FullPressGPT

Developers and contributors can find a link to our github page and engage in development project planning here.
Post Reply
Message
Author
teacon7
Posts: 1253
Joined: Fri Oct 06, 2017 7:48 pm
Location: Midwestern USA
Karma: 141
Contact:

FullPressGPT

#1 Post by teacon7 » 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.
kestasjk wrote:
Tue Apr 11, 2023 9:04 am
... first priority re: bots is to get a full-press bot working, and to add difficulty levels to classic.
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).
kestasjk wrote:
Tue Apr 11, 2023 9:04 am
Meta 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)
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?

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? :-)

jamesa7171
Posts: 28
Joined: Sat Jun 02, 2018 4:50 am
Karma: 27
Contact:

Re: FullPressGPT

#2 Post by jamesa7171 » Mon Apr 17, 2023 8:52 pm

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.
You might be interested in this if you haven't seen it:

https://dl.fbaipublicfiles.com/diplomac ... ND_AG.html

It's a partial press transcript from a game involving Meta's bot Cicero (playing as England) against human opponents. Austria and Germany both consented to releasing their conversations with Cicero throughout the game, so you can see for yourself how it incorporates outside information.

User avatar
JECE
Posts: 1478
Joined: Sun Dec 31, 2017 4:35 pm
Karma: 393
Contact:

Re: FullPressGPT

#3 Post by JECE » Tue Apr 18, 2023 12:05 am

1
See my full Profile:
http://www.webdiplomacy.net/profile.php?userID=17421

User avatar
kestasjk
Developer
Developer
Posts: 518
Joined: Tue Mar 14, 2017 8:13 pm
Location: Perth, Australia
Karma: 775
Contact:

Re: FullPressGPT

#4 Post by kestasjk » Tue Aug 29, 2023 7:10 pm

Hi Teacon; yep some of what you're critcising / suggesting is way out of take compared to the latest AI developments. I would suggest reading the articles on the recent bot developments in Science, but I'm hoping that with a bit of luck we'll have full press bots running here at a level that's ready for end users to enjoy playing against, and people will get a much better understanding of the strengths / weaknesses / next steps from that.

Post Reply

Who is online

Users browsing this forum: No registered users and 35 guests