LLMs are not general--they're only general with respect to language. It's surprising to me that they are as okay as they are at e.g. chess (I had GPT4.5 beat my ELO 1000 cousin in a chess game this last week, but I was reminding it of every move taken every single turn, and once warned it that its reasoning was wrong)... but language is a very inefficient way of thinking about something geometric like chess or Pokemon. On the other hand, a reinforcement-learning-based AI like MuZero is (as I understand) general with respect to untimed games that might have hidden information, but you would need to feed it information about the game state in a MuZero-native format. If someone wrote devised an LLM that could recognize when it was playing a game with definite rules and call MuZero as a subroutine, would that be "general" for you?
>If someone wrote devised an LLM that could recognize when it was playing a game with definite rules and call MuZero as a subroutine, would that be "general" for you?
Probably. I would have to see what the end product looked like to be sure, but I see no reason why daisy-chaining AIs together should be off limits. Granted, ChatGPT would need to examine the game itself and translate the gamestate to MuZero format itself. But if it can do that, then yeah, why not?
LLMs are not general--they're only general with respect to language. It's surprising to me that they are as okay as they are at e.g. chess (I had GPT4.5 beat my ELO 1000 cousin in a chess game this last week, but I was reminding it of every move taken every single turn, and once warned it that its reasoning was wrong)... but language is a very inefficient way of thinking about something geometric like chess or Pokemon. On the other hand, a reinforcement-learning-based AI like MuZero is (as I understand) general with respect to untimed games that might have hidden information, but you would need to feed it information about the game state in a MuZero-native format. If someone wrote devised an LLM that could recognize when it was playing a game with definite rules and call MuZero as a subroutine, would that be "general" for you?
>If someone wrote devised an LLM that could recognize when it was playing a game with definite rules and call MuZero as a subroutine, would that be "general" for you?
Probably. I would have to see what the end product looked like to be sure, but I see no reason why daisy-chaining AIs together should be off limits. Granted, ChatGPT would need to examine the game itself and translate the gamestate to MuZero format itself. But if it can do that, then yeah, why not?