Comment by Jensson

The best way to describe the LLM intelligence is "an expert system that works the way people thought expert systems would work". You can encode a massive amount of skills into an LLM, and then the LLM uses those to navigate problems.

But the LLM is still dumb where those skills doesn't have good coverage, since unlike the expert systems it maps fuzzily to its skills, and they are tuned to produce results over rejecting the request when its unclear if coverage is good. As long as that is true you have to treat them as dumb even if they sometimes produce brilliant results.