Will general purpose AI models beat average score of human players in Diplomacy by 2028?
Plus
10
Ṁ11712027
60%
chance
1D
1W
1M
ALL
General purpose (not trained for a specific task) language models demonstrated chess playing ability. They are also capable of deception and lie detection. Will language models or visual-language models* beat the average score of human players during a series of 40 games on webDiplomacy.net by 2028? (question modeled after Meta's Cicero result).
[EDIT: Please notice that while "CICERO achieved more than 2x the average score of its opponents" this question requires only achieving the above-average score]
*models or agents trained on different modalities (so e. g. models capable of controlling robotic arm like PaLM-E) would also qualify as long as they weren't trained specifically to play Diplomacy
This question is managed and resolved by Manifold.
Get
1,000
and3.00
Related questions
Related questions
Will a large language model beat a super grandmaster playing chess by 2028?
59% chance
Will an AI achieve >85% performance on the FrontierMath benchmark before 2028?
53% chance
In 2028, will an AI be able to play randomly selected computer games at human level without getting to practice?
58% chance
Will AI beat the best humans in competitive programming before the end of 2024?
7% chance
Will any OpenAI model win a chess match against IM by the end of 2024?
8% chance
Will AI surpass human intellect by 2030?
92% chance
Will an AI be capable of achieving a perfect score on the Putnam exam before 2028?
66% chance
Will an AI score 1st place on International Math Olympiad (IMO) 2025?
26% chance
Will an AI by OpenAI beat a super grandmaster playing chess by 2028?
69% chance
In 2028, will an AI be able to play randomly-selected computer games at human level, given the chance to train via self-play?
76% chance