The latest AI development is: AI achieves human level in (blitz 5-minute-turn) full-communication anonymous online Diplomacy (paper). Why not? I mean, aside from the obvious. A take I saw multiple times was that AI labs, or at least Meta, were intentionally going for the scariest possible thing,
If anything, I'd rather we played with Diplomacy-like AI while it's in infancy and we can learn a bunch of stuff. The alternative is to have all the mature tech and BAM, somebody applies it the right way.
I'm an author on the paper. I want to point out what I think are a few mistakes with the blog.
First, Zvi says "The strategic engine, as I evaluated it based on a sample game with six bots and a human, is mediocre at tactics and lousy at strategy." This is not the feedback we have gotten from expert Diplomacy players. The general consensus among expert Diplomacy players is that the strategy/tactics are extremely strong, perhaps expert level, but that there is room for improvement in the dialogue. I'm not familiar with Zvi's experience with the game of Diplomacy, but I did place 3rd in the North American championship this year and have learned a lot about the game of Diplomacy, including how humans play it, over the past 3 years, so I feel somewhat qualified to comment on Diplomacy strategy.
In particular, I disagree with Zvi's opinion on the bot's strategy. Zvi says "I hate France’s tactical play, both its actual plays and the communications with Russia that are based on its tactics, dating back to at least 1903. The move here to Irish Sea needs to be accompanied by a convoy of Picardy into London or Wales, fighting for Belgium here is silly."
It's well-known among experienced Diplomacy players that France needs 3 armies on its mainland in order to defend itself well from a hostile Germany. Moreover, a convoy to London or Wales is unnecessary here. By moving to Irish Sea, France is setting themselves up for the option to convoy directly into Liverpool next turn. England can't block it because they have no army on their mainland. In short, I feel pretty confident saying that France made the right move here. If Zvi still disagrees, we could pull in some consensus expert Diplomacy players to get their opinions on this.
Second, we did a 200-game tournament for no-press (that is, no-dialogue) Diplomacy back in January where players were informed that one of the players in each game was a bot. Our bot, Diplodocus, placed first in this tournament. There's a video on it here: https://www.youtube.com/watch?v=AWQFhYSD7h4&ab_channel=DiploStrats . You can see in the Youtube comments that one expert Diplomacy player (Sploack) described the bot as "currently the best gunboat player in existence, or at least in the top 5." The strategy/tactics in no-press Diplomacy don't match up perfectly with full-press, but they do carry over to some extent.
Third, Zvi criticizes the fixed 1908 end date as working to the bot's advantage. The no-press tournament mentioned above did not have fixed end dates. Also, I think the fact that the full-press games ended in 1908 rather than a later date (say, 1910) actually hurt the bot. The bot handles endgame tactics quite well.
Agreeing with Daphne_W at the LW discussion, my main update on this is to reinforce my belief that human intelligence is a much lower bar than it seems, because most people are operating at a low cognitive level for most tasks, most of the time. We don't need uniformly expert level performance from an AI to pass as human-equivalent, but we do need the General part. Training special-purpose bots therefore seems to be shooting fish in a barrel: we should expect AI to do well at this when the task is well defined enough for optimization approaches to work well. We instead need to assess progress on how AI agents do at generalist tasks where the objectives are fuzzy and feedback is indirect and noisy. Agreeing with Zvi, I'm not going to worry about AGI more as a result of this work.
If anything, I'd rather we played with Diplomacy-like AI while it's in infancy and we can learn a bunch of stuff. The alternative is to have all the mature tech and BAM, somebody applies it the right way.
I'm an author on the paper. I want to point out what I think are a few mistakes with the blog.
First, Zvi says "The strategic engine, as I evaluated it based on a sample game with six bots and a human, is mediocre at tactics and lousy at strategy." This is not the feedback we have gotten from expert Diplomacy players. The general consensus among expert Diplomacy players is that the strategy/tactics are extremely strong, perhaps expert level, but that there is room for improvement in the dialogue. I'm not familiar with Zvi's experience with the game of Diplomacy, but I did place 3rd in the North American championship this year and have learned a lot about the game of Diplomacy, including how humans play it, over the past 3 years, so I feel somewhat qualified to comment on Diplomacy strategy.
In particular, I disagree with Zvi's opinion on the bot's strategy. Zvi says "I hate France’s tactical play, both its actual plays and the communications with Russia that are based on its tactics, dating back to at least 1903. The move here to Irish Sea needs to be accompanied by a convoy of Picardy into London or Wales, fighting for Belgium here is silly."
It's well-known among experienced Diplomacy players that France needs 3 armies on its mainland in order to defend itself well from a hostile Germany. Moreover, a convoy to London or Wales is unnecessary here. By moving to Irish Sea, France is setting themselves up for the option to convoy directly into Liverpool next turn. England can't block it because they have no army on their mainland. In short, I feel pretty confident saying that France made the right move here. If Zvi still disagrees, we could pull in some consensus expert Diplomacy players to get their opinions on this.
Second, we did a 200-game tournament for no-press (that is, no-dialogue) Diplomacy back in January where players were informed that one of the players in each game was a bot. Our bot, Diplodocus, placed first in this tournament. There's a video on it here: https://www.youtube.com/watch?v=AWQFhYSD7h4&ab_channel=DiploStrats . You can see in the Youtube comments that one expert Diplomacy player (Sploack) described the bot as "currently the best gunboat player in existence, or at least in the top 5." The strategy/tactics in no-press Diplomacy don't match up perfectly with full-press, but they do carry over to some extent.
Third, Zvi criticizes the fixed 1908 end date as working to the bot's advantage. The no-press tournament mentioned above did not have fixed end dates. Also, I think the fact that the full-press games ended in 1908 rather than a later date (say, 1910) actually hurt the bot. The bot handles endgame tactics quite well.
Agreeing with Daphne_W at the LW discussion, my main update on this is to reinforce my belief that human intelligence is a much lower bar than it seems, because most people are operating at a low cognitive level for most tasks, most of the time. We don't need uniformly expert level performance from an AI to pass as human-equivalent, but we do need the General part. Training special-purpose bots therefore seems to be shooting fish in a barrel: we should expect AI to do well at this when the task is well defined enough for optimization approaches to work well. We instead need to assess progress on how AI agents do at generalist tasks where the objectives are fuzzy and feedback is indirect and noisy. Agreeing with Zvi, I'm not going to worry about AGI more as a result of this work.
Published just yesterday (06 Dec) Noam Brown and Lex Fridman had a characteristically deep, thorough conversation:
Noam Brown: AI vs Humans in Poker and Games of Strategic Negotiation:
https://www.youtube.com/watch?v=2oHH4aClJQs
OUTLINE:
0:00 - Introduction
1:09 - No Limit Texas Hold 'em
5:02 - Solving poker
18:12 - Poker vs Chess
24:50 - AI playing poker
58:18 - Heads-up vs Multi-way poker
1:09:08 - Greatest poker player of all time
1:12:42 - Diplomacy game
1:22:33 - AI negotiating with humans
2:04:58 - AI in geopolitics
2:09:43 - Human-like AI for games
2:15:44 - Ethics of AI
2:19:57 - AGI
2:23:57 - Advice to beginners