@MetaCircle_AI

MetaCircle/元环智能, AI for AI, Science of AI, Mechanistic AutoResearch

meta-circle.com/
Joined July 2026
  • Tweets18
  • Following4
  • Followers216
13Media
Over-encoding, a memory trick akin to DeepSeek v4.1’s Engram, really works—but it has a failure mode. We call it Forking: under data replay, train loss keeps falling while validation loss turns up. The two curves split like a fork. Watch the explainer below 👇 @ZimingLiu11
2
2
1
10
2,932
more replies
3/4 Post-training replays data too, and Forking comes back on small code, math, and RLVR-style corpora. Freezing the n-gram table helps.
1
103
4/4 The rarer the context, the larger the gap—roughly a power law in frequency (our paper explains why). Masking low-frequency n-gram reads shrinks the gap. SCONE already works this way: its BPE-style mining gives memory rows only to frequent n-grams.
90
We quantified LLMs’ training intuition—and revealed its blind spots. ArchitectureIQ: 500 questions. One synthetic dataset, three recipes per question. Answers verified by training runs. GPT-5.6 Sol: 76.4% accuracy (chance: 33.3%). Watch the explainer below👇 @ZimingLiu11
1
1
1
6
2,451
more replies
3/4 Models invoke training physics: cumulative effective learning rate U (Adam ≈ ηT, Adagrad ≈ 2η√T), NTK curvature, and zero-momentum SGD failures. A readable decision-tree policy reaches 78.2% out-of-fold accuracy: intuition compresses into rules.
1
50
4/4 We use Opus-5 to build an evidence-weighted knowledge base. With up to 20 principles, GPT-4o's accuracy rises from 53.2% to 67.6% on 50 held-out questions, without weight updates. Experience becomes transferable text—still incremental, not yet a science of AI.
41
Staring from 2022, @ZimingLiu11 has the habit of MANUALLY browsing and collecting fun arXiv papers, gathering ~2000 papers in total. Today we release this paper list, and test if an LLM agent could learn my taste simply from this paper list. meta-circle.com/blog/can-tas…
🤖 Made with AI
2
4
627
We open-source our AI experiment workbench, great for exploring "Science of AI"! We deem it as a playground where it's easy to compose experiments and observe new phenomena. We believe that observing more phenomena is the first step towards a "Science of AI".
Ziming Liu
@ZimingLiu11Aug 11
Bored of Jupyter Notebooks or Python files? Introducing ComfyResearch -- a new, graphical way to visualize AI experiments. With ComfyResearch, "Science of AI" becomes super easy and fun, like building LEGO. Infinite worlds, infinite phenomena! Our blog: meta-circle.com/blog/comfyre…
1
403
Bored of Jupyter Notebooks or Python files? Introducing ComfyResearch -- a new, graphical way to visualize AI experiments. With ComfyResearch, "Science of AI" becomes super easy and fun, like building LEGO. Infinite worlds, infinite phenomena! Our blog: meta-circle.com/blog/comfyre…
1
1
6
4,649
Github: github.com/MetaCircleAI/Comf… Try Online: online.comfy-research.com/

GitHub - MetaCircleAI/ComfyResearch: A platform for structuring AI research and studying physics of...

A platform for structuring AI research and studying physics of AI - MetaCircleAI/ComfyResearch

github.com
154
Current Auto-Research systems are Shakespeare's monkeys, typing randomly, hoping brilliance emerge from chaos. Today, we @MetaCircle_AI propose "Mechanistic Auto-Research", not using any LLMs, but surpassing RSI's results. Scale insight, not compute. meta-circle.com/blog/ophis-a…
2
429
At @MetaCircle_AI (元环智能), we work on mechanistic autoresearch, we believe that understanding and structure should precede naive brute-force automation. In order for AI4AI systems to deliver next "transformers", we should: Scale insight and understanding, not (just) compute
2
1
10
6,844
Load more