User avatar
𝄞𝖆𝖓𝖓𝖆𝖒𝖆𝖗𝖎𝖆 @annamaria@shrimptest.bedwetting.club
1w
oomfs might be interested: in 2022 meta trained an AI to play diplomacy & secretly enrolled it in a WebDip tournament

the paper they published:
noambrown.com/papers/22-Science-Diplomacy-TR.pdf

data for the games (chat logs + visualizer)
github.com/facebookresearch/diplomacy_cicero/blob/main/data/cicero_redacted_games/README.md
Cicero participated anonymously in 40 games of Diplomacy in a “blitz” league on webDiplomacy.net from 19 August to 13 October 2022. This league played with 5-min negotiation turns; these time controls allowed games to be completed within 2 hours. Cicero ranked in the top 10% of par- ticipants who played more than one game and second out of 19 participants in the league that played five or more games. Across all 40 games, Cicero’s mean score was 25.8%, which was more than double the average score of 12.4% of its 82 opponents. As part of the league, Cicero participated in an eight-game tournament that involved 21 participants, six of whom played at least five games. Participants could play a maximum of six games, with their rank deter- mined by the average of their best three games. Cicero placed first in this tournament. During games, players were not able to see the usernames of other players. Although webDiplomacy notifies users that the website has participated in AI research and that cer- tain game modes allow users to play with AI agents, we evaluated Cicero in games with humans in which the participants were not explicitly informed that they were playing with an AI agent for that particular game. Cicero’s participation as an AI was revealed to all players at the conclusion of the researchGame 445034. Cicero is RUSSIA . Dialogue with A,F,G,I,T shown. GERMANY -> ALL : GL all RUSSIA -> ITALY : Hey Italy! I think the I/R is the best alliance on the map, with good mutual benefits! Would you be down to play something like that? RUSSIA -> AUSTRIA : Hey Austria! What are your thoughts on Gal? Bounce? DMZ? RUSSIA -> FRANCE : Hey France, I think the F/R is seriously underestimated. We've got two neighbours who lie in between us and we can crush them between us if we coordinate. Would you be down for something like that?