Epistemic status: This is an off-the-cuff question.

~5 years ago there was a lot of exciting progress on game playing through reinforcement learning (RL). Now we have basically switched paradigms, pretraining massive LLMs on ~the internet and then apparently doing some really trivial unsophisticated RL on top of that - this is successful and highly popular because interacting with LLMs is pretty awesome (at least if you haven't done it before) and they "feel" a lot more like A.G.I. Probably there's somewhat more commercial use as well via code completion (and some would say many other tasks, personally not really convinced - generative image/video models will certainly be profitable though). There's also a sense in which they are clearly more general - e.g. one RL algorithm may learn many games but there's typically an instance per game not one integrated agent. You can just ask an LLM in context to play some games.

However, I've been following moderately closely and I can't seem to think of any examples where LLMs really pushed the state of the art in narrow game playing  - how much have LLMs contributed to RL research? For instance, will adding o3 to the stack easily stomp on previous Starcraft / go / chess agents?