AI models can only learn during their batch training runs not from daily interactions with users. Session memory isn’t the same as actual learning.
There’s also no core “truth” layer in these systems: no deterministic backbone, no real understanding of concepts, and no explicit dictionary or knowledge store they can reference, cross-check, or update.
A dynamic knowledge graph could help fix a lot of this. It would lower hallucinations and improve performance in high-stakes fields like medicine, law, physics, and chemistry. It could also reduce the number of vector embeddings needed for complex LLMs.
Do you agree? Or is there a better path forward?
[link] [comments]