Thursday, January 1, 2026

AI progress

It remains to be seen if transformer neural networks (LLMs) can learn (approximate) the human cognitive architecture.  Even if they could, humans are not, themselves, rational* nor do they possess a sound value system**.  Noorbakhsh Amiri Golilarz et al have a good discussion*** of some of the things that are missing with LLMs. With my A.s.a. H. architecture I have tried to address the shortcomings Golilarz et al outline. Self-monitoring, allocation of resources, and meta-cognitive awareness are discussed in my "Experiments with Asa H" in my book Twelve Papers.**** The learned case-base (knowledge representation/concept structure/memory) is corrected/repaired, grows, updates, and evolves over time. Deployed on robots Asa is fully embodied (sensorimotor grounding). The hierarchical case-base is distributed across multiple timescales.

* see Predictably Irrational, Dan Ariely, Harper, 2008.

** We are, after all, apes. (Gentle, very modern apes as Erika would say.)

*** Bridging the Gap: Toward Cognitive Autonomy in Artificial Intelligence, arXiv:2512.02280v1, 1 Dec. 2025.

**** Available under "book" at www.robert-w-jones.com

No comments:

Post a Comment