Non-neural AI assistant CAEL grows from 77 KB to 1.4 GB, still loses 76–126 to Qwen 0.5B
Silver_Employ2617 · reddit · 2026-10-22
A Reddit user presented CAEL (Cellular Automaton Event Language), an AI assistant with no transformers or neural networks at inference time.
- The first version was just 77 KB of learned model data: integer cellular automata, learned event transitions and a small grammar system. It could generate simple stories, track characters/objects, and scored 13,029/13,029 on supported factual questions in a closed-domain synthetic story eval.
- Later versions added modular engines — symbolic logic, statistical language models, retrieval, deterministic algorithms — with math solving, contradiction detection, code generation/repair, persistent rule memory and an orchestrator chaining them, all running locally on CPU. The unified bundle is now 1.4 GB.
- Benchmark vs Qwen2.5-0.5B-Instruct on 174 locked questions: CAEL 0.2.8 got 67, the dev branch 76, Qwen 126. Ties on logic (5/5) and code (7/12); large gaps in math, multi-step reasoning and writing.
The author is explicit this is a closed-domain experiment, not evidence of general language understanding.
More from Research
- Microsoft Research's PDD Kit formally proves AI-generated code meets your intent — omojumiller · 2026-10-22
- LAGEN: one frame of latency can cut real-time agent performance by 95% — wzenus · 2026-10-22
- Angler: small sequence model challenges larger models on RNA splicing and variant effect prediction — anshulkundaje · 2026-10-22
- Oopsie Data: multi-lab effort to build first large-scale robot failure dataset — micoolcho · 2026-10-22
- Terence Tao's "Math 2.0" lecture slides map how humans plus AI expand the research frontier — tak3sh8 · 2026-10-22
- JapanFold launches: low-cost AI drug discovery on Tenstorrent chips at $0.26/hour — DavidBennett__ · 2026-10-22