premature when first articulated. - -Personal AI OS is the next one. - ---- - -## The definition - -A Personal AI OS is an intelligence layer that: - -- Runs entirely on hardware you own -- Has access to your full personal context - messages, calendar, files, health, location -- Can act on your behalf across apps and devices -- Operates offline by default, with no data sent to external servers -- Persists context between sessions, building a working model of your life and work -- Is open and auditable - no black-box telemetry, no hidden data collection - -That's the category. Everything else currently called AI (cloud assistants, chatbots, autonomous agents) is something different. - ---- - -## Why this is a new category - -The dominant AI products today are cloud services. You send them a query. They process it on a remote server. They return a response. Your data passes through infrastructure you don't control, gets logged, and contributes to models you can't inspect. - -This works for general-purpose tasks where your personal context doesn't matter. Ask about the weather in Tokyo or summarise a Wikipedia article. It doesn't matter that the request went to a server. - -But the tasks where AI becomes useful are the ones that require knowing you. Triaging your inbox. Preparing for your next meeting. Noticing that you have three conflicting commitments next Thursday. Drafting a message in your tone, not a generic one. - -For those tasks, the AI needs your data. Handing your most personal data to a server you don't control, in exchange for a subscription, is a trade most people haven't consciously agreed to. - -A Personal AI OS resolves this by keeping the intelligence local. The model runs on your device. Your context never leaves. The most capable AI for your life is also the most private: not by policy, but by architecture. - ---- - -## The 7 principles - -These are the properties that define a true Personal AI OS. They are structural requirements. An AI product that fails any one of them is something else. - -**1. Runs on-device.** Inference happens on your hardware: CPU, GPU, or NPU. No query is sent to a remote model. No response comes back from a server. - -**2. Never phones home.** No telemetry. No usage logs. No data collection of any kind. What happens on your device stays on your device. - -**3. Persistent context.** The AI maintains a working model of your life across sessions. It knows your calendar, your recent messages, your open tasks, your work patterns. Context is the primitive, not queries. - -**4. Acts on your behalf.** The AI can take actions (draft messages, set reminders, summarise documents, search your files), not just answer questions. Agency, with your consent as the operating principle. - -**5. Works across your devices.** Your phone and laptop are used by one person. The AI should have a unified view across both, synced over your local network without a cloud relay. - -**6. Open and auditable.** The model weights and application code are inspectable. You can verify what the AI does and does not do with your data. Trust through transparency, not through policy. - -**7. No cloud compute rent.** You do not pay ongoing fees for someone else's servers to process your queries. The model runs on your hardware. There is no server cost to recover from you. Software may have a price, because building it takes work, but the AI itself is not metered. - ---- - -## What it is not - -A Personal AI OS is not an autonomous agent. It does not make decisions on your behalf without your knowledge. It does not connect to external services without your explicit direction. It does not run in the background taking actions you haven't approved. - -It is also not a walled garden. The category requires openness: open models, open source code, open protocols for cross-device communication. A closed Personal AI OS is a contradiction in terms. - -And it is not a product tied to a hardware platform. The AI features built into operating systems are constrained by the platform's architecture and commercial interests. A Personal AI OS is an independent layer that runs on your hardware regardless of who made it. - ---- - -## Why it matters - -800 million knowledge workers use a phone and a laptop every day. Both devices hold the context that would make AI useful. Neither does anything meaningful with it. - -The Personal AI OS is the software category that closes that gap. It is the first architecture that earns the right to your full context, because the data never leaves your hands. - -That's what we're building with [Off Grid]({{ '/' | relative_url }}). - ---- - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/what-personal-ai-should-know.md b/website/writing/what-personal-ai-should-know.md deleted file mode 100644 index 9a624e12..00000000 --- a/website/writing/what-personal-ai-should-know.md +++ /dev/null @@ -1,75 +0,0 @@ ---- -layout: default -title: "What a Personal AI OS Should Know About You - And What It Shouldn't" -parent: Perspectives -nav_order: 9 -description: The right level of context makes a Personal AI OS useful. The wrong level makes it something you don't want near your life. Here's where the line is and why it matters. ---- - -# What a Personal AI OS Should Know About You - And What It Shouldn't - -Context is what makes a personal AI useful. The more it knows about your patterns, commitments, and working style, the better it can help you. - -But context without limits is surveillance. Even self-directed surveillance produces a system that knows too much about you in ways that change how you behave around it. - -The question is not "how much context should a Personal AI OS have?" It's "what kind of context serves you, and what kind creates a system you'd rather not live with?" - ---- - -## What a Personal AI OS should know - -**Your schedule.** The pattern of your week: when you do focused work, when you take calls, when you're typically unavailable, how often plans change. This lets the AI reason about your time in ways that a simple calendar view can't. - -**Your communication patterns.** The rhythm of your messages: how quickly you typically respond, which conversations you prioritise, which you defer. Not the content of every message, but the structure of your communication life. - -**Your active work context.** What you're currently working on, what's blocked, what's coming due. The AI should know enough about your work to help you prepare, prioritise, and not miss things. - -**Your preferences and style.** How you write. What you consider important. How you prefer information presented. These don't need to be explicitly programmed. They emerge from observing how you interact with the system over time. - -**Your recent activity.** What you've been doing in the last few hours and days. Not a permanent record, but enough recent context to understand where you are in your work and what's front of mind. - ---- - -## What a Personal AI OS should not know - -**Historical records you don't need it to have.** The value of persistent context comes from understanding patterns, not from storing everything indefinitely. A Personal AI OS that retains five years of messages and location history is building a liability, not a feature. Context should have a horizon: enough to be useful, not so much that it becomes a permanent record of your life. - -**Sensitive personal domains you haven't explicitly opened.** Your financial accounts, your medical records, your private relationships: these require explicit, intentional access grants. The AI should not assume that access to your calendar means access to everything connected to it. - -**Inferences you haven't verified.** A Personal AI OS can notice patterns ("you seem to do your best work in the mornings") but it should surface those observations for your confirmation rather than silently acting on them. Inferences about your mental state, your relationships, or your intentions are especially dangerous to act on without verification. - -**Enough to manipulate you.** The line between a helpful personal AI and a manipulative one is whether it's optimising for your outcomes or for its engagement with you. A system that knows your emotional patterns well enough to time notifications for moments of vulnerability is not an assistant. It's an adversary. The Personal AI OS should have this line built in from the start. - ---- - -## The consent principle - -The right framework for a Personal AI OS is explicit consent for each category of access, with the ability to revoke at any time. - -Calendar access is not messages access. Messages access is not health data access. Each extension of context should be a deliberate choice: not an opt-out buried in settings, but an opt-in made with a clear understanding of what the AI gains and what you gain from the exchange. - -This is a design principle as much as a privacy one. A system you don't fully trust is a system you won't give full context. A system you've explicitly consented to is one you can actually use without reservation. - ---- - -## The audit principle - -A Personal AI OS should show its work. - -Not in a technical sense, not raw logs of every inference step. But in a legible sense: if the AI says "I prioritised this message because you typically respond to this contact quickly," that reasoning should be accessible and correctable. - -Opacity breeds distrust. A system that makes recommendations without explanation creates anxiety about what it might know or infer that it isn't saying. Transparency about what context the AI has used, and the ability to correct its model of you, is part of what makes it a tool rather than something that's just happening to you. - ---- - -## The minimalism principle - -A Personal AI OS should know as much as it needs to help you and no more. - -This is good design, not just privacy hygiene. A system with too much context becomes slow, noisy, and prone to surfacing irrelevant information. A system tuned to the right level of context is fast, accurate, and feels like it actually understands you. - -The goal is a system that knows the right things (your schedule, your priorities, your working style) well enough to reduce friction and surface what matters, without becoming a burden of its own. - ---- - -*Off Grid processes context locally. You control what it can access. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/whatsapp-moment-for-ai.md b/website/writing/whatsapp-moment-for-ai.md deleted file mode 100644 index 3d21b4cd..00000000 --- a/website/writing/whatsapp-moment-for-ai.md +++ /dev/null @@ -1,69 +0,0 @@ ---- -layout: default -title: "The Encrypted Messaging Moment for AI: Why Privacy Will Define the Next Platform" -parent: Perspectives -nav_order: 4 -description: Encrypted messaging went mainstream because the market demanded it. AI is the next communication infrastructure. The arc is the same - and the outcome will be the same. ---- - -# The Encrypted Messaging Moment for AI: Why Privacy Will Define the Next Platform - -Around 2016, encrypted messaging crossed into the mainstream. - -The major messaging platforms, which had resisted encryption for years because it limited their own data access, shipped end-to-end encryption anyway. Not because the engineers pushed for it. Because the market demanded it. - -After years of high-profile data breaches, after growing public awareness that messages on unencrypted platforms were readable by the platforms themselves, users started to care. Encrypted-first apps had been gaining ground on privacy. The market signal was clear enough that even platforms with no obvious incentive to reduce their own data access made the switch. - -AI is the next communication infrastructure. The same arc applies. - -## Why AI is communication infrastructure - -The canonical communication technologies (telephone, email, SMS, messaging apps) all share a defining property: they carry the most private content in a person's life. - -Your phone calls are where you talk to your doctor, your lawyer, your family. Your messages are where you say things you wouldn't say in public. The history of communication technology is a history of fighting for the right to have those conversations without a third party listening. - -AI is becoming the next layer of that infrastructure. An AI assistant that knows your messages, your health, your finances, your work, and that can act on your behalf, is the most intimate piece of software ever built. It has more context than any communication app because its entire value proposition is having more context. - -That makes the privacy question central. The same concerns that drove demand for encrypted messaging are, at higher stakes, the concerns that will drive demand for on-device AI. - -## The demand arc - -Privacy doesn't win in technology markets because people are principled. It wins because a critical mass of users has a concrete bad experience, understands what caused it, and has an alternative to switch to. - -For messaging, that moment came gradually. Breaches, then acquisitions with changed terms, then enough mainstream coverage that ordinary people understood their messages were readable by the platforms carrying them. Encrypted messaging went from a niche concern to a mainstream expectation. - -For AI, the trigger events will be different in their specifics but identical in their structure. Moments where users experience the consequences of their most personal data sitting on someone else's server. Moments where they understand the alternative exists. Moments where they switch. - -Some of those moments will involve breaches. Some will involve policy changes that remove access to user data. Some will involve acquisitions where the terms change after users have already given years of context to a product, and they then lose access to their own data overnight when the company shuts down or changes hands. The specifics will vary. The outcome will not: a market that demands on-device AI. - -## The structural difference - -The encrypted messaging story has one important caveat: encryption protects data in transit, but the platform still knows who you're talking to, when, and how often. Metadata remained. The key privacy property was delivered, but the full picture is more complicated. - -On-device AI can be structurally cleaner. If inference runs locally and context is stored on-device, there is no transit to encrypt. There is no server that sees metadata. The architecture doesn't produce the data that would need to be protected in the first place. - -This is what "private by architecture" means in practice. Not better encryption. Not stronger policy. An architecture that eliminates the exposure surface entirely. - -## What has to be true for this to happen - -The encrypted messaging moment for AI requires three things to align. - -The technology has to be good enough. When encrypted messaging went mainstream, it was already as fast and reliable as unencrypted messaging. Users didn't have to sacrifice quality for privacy. On-device AI is approaching that inflection point. Models like Qwen 3.5, Gemma 4, and Phi-4 run in real time on current flagship phones. The gap with cloud models is closing. - -The alternatives have to be visible. Users can't demand what they don't know exists. The role of products like Off Grid is partly technical and partly demonstrative: showing that capable AI running entirely on-device is a present reality, not a future possibility. - -The consequences have to be understood. For AI, the equivalent is users understanding that the context they hand to cloud AI (the full text of their messages, their health records, their financial patterns) is being stored, potentially used for training, and potentially accessible to parties they didn't intend. That understanding is spreading. - -All three conditions are converging. The technology is ready. The alternatives exist. The consequences are becoming legible. - -## The platform question - -When the market demands private AI at scale, the question becomes: who built the infrastructure for it? - -The major platforms are late to the architecture. Their on-device AI efforts are features on top of existing cloud platforms, not genuine on-device intelligence layers. The openness required for a true Personal AI OS (open models, inspectable code, no platform lock-in) runs against their economic interests. - -The opportunity is for software that builds the intelligence layer independently of the platforms. Software that runs on the hardware you already own. Software that treats the privacy guarantee as an architectural property, not a marketing claim. - -That's the bet Off Grid is making. Not on a new device or a new platform. On the architecture being right. - -*[Download Off Grid for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/who-owns-your-ai-memory.md b/website/writing/who-owns-your-ai-memory.md deleted file mode 100644 index 0b44b135..00000000 --- a/website/writing/who-owns-your-ai-memory.md +++ /dev/null @@ -1,93 +0,0 @@ ---- -layout: default -title: "Who Owns Your AI's Memory? The Question Nobody Is Asking." -parent: Perspectives -nav_order: 20 -description: When your AI remembers everything about you - your patterns, your preferences, years of your context - who owns that memory? It's the most important digital rights question of the decade, and almost nobody is asking it. ---- - -# Who Owns Your AI's Memory? The Question Nobody Is Asking. - -AI products with persistent memory are becoming common. The system remembers what you told it last month. It knows your preferences, your patterns, your history. It uses that knowledge to give better responses. - -This is useful. An AI that knows you works better than one that starts from scratch every session. - -But nobody is asking the obvious question: who owns that memory? - ---- - -## What memory means for AI - -Persistent AI memory is not a simple data store. It is a working model of you. - -Over time, a system with persistent memory learns: how you communicate, what you care about, what your work involves, what your relationships are like, what decisions you've made and why, what you're worried about, what you find funny, what you avoid. It learns things about you that you haven't told anyone: patterns in your behaviour that emerge from the data rather than from explicit disclosure. - -This is the promise of persistent AI: it becomes more useful the longer you use it, because it knows you better. - -It also makes the ownership question significant. A memory this rich and detailed is the most thorough model of a person that has ever existed in software. - ---- - -## The ownership question - -When that memory lives on a company's server, the ownership is unclear. - -The data originated with you. The patterns were derived from your behaviour. But the storage, the infrastructure, and the model of you that was built all sit on infrastructure owned and controlled by the company. - -You cannot easily export it in a form another system can use. You cannot verify what is stored. You can request deletion, but you cannot verify it was deleted. If the company is acquired, the memory transfers to the acquiring entity under whatever terms were agreed. - -The memory that was supposed to be yours, built from your most personal data, is an asset a corporation can buy and sell. - ---- - -## What happens when the service changes - -The most concrete version of this problem appears when a service changes terms, is acquired, or shuts down. - -Users who have spent months or years building up context with an AI product, who handed over the context of their professional and personal lives, find that access to that context is controlled by someone else's business decisions. - -The AI that knew them is gone. Or it is now owned by a different company. Or it continues under terms that include training on their data in ways the original product did not allow. - -The memory they thought was theirs turns out to have been held by a company. Companies are bought, sold, and shut down. - ---- - -## Memory on your device - -The alternative is memory that lives on your device. - -Your context (your messages, your preferences, your work patterns, the model of you the AI has built) is stored locally. It moves with you to new devices over your local network. It does not require a server to exist. It does not disappear if a company is acquired. - -You can inspect it, because it is on your storage and the software that accesses it is open. You can delete specific things from it. You can export it. You can run it with a different AI model if you switch software. - -The memory is yours in the same way your documents are yours. It is on your hardware, under your control, not held by a third party. - ---- - -## Why this question will become central - -Persistent AI memory is still relatively new. Most users have not been using memory-enabled AI products long enough for the ownership question to feel urgent. - -It will. - -As AI memories get richer, as they start to include conversation history, your messages, files, and health data, the value of that memory increases. So does the risk of having it on someone else's server. - -The first wave of high-profile incidents around AI memory ownership will make this question visible to a mainstream audience: an acquisition where users lose access, a breach that exposes a detailed profile of millions of people, a terms change that makes historical memories available for training. - -When that happens, the products built with on-device memory from the start will have a significant advantage. Not because they were more capable, but because they were built on the right assumption: the memory belongs to the user. - ---- - -## The data rights frame - -Privacy regulation has spent a decade establishing the principle that personal data belongs to the person it is about. The right to access, correct, and delete your data. The right to portability. The right not to have your data sold without your consent. - -AI memory is a new form of personal data. It is arguably the most personal form that has ever existed, because it encodes a model of how you think and behave. - -The same principles apply. Your AI's memory of you is yours. You should be able to access it, move it, delete it, and ensure it does not end up somewhere you did not intend. - -On-device architecture is the only architecture that delivers on these principles without requiring a regulatory framework to enforce them. The memory runs on your device. You already own it. - ---- - -*Off Grid stores all context on your device. Your AI's memory is yours. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).* diff --git a/website/writing/why-personal-ai-should-never-live-in-cloud.md b/website/writing/why-personal-ai-should-never-live-in-cloud.md deleted file mode 100644 index 2a914881..00000000 --- a/website/writing/why-personal-ai-should-never-live-in-cloud.md +++ /dev/null @@ -1,87 +0,0 @@ ---- -layout: default -title: Why Your Personal AI Should Never Live in the Cloud -parent: Perspectives -nav_order: 8 -description: This is not a privacy rant. It's a structural argument. Cloud-dependent personal AI is broken by design - not because the companies building it are untrustworthy, but because the architecture makes the most important guarantees impossible. ---- - -# Why Your Personal AI Should Never Live in the Cloud - -This is not an argument about whether cloud companies are trustworthy. Assume they are. Assume the privacy policies are genuine, the security is excellent, and the intentions are good. - -The argument against cloud-dependent personal AI is structural. The architecture makes certain guarantees impossible - not unlikely, not risky, but impossible. For personal AI specifically, those are exactly the guarantees that matter. - ---- - -## The three structural problems - -### 1. The data has to leave your device - -A cloud AI processes your queries on a remote server. For that to happen, your query - and any context attached to it - has to travel across a network. - -For general-purpose queries, this is fine. Asking about the weather in Tokyo or summarising a Wikipedia article carries no personal risk. - -But a personal AI's value comes from personal context. The AI that can help you is the one that knows your messages, your calendar, your financial patterns, your health history. When that context rides a network request to a cloud server, it is no longer under your control. From that point, its fate is governed by policy, not architecture. - -Policy can change. Architecture cannot be changed retroactively. The moment the data leaves your device, the structural guarantee - "nothing can access this except you" - is gone. - -### 2. Continuity depends on the vendor - -Cloud AI products are services. Services have lifecycles. They get acquired. They change pricing. They pivot. They shut down. - -For a todo app or a news reader, this is a manageable risk. You might lose your data or have to migrate. Inconvenient, but recoverable. - -For a personal AI that has built a model of you over months or years - your patterns, your preferences, your context - service discontinuity is not an inconvenience. It's the loss of a system that has become load-bearing for how you work. - -On-device AI has no such dependency. The model runs on your hardware. The context is stored locally. If the company that shipped the software disappears tomorrow, you still have the model, the context, and the ability to run inference. Nothing about your setup depends on a server staying online. - -### 3. The incentive structure is misaligned - -A cloud AI business recovers its compute costs through subscriptions, API fees, or advertising. The marginal cost of inference scales with usage. The business needs your ongoing engagement. - -This creates incentives that are structurally misaligned with yours. You want an AI that makes you more efficient - that handles things quickly so you can move on. The business wants an AI that keeps you engaged. - -On-device AI has different economics. The compute runs on your hardware. There is no server cost to recover. The product can be designed entirely around your outcomes rather than around metrics that proxy for revenue. - -A subscription for on-device AI is not impossible, but it is a choice - not a requirement. The architecture allows for a one-time purchase or an open-source model in a way that cloud AI fundamentally cannot support. - ---- - -## The context problem - -There is a subtler structural issue specific to personal AI. - -A cloud AI assistant gets better for you as it learns your context. But collecting your context - your messages, health data, location history - at scale creates an asset that is worth money to people other than you. - -An AI product that has collected the full personal context of millions of users has something extraordinarily valuable: a detailed model of how those people think, what they care about, how they spend their time and money. Even with the best intentions, that asset exists, and it creates incentives and vulnerabilities that on-device AI does not. - -On-device AI has no aggregate context asset. The data is distributed across individual devices. There is nothing to monetise, sell, or lose in a breach. The architecture eliminates the asset - and with it, the incentives and vulnerabilities that come with holding it. - ---- - -## What changes with on-device - -On-device AI is not cloud AI minus the privacy risks. It's a different architecture with different properties. - -Latency drops to zero - inference is local. Availability improves - the model works on a plane, in a tunnel, in a dead zone. Context can be richer - local data sources that would never be sent to a cloud service (your full message history, your local files, your health data) are accessible to the model. - -The privacy guarantee is structural - not "we promise not to misuse your data" but "the data never left your device." The continuity guarantee is structural - your AI survives any change to the vendor's situation. - -These are not marginal improvements. They are different properties that the cloud architecture cannot provide. - ---- - -## The objection - -The obvious objection is capability. Cloud models are large. They were trained on more data with more compute than can be replicated on-device. They can do things local models cannot. - -This is true today and was more true two years ago. The gap is closing faster than most people expect. - -Models like Qwen 3.5, Gemma 4, and Phi-4 Mini run on current phones at 20-30 tokens per second. For the tasks that define personal AI - context-aware assistance, summarisation, drafting, search over your own data - the quality difference between a capable local model and a large cloud model is already small and getting smaller. - -The capability argument for cloud AI weakens with every model release. The structural arguments against it don't change. - ---- - -*Off Grid runs on-device. No cloud. No subscription required. [Download for iPhone](https://apps.apple.com/us/app/off-grid-local-ai/id6759299882?utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing) or [Android](https://play.google.com/store/apps/details?id=ai.offgridmobile&utm_source=offgrid-docs&utm_medium=website&utm_campaign=writing).*