Well, the new models are certainly better at math and at vulnerability finding than previous models. My guess is that while information that gives human-language-based heuristics to understand the world, coding and so-forth is limited, information in the form of what allows the proof of what is unlimited and so LLMs will be moving in the (simulation of) "reasoning" direction at this point.

The final stage, I would guess, is having humans with data suits living their daily lives and giving the "AIs" the full information needed for a semi-complete simulation of "intelligence". Whether people will put up with that remains to be seen.

How much this will help models deal with the "real world" remains to be seen.