Independent-Researcher
Never work for any country, any orgnization, any company, any individual.
Never be used for 'Beijing China'.
Just for my dream: approach to the demiurge, through super-intelligence.
I have resigned from my last company, financial independence, and full-time for super-ai researching and implementation.
I will record my key insights in the process of researching and implementation to super intelligence in Chinese. (If you need to read, please translate them into your own language.)
The first version of the ultimate visual model is completed.
input: single-image, camera-free/train&infer, mask-support, incremental-hash-priori (so-far: single-object)
output: explicit-stereo-representation (only: stereo; todo: dynamic and interactive.)
With accompanying features: incremental/online learning.
2024-11-21: The 1st version of the non-BP learning (block-independent-leaning, w/ or w/o target). Effitive! Tested on: MLP, dimension-reduction, classification.
2024-12-14: The 2nd version of the non-BP learning (block-independent-leaning, w/ or w/o target). Effitive! Tested on: MLP, regression (difficult-task-type, a little different with classification), generation should be easy. Not implement on larget attention/transformer network and cifar10/100 or other big dataset task.
2025-03-01 to 2025-08-31: In the summer of 2025, collaborated with Justin(https://github.com/yuenuting) to develop the Neural Octree Mesh representation algorithm, addressing highly detailed very large scale 3D representations and enabling neural network learnable.
2025-09-01 to 2025-09-21: Completed the first version of the 'ASI' program, capable of confidently handling tasks such as ARC-AGI, Math Solver Tasks (SAT/IP/MIP/...), Sudoku, and more. (arcprize/hierarchical-reasoning-model-analysis#2)
Conclusion: "There is NO such thing as so-called AGI/ASI" -- what truly matters is diligent, systematic task-solving. However, problems can be abstracted into common solution frameworks and prepared system; yet, both learning and searching remain indispensable. The common components include but not limited:
Problem representation
Transformation of problem solution space (compression/abstraction, decomposition)
General representation of solutions (one-shot function) and structured representation (directed graph)
Automatic construction of atomic solution functions
Brute-force search based on atomic solution functions, and
Heuristic search guided by prior learning as state-action mappings (generalized reasoning: DG{A|S}, ...)
...
2025-10-31，complete the ARC-AGI task solving, including algorithm and code. detail pls refer to : arcprize/hierarchical-reasoning-model-analysis#2
浪费了长长的光阴，得到的痛苦的教训，关于人工智能的误解。 2010特别是2012深度学习爆发以来，作为一个洞察事物本质缺乏深度的人，很容易受到行业里面辉煌的东西（比如DL@Vision#CNN/Transoformer/xNN@ImageNet, DL@Symbol#LLM/GPT, ...的蓬勃发展)的东西影响。 在灿烂的烟花下，往往忽略了对事物本质的持续洞察，记录下2025年末，我得到的几个苦涩的教训：
- AI任务不是“函数拟合问题”，而是“算子在受约束范畴中的构造问题”。 AI Task: T=(X,Y,S,L,C, ...), S是Latent Space，好的S的构造有助于“智能”的解决问题，我们应该弱化L/Loss，Loss可能是难以完美构造；我们应该强化C约束的构造（含学习）。
- 约束很重要，约束很重要，对于本身是强约束的任务，如果只是一味的面向Y/target的优化，得到的东西终究只是奇葩。比如LLM本身对应的任务本身应该是强约束的，所谓幻觉只是约束不够。 行业里面喧嚣的World Model，其实是不可能完美构造的，最终的Validator也很重要。世界模型被提及但缺乏有深度的理解。 面向优化拟合而设计的损失（广义的评估）是受限的。关于如何系统的表示和构造约束是空白。
- 目前为止的几乎所有的AI算法，更多的强调了学习Learn/Train（先验），忽略了搜索/Search（除了MuZero类是有比较强搜索的）。
- 个人的一个理解，ASI面对的是可计算性的墙，而AGI的墙，可能需要一种“Language++的表征与操作体系”才能完成。现在还是空白。自然语言不够，纯数学的表达也不够，程序语言只有具体实例一层缺乏层次化的抽象。
- 当前人工智能领域，其实缺乏真正的大师了，像图灵那样的人。 能够一眼洞穿这些问题的本质，能够高瞻远瞩，在50年100年后，见解还足够深刻和正确的。 工程和算法上的繁荣是不够的。 总体来说，当前的热闹，还是回避了很多不应该回避的东西，迟早要面对： 本体论（ontology） 约束的第一性 不可学习性 假设与否证 ...等等 还是要面对: Gödel 不完备性定理 Church–Turing thesis Rice 定理 程序等价性不可判定 Proof vs Verification 的非对称性 ...等等中的一些基本问题
(not generative direction; Fig.2 strcuture/algorithm is still on developing ...)
2022-08-28 My thoughts, under LeCun's Paper on OpenReview: https://openreview.net/forum?id=BZ5a1r-kVsf¬eId=8g5X9wi4HX
(Chinese)
(English)
/physical/vision/interactive-dynamic-stereo (https://github.com/yuedajiong/super-ai-vision)
/symbol/thinker/unified-thinker(mathink & math-solver & arc-agi) (https://github.com/yuedajiong/super-ai-symbol)
