CAEL 0.4.1: a chat assistant with no neural network and no transformer

A hobbyist developer released CAEL 0.4.1, a chat assistant built in plain Python with no neural network, transformer or LLM weights, now published on Hugging Face. On five Open SLM Leaderboard tests run locally, CAEL scored a 49.97% mean versus SmolLM2-360M's 58.86%, ahead on ArithMark-3 (69.5% vs 59.6%) but far behind on HellaSwag and PIQA, and in blind chat rounds against Qwen3.5-0.8B (set to 1000 Elo) CAEL scored about 796. The system combines symbolic, count-based and cellular-automaton engines with an index of about 285,000 human-written answers from Stack Exchange, OpenAssistant and databricks-dolly-15k, and is source-available for non-commercial use; the developer notes it is weak on open advice, how-to questions and creative writing.

Hi all. I’d like to share CAEL, a hobby project: a chat assistant built without any neural network. No transformer, no neural weights, no LLM behind the replies. It runs on a CPU in plain Python. How it works - Symbolic, count-based and cellular-automaton engines for reasoning and state. - Exact skills for conversation: quantities, rates, clock times, syllogisms, schedules and changed values. - Chat memory: it answers from what you said earlier in the conversation. - Checked web answers: when it doesn’t know, it searches and answers only if a second independent source agrees. - An index of about 285,000 human-written answers Stack Exchange, OpenAssistant, databricks-dolly-15k for open questions. Honest results - On the five Open SLM Leaderboard tests, run locally: 49.97% mean against SmolLM2-360M’s 58.86%. It is ahead on ArithMark-3 69.5% vs 59.6% and far behind on HellaSwag and PIQA. This is a local comparison and not a leaderboard entry, since CAEL uses offline Wikipedia and the ARC Corpus while answering. - In blind chat rounds against Qwen3.5-0.8B set to 1000 Elo , CAEL scores about 796. It does not match a small LLM at general chat. I judged these rounds myself, so treat them as a rough guide. Where it is weak: open advice, how-to questions and creative writing. It often declines to answer where an LLM would say something plausible. Links It is source-available for non-commercial use, and it was built with AI assistance. I’d welcome feedback, especially ideas for non-neural approaches to open-ended chat. Edit: 0.4.1 is now published on Hugging Face; the link above replaces the earlier 0.4.0 model card link.