FullPressGPT
Posted: Mon Apr 17, 2023 2:57 pm
This is a discussion and wondering, not a request or nag, hoping to hear from people who know about such things.
My non-CS background take: Current generations of language models seem like they could translate the intentions of an AI back and forth into human-readable press. What seems like a necessary new step here is how to get a bot to change its mind based on receiving press from the outside. It would have to have the game-mechanics AI output a 100% confidence stream of press to itself to tell it what it thinks is a good idea, then use language model tools to assess motivations and trustworthiness based on move history, press history, and a comparison of two. Then it would have to assess whether it wanted to accept suggested moves, or propose its own, and communicate those back. That seems like an entirely different generation of strategy than simply determining what's the best data-driven move for power X given board position Y.
Is any of this remotely close to how it works? Is there a better place to go for discussion of AI + diplomacy?
Imagine:
How would a full press bot work... what internal flow of inputs would allow that to happen? I remember playing diplomacy against the Albert bots wayyy back in the days of DAIDE. There was a rudimentary sort of press there - you could propose moves for the bot to do, and it would accept or reject them. It couldn't parse text so you'd go through the tree of orders (like this: 'I suggest' > F > Sev > H).
Did meta's bots actually interact with one another in the sense that they changed each other's minds and set up strategies, etc? Where can I learn more about that?kestasjk wrote: ↑Tue Apr 11, 2023 9:04 amMeta trained a set of bots that could communicate and play with each other, and they invented their own incomprehensible language, and were great at playing against each other, great at playing against 1-2-3 humans, but were terrible at playing against 4(?)-5-6 humans. (Even gunboat bots followed this pattern)
My non-CS background take: Current generations of language models seem like they could translate the intentions of an AI back and forth into human-readable press. What seems like a necessary new step here is how to get a bot to change its mind based on receiving press from the outside. It would have to have the game-mechanics AI output a 100% confidence stream of press to itself to tell it what it thinks is a good idea, then use language model tools to assess motivations and trustworthiness based on move history, press history, and a comparison of two. Then it would have to assess whether it wanted to accept suggested moves, or propose its own, and communicate those back. That seems like an entirely different generation of strategy than simply determining what's the best data-driven move for power X given board position Y.
Is any of this remotely close to how it works? Is there a better place to go for discussion of AI + diplomacy?
Imagine:
diploGPT wrote: Hello! As an AI language model, It would be unethical for me to misrepresent anything or lie. Therefore I can let you know that your press has a high degree of 'pitiful pleas for mercy' rhetoric. I also notice a pattern of disregard for our previous agreements. I can also let you know that your suggestion has been considered and rejected, on the basis that it will in no way further my power's win conditions. I will be ordering A Tyr s Tri-Ven as planned. Please see the attached link to teletherapy sessions available in your healthcare network. Is there anything else I can do to assist you today?
