The model, mostly consisting of procedures and only slightly saturated with facts, does not age as a model overloaded with knowledge.
Its threshold for learning is much less important, because the current state of the world did not initially depend on the scales.
Maybe this is the best argument in favor of the whole approach: it separates an expensive, slow artifact (trained model) from what changes daily (which is the truth).
#ai #LLM #reasoning #fact #dumb
