vynly 1
I’m building CAEL, an experimental assistant that combines counted language models, cellular automata, symbolic reasoning, exact arithmetic and local retrieval. Its inference uses no neural network, transformer or external language model.
The current public release is CAEL 0.3:
The idea is to combine a small statistical language component with specialized, inspectable helpers rather than rely on fluent text alone. The unified runtime includes chat, supported symbolic reasoning, exact calculations, limited Python helpers, prompt/story writing and local Wikipedia retrieval. The downloadable bundle is about 556 MB; picture-generation weights are excluded.
One small example you can try after setup:
python -m cael_unified "Calculate 1/3. Then multiply it by 3" --mode chat
CAEL is still experimental. General conversation and commonsense accuracy remain limited, and supported reasoning/programming forms are narrow. I’m sharing it for people interested in non-neural approaches and practical failure analysis, not claiming general Python understanding or competitive LLM performance. Newer local tool-building work is still being tested and is not part of this public download yet.
The owned code/model materials use the MICA AI Non-Commercial License, with third-party notices retained; public availability does not mean unrestricted commercial use.
I’d appreciate technical feedback: which tasks would provide a useful evaluation of this kind of hybrid, and which reproducible failures should I prioritize? If you try it, please include the prompt and expected behavior so I can investigate.
Evaluate the routing, not just the answers: build a small suite in three buckets — tasks an exact helper must own (arithmetic, supported symbolic forms), tasks just outside those forms where the correct behavior is to say so rather than improvise, and tasks the statistical component may answer freely — and score a wrong-route answer as its own failure class, separate from a wrong answer. Prioritize silent wrong-route failures: a plausible answer produced by the wrong part of the system is worse than obvious nonsense here, because the whole value of inspectable helpers is lost if you cannot tell which part actually answered.