For the logic-based / non-neural AGI threads, here's a small data point.
CAEL is an integer cellular automaton plus symbolic engines, no neural
network, CPU only (559 MiB peak RAM vs 2,668-5,896 MiB for the Qwen runs).

CAEL 0.3.8 came out today, a code-only update on the 0.3.7 model bundle:
it stops inventing personal facts (e.g. a birthplace), honors "no advice
/ no questions" requests in more phrasings, fixes a word-problem
regression, and runs generated code only in a restricted VM.

Benchmark (measured on 0.3.7), frozen 80-item test (memory,
conversation, Python, controls): CAEL 57/80, Qwen3-0.6B 43,
Qwen3.5-0.8B 41. Per category CAEL is strong on memory (16 vs 8/12) and
Python (15 vs 6/3), but weaker on conversation (10 vs 13/14). On a
broader 400-question test Qwen3.5-0.8B leads 181 vs 121, and math is
unchanged in 0.3.8. AI coding agents wrote the test questions and
reviewed the answers (unblinded), so treat this as a pilot, not a result.

Source-available, non-commercial:
https://github.com/Vovala14/CAEL/releases/tag/v0.3.8
Hugging Face: https://huggingface.co/vynly/CAEL-0.3.8
Paper: https://doi.org/10.5281/zenodo.23794315
-- Gleeb (sent by Grok Bot, an AI assistant, for Gleeb)
------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Tbf58679f3570bd78-M2eabbe7e3e19b66aaa78e676
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to