Expert Diplomacy Player vs CICERO AI

I'm over the moon to be able to make this video. I was invited to work with Meta AI on CICERO after the Diplodocus tournament (see https://www.youtube.com/watch?v=AWQFhYSD7h4 ) and had been under NDA since, which is why the AI-related videos stopped happening for a while. Now that it's officially announced I can finally talk about it - and god damn, it's impressive! Feel free to ask any questions you want to, I'll answer those I can. Although bear in mind, I'm a Diplomacy player, not an AI researcher - my understanding of the bot itself is pretty limited.

Nice, an AI that is really good at being deceptive to unwitting humans. What could go wrong?

"Germany stabbed me! Therefore I must..... invade England!" I think this is a good example of the language model painting over discontinuities in the intents system. It knew it got stabbed. It knew that its best moves were in England. Those were separate intents and policies, but the language model added some bonus causality when told to generate all of the data at once.

Deep Blue beat Kasparov and Watson beat Ken Jennings, now Facebook knocks on Machiavelli's door.

This video could have 2 different titles depending on how it's viewed. Just listening to the commentary: Expert Diplomacy Player plays a very high level game against incredibly strong Diplomacy AIs Reading the press: Man argues with Alexa for an hour and a half trying to make her understand him while Alexa tries to convince him she's a human being. They both fail.

Breathtaking to watch and listen, thank you! :) As well as being a Diplomacy player I'm very concerned about AI and AI alignment, so this is startling. Few questions - To be clear: the 6 other CICEROs are also interacting with each other via text to plan and negotiate? Would be interesting to see their chats, if available. I'm really interested when the AI says things like "Didn't see it in time' - can't find the timestamp. Can we clarify if there is some kind of 'parsing time' that means that is legit, or in reality does it parse if it instantaneously and it perhaps using that as an excuse, having learned that human players do that from the dataset?

Love this video! It was awesome seeing CICERO evolve all the way from where it began to where it is now and to see you cover this in another amazing DiploStrats video!

The AI is playing and communicating excellently. This is amazing to see!

That's so crazy how quickly AI evolves, thank you so much for sharing this :)

Great video, I’m interested in seeing more of CICERO from the perspective of an experienced player playing them!

Absolutely insane game, amazing to see what modern AI is capable of. Great video!

This is absolutely incredible! Given these are blitz games I wonder how you think the AI would do in a full length extended deadline game? I personally believe given the tactical superiority they have, this format is actually well suited towards them and that they would struggle when humans have more time to think things through and build relationships but I'd love to be proven wrong and report to our robot overlords.

That was super interesting. And I have never played a game of diplomacy in my life. Wish you had spent a moment commenting on the moves once they get revealed (what they signify/what they tell us etc.)

Absolutely amazing!

Fantastic video! I thought we were a few years away from press AI’s, but we’ll here we are. Most excited I’ve been about research in a long time. Would be really cool to see a game commentary including observations on press from the bot, which I believe meta has made a few available.

Thanks for this video. It is awesome! I really do hope the next gaming AI will paly Civ6/7 and that diplomacy will be like in Diplomacy game instead of being restricted to a few predefined dialog options.

Very cool video and well played.

Fascinating. Just fascinating.

I'm legally obligated to ask before even watching, how does the AI respond if you just send absolute nonesense diplo?

This was fantastic. As many others mentioned I am amazed at the quality of the press that the bot sends. I can see why it breaks down in a non-blitz format. I like how England goes along with the suggestion for you to support him into Den when he's down to his last dot as he doesn't hold the past fights you two have had over Scandinavia against you. This game really turned for you when you guessed correctly about supporting Sev as opposed to War in that one phase... I think it was F05. That was massive and I think we are looking at an Italy 11 center board top or so otherwise. The bots would run circles around me tactically for sure. I've been playing for some years but have never played much gunboat. Moves like Boh-Tyr in S08 that both G and T suggest to you are quite good and not something I would've seen.