11 ms·
From the article: > You could also say that GPT-4 was playing the game blindfolded as it had no access to refreshing its memory of the current state of the boa
by cjbprime 4y ago
From the article:
> You could also say that GPT-4 was playing the game blindfolded as it had no access to refreshing its memory of the current state of the board.
That doesn't sound right. Under the hood, I think everything inside the current conversation is just appended to the prompt each time you add a message, and the whole conversation re-evaluated? Or is there a limit when the conversation gets long which means that the LLM forgets how the conversation started..?
- anonzzzies 4y agoIt has limited token length; I think the chatgpt one has 8k, but they have a version with 32k now as well. That includes input and output in that length. It will ‘forget’ the start of the conversation if the conversation is too long compared to that length. They maybe use tricks like summerization, but that won’t help with chess or other precise conversations.
- nullc 4y ago> but that won’t help with chess or other precise conversations. Not necessarily true: the summerization might happen to give the board state! Certainly you can come up with an input that would make it summarize that way, though getting that outcome from a generic summerization prompt would be more surprising.
- anonzzzies 4y agoYes, certainly if you prompt it to do so. Or pass the board efficiently. I just don’t think it’s set up like that currently seeing it’s amnesia of the positions.