From my limited understanding, a smaller, more specialized model could produce more accurate responses, but the main thesis of the article would still be true.

From my limited understanding, a smaller, more specialized model could produce more accurate responses, but the main thesis of the article would still be true. Such LLMs would still be statistic prediction models, with no concept of reality. We already know that specialist systems, such as those tried in the 80's are not very good in modeling reality, either. The thing is, using LLMs as knowledge bases or as a sort of modern oracles seems to me like using a powerful tool for something they are horrible at. They are good enough (but not perfect) text summarizers, text translators and text transformers, but are far from the so expected "AGI" some seem to be expecting them to be.