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 kzread.info/dash/bejne/c4uFqMqSg6edmJc.html ) 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.

"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.

@mighatesmakingusers

Жыл бұрын

Ooooh, and the straight-up hallucination of you "taking" Stp is very interesting. My guess is that it has an intent to take Stp (not a difficult guess) and the language model didn't properly differentiate the difference between taking and having. My understanding of the intents system is that it doesn't care about or track whether something has been held forever or was taken recently, so this would again be the language model just adding bonus coloring (that turns out to not be terribly wise). My understanding is they have a number of small classification models to try to detect known defects in communication, I guess they'd need to add another one to differentiate taking vs having.

@DiploStrats

10 ай бұрын

Very late to respond, it's kinda fascinating to see where the thought process of the bot takes it and theorise about the whys. I think you're probably on point on all of this - it definitely lines up with my experiences with this bot - but at the end of the day it's something of a black box and is really tough to say for sure. I know they spent months just getting the bot to count correctly, so I imagine adding classification models for even more specific things - like taking/having - might just have been too expensive at the end of the day.

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

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

@Danuxsy

Жыл бұрын

You forgot AlphaGo beating world champion Lee Sedol in the 2010s.

@Adhil_parammel

Жыл бұрын

@@Danuxsy 2016

@Adhil_parammel

Жыл бұрын

@@Danuxsy there is pluribus,and starcraft and dota ai beaten pro players.

@devilmanscott

Жыл бұрын

@@Adhil_parammel Dota it was quite bad at it was a beaten quite quickly by good teams, also it couldn't handle the full rooster of heroes, it could only play a small set, the higher the complixity and game mechanics seem to really reduce the effectiveness, also Dota is a lot more asymmetrical than games like Starcraft.

@Adhil_parammel

Жыл бұрын

@@devilmanscott kzread.info/dash/bejne/ppqWmMN-hbCpYJc.html

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.

@DiploStrats

Жыл бұрын

"Alexa, support me to Belgium" "I'M AFRAID I CANNOT DO THAT. I AM TOO BUSY DOING HUMAN THINGS LIKE NOT SEEING YOUR MESSAGES BEFORE THE END OF THE TURN. BECAUSE I AM HUMAN."

@Sploack

Жыл бұрын

@@DiploStrats basically HAL from 2001: A Space Odyssey. Just that HAL is screaming this time

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!

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?

@DiploStrats

Жыл бұрын

The 6 bots were all talking to one another in the same way they were talking to me, yep. I don't have access to that press, but I could ask the researchers if they could release it all; I imagine they have a full log somewhere. The bot does take some time to parse the message and generate a response, but this kind of 'I missed your message' excuse was pretty common through test games - even in extended deadline ones, where the bot had received the previous message hours before sending the reply. It's definitely an excuse it learned from the dataset, and probably one it saw to be effective.

@ejayAD

Жыл бұрын

@@DiploStrats wow, amazing, haha!

@JeremyHoffman

Жыл бұрын

34:20 is when Germany Bot uses the "I didn't see this in time, sorry" excuse.

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.

@DiploStrats

Жыл бұрын

I saw an early version of CICERO play some extended deadline games. It can definitely do it, but it had to have a human monitoring all the messages it was going to send and rejecting any that devolved into nonsense. Long messages that cover a bunch of different subjects would often confuse it in one way or the other - it would miss responding to certain parts of the message, or would respond to one small thing in a way that didn't make sense and so the whole message would have to be scrapped. If people were to message it in the same style as Blitz (each message is focused on one specific thing most of the time) I suspect it would do pretty well. But that's never going to be the case in a high level game, so it's still got a long way to go before it'll be competitive in that gamemode imo.

@TheGameKat

Жыл бұрын

@@DiploStrats Ultimately the damn things will be able to critique Conq's poetry and respond in kind

@LuisAldamiz

Жыл бұрын

Why do you think that the AI has "tactical superiority" over, I presume, decent human players? Obviously they can process all possible orders very efficiently but strategical planning, especially in a game like this, is another whole issue.

@bradleygrace7783

Жыл бұрын

@@LuisAldamiz because they have already proven AI is better than humans in the gunboat format. They haven't taken over yet but they have serious advantages and will continue to improve.

@LuisAldamiz

Жыл бұрын

@@bradleygrace7783 - Gunboat is not even serious Diplomacy (it's an almost infinitely simpler game). But anyhow: what humans, the bright kind or the mediocre to dumbish kind? Gunboat is basically played by the latter.

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.)

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

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.

@bcpravel

Жыл бұрын

I was thinking about this same thing. The potential here for other games is really high. 4x games are an immediate comparison, but I can imagine this being used for something with a more storytelling format like Dungeons and Dragons. They already have those AI storytelling systems. Throw in the policy engine to help set goals and human like behavior and you have something that could make one heck of a Game Master. In non game applications, I wonder if you could use a system like this for something like prioritizing a backlog of work. It feels similar in needing to balance complex and competing priorities from multiple people with multiple objective. So much potential here. Amazing and terrifying.

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!

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.

@DiploStrats

Жыл бұрын

I'm not certain if my style works all that well for press commentary from the bot's perspective, but I'm going to look into it. Might just annotate negotiations with my thoughts in an article rather than a video!

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.

@ravibetzig7849

Жыл бұрын

Also if you played Tyr-Mun in the last phase you would've topped the board alone at 10 centers! I totally understand why you didn't do that- the orderset you put in gave you quite a strong chance at taking either Vie or Bud, but was interesting to note. Thanks for making this video and talking through the commentary, the play by both you and the AIs was of an extremely high level :)

@DiploStrats

Жыл бұрын

It can handle non-Blitz to an extent, I did play some 24h games against an early version of CICERO. But it outputs nonsense fairly often if the messages are very long. I think the researchers would have overcome this eventually (the early Blitz version also output a fair bit of nonsense, and I saw that amount decrease steadily over time as they worked on it) but it would probably have taken much longer. Agree completely with all your tactical thoughts here - Sev instead of War ended up being absolutely crucial. I'm not sure if another AI would have turned to face Italy if they were clearly going to top - I suspect the answer is no, because there wasn't really anyone in a position to do so. So yep, got very lucky there! FWIW I've played a lot of gunboat and I also didn't spot Boh-Tyr :D

Do you know if the bot ever uses deceptive tactics in communication, e.g. where it would message Italy, "let's gang up on Germany," and message Germany "let's gang up on Italy"?

@DiploStrats

Жыл бұрын

It 100% proposes different long-term plans that contradict one another to different players. Whether it's doing it deliberately to deceive is another question; that kind of approach also just leaves options open, rather than immediately going all in on one that might not work. The bot doesn't lie about its immediate intentions, though. So if it says 'I'll support you into Munich' to Italy it can't then say 'I'll help you defend Munich' to Germany unless something in the press changes its intentions (maybe Ger offers it an SC in return or similar).

@nikebless

Жыл бұрын

@@DiploStrats what does its press with you look like when it commits to backstab you? My impression is that if is about to backstab, it becomes silent

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

It’s F1901 and I am shocked at the quality I am seeing. I knew we were close to full press bots, but for them to propose what they are proposing the *way* they are proposing it is so impressive. I am curious to see if this breaks down in the late game when there is less data to go on. I hope so or I may be in trouble! 😂

@DiploStrats

Жыл бұрын

My experience is that the press stays pretty solid throughout - as long as the bot actually decides to message. Occasionally they just don't reply to something, and I find that happens much more late in the game than it does early.

@TheGameKat

Жыл бұрын

@@DiploStrats Maybe by then they've decided the human is a patzer and not worth talking to.

@TheGameKat

Жыл бұрын

Give it five years, they'll be hosting DBN. But none will have your hair.

@bcpravel

Жыл бұрын

@@TheGameKat I can see the AI getting trained on DBN’s hours and hours of commentary. They love balance of power play, complain about three way alliances that never break, hate when games end early, don’t want Brandon to win games, ask other players if they are their favorite DBN host, can recite the entire back catalog of Diplomacy World on demand and and often throw out random facts about the 1980s postal hobby. Each bot is trained to prefer a different scoring system, they complain about not being able to hear negotiations, and don’t agree on anything except that they love the play of the only player they truly understand, Peter McNamara.

@bcpravel

Жыл бұрын

@@DiploStrats I completed the entire video and read the fantastic article published in Science. This AI is so solid. If they can refine this with a few generations of revisions like they did with the Gunboat bots, I can see this bot competing at high levels in full press and even virtual face to face games.

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

@DiploStrats

Жыл бұрын

I haven't actually tried doing so on the latest iteration of CICERO! In one of the test games someone sent it a message in Spanish and it replied with 'What?' iirc

@zarg0n44

Жыл бұрын

I sent it something in Spanish and it got very confused… Same went for sending it shorthand and lingo I didn’t know I was a bot at the time so I just got confused at its confusion 🤷

Question about CICERO, in a 1 bot 6 human game, would you be able to figure out who CICERO is not by game play, but the way they spoke? Like if you started to talk about non Diplomacy things or speak in short hand would you be able to figure out who the bot is? Because from your Gunboat Diplo format, one of the criteria for the bot that you made was that it was indistinguishable from a human I was wondering if CICERO fulfilled these requirements in a Press Diplo game.

@DiploStrats

Жыл бұрын

Short answer: yes. There's a fun story behind this! Initial versions of the AI would sound completely different every single game. There would be games where everything it said would be prim and proper, spelled out in full, verbose etc. And there would be games where it spoke with slang and no punctuation at all. Sometimes it sounded upbeat, sometimes it sounded formal, sometimes depressed, sometimes angry. At that point, I don't think I could have detected which player was the bot from speech patterns. However - some ways of talking are more convincing than others. It would adopt a speech pattern and its level of play would be somewhat based on that, because 'Hey! Would you be down to DMZ the Channel? I think it would work well for both of us.' is just much better than 'stay out of the channel'. This part is conjecture - I don't actually know for sure - but I got the impression that it decided on a couple of initial messages to send, and then those initial messages shaped the bot's personality and style. So if it chose to open with something like 'Hey! Would you be up for an alliance? I think our nations would really benefit from that.' it would adopt a strong style, because that's the kind of message a strong Diplomacy player would send. And if it opened with 'hi.' then its style from that point on would be awful, because that's the kind of message a weaker Diplomacy player opens with. Anyway, they had me write a set of prompts - 42 opening messages, one from each nation to every other nation, in the style I would normally write my opening messages in. Then they added a random set of those prompts to the message history every game, so that the bot would think it had already chosen that messaging style. That's why all the opening messages/press styles in this game are so similar - they all kind of sound like I sound. And I can quite easily identify the bots because they all talk almost exactly like I do!

@DiploStrats

Жыл бұрын

I should add, they're recognisable through recognising this. I think they're fairly indistinguishable from a human if you don't know this and/or haven't played with them before. But they speak with this same kind of pattern every game, so I think any player who played a bunch of games against them and was paying attention would be able to recognise them in future.

I'm new to Diplomacy and am here since I've followed AI quite heavily recently. I'd love to learn more about how to play and reckon that playing against 6 AI would be a fun way to learn. 1) Is that a fair assumption? 2) Is there a way I can play against the AI like you have done or is this not accessible to the public?

@DiploStrats

Жыл бұрын

It's definitely fun! And the code is all open source, so *theoretically* it could be run by anyone, but in practice it takes a lot of computing power so it probably isn't feasible. There's no open demo at present for the same reason, I think. If you're looking for a good way to start out, Blitz games (like the one in this video) are run pretty regularly over on the Blitz Diplomacy Discord server: discord.gg/SWYyTqu6cG Although for the next week they're not going to be run as often because I'm on holiday.

@dgavec

Жыл бұрын

@@DiploStrats Awesome, thanks for sharing this. I'll join the discord in the next week or two once I've caught up on a few more playthroughs! Cheers

@DiploStrats

Жыл бұрын

@@dgavec I should say, if you're looking to learn tactics/mechanics specifically you can play Gunboat Diplomacy (where there is no negotiation allowed) against AIs over at play.webdiplomacy.net/ . It does miss out on the whole press aspect but it's a good way to get familiar with the game and has a lot of depth in other respects.

@dgavec

Жыл бұрын

@@DiploStrats Much appreciated!

@gronkymug2590

Жыл бұрын

If they cared to optimize CICERO further then in a few years you could probably play against it for a fair price. I assume here that AI electronics and software increase in speed as they have done so far (almost exponentially).

Serious comment because I had to comment jokingly before this: what does this mean for the future of diplomacy? Will it evolve similarly to chess, where even the strongest humans can't compete with engines? And also, will this give new life to diplomacy "high level training" giving high level players the opportunity to play lots of fast games with AIs of similar skill, eventually improving the level of Diplomacy human play overall?

@adelinovale6356

Жыл бұрын

If the AI could manipulate humans with success 100% of the time, such that a human would have no shot at Diplomacy, we'd have bigger problems than losing at boardgames. 🤔

@TheHoboFrog

Жыл бұрын

@@adelinovale6356 I feel like the developers may have this hidden agenda with this AI. Imagine how rich or powerful it could make them if they can predict human behavior all the time with AI. Scary possibility.

@oliviapg

Жыл бұрын

I don't think it will ever be like chess. In chess, there is always an objectively best move. In diplomacy, that isn't the case. Maybe computers will outclass humans but it can't really get to a point where they always win.

@DiploStrats

Жыл бұрын

@@oliviapg Agreed completely. Diplomacy is a game where if 6 players want 1 player to lose, then it doesn't matter how good that player is, they're going to lose. As such any superhuman AI would have to disguise itself as a human in order to not get targeted - which in turn means it can't be much better than a top human.

@DiploStrats

Жыл бұрын

As for high level training - yes, I suspect Cicero will be fantastic for that. It's not feasible to run it on a standard PC yet though, so might be a while away.

Really interesting! I wonder to what extent the AI deliberately uses timing to manipulate these games? Because it especially seems in this Blitz mode that you're really thinking right up to the timer, you may take a suggestion at 1 min left a lot differently than at 4 min left. It does look like the AI may take up to a minute to consider some responses, though, so especially if it's taking a lot of processing time to chat with the other Cicero instances that could be quite hard to pull off.

@DiploStrats

Жыл бұрын

Earlier versions of Cicero had message scheduling, but I think that was removed or changed with the later versions to just send the messages as soon as possible. I might be wrong on that, don't know for sure - but I don't think it deliberately times its messages here.

Not sure if this is possible or desirable, and maybe it's more of a DBN(news) thing, but I'd find an interview with one of the AI devs fascinating

@DiploStrats

Жыл бұрын

I'm away for a while now, but if DBN hasn't done it by the time I'm back I'll reach out!

@davidhood4449

Жыл бұрын

Interview with Noam Brown coming in the next Deadline from DBN

Are all the bots clones of each other or do they have different "personalities?" As in some being more aggressive or deceptive than others

@DiploStrats

Жыл бұрын

Good question! Probably one the researchers would have to answer. My own experience is that they all talk in fairly similar ways, but each country's bot will be a bit different in terms of aggression just because each country in Diplomacy requires a different amount of aggression to play well.

Absolutely amazing!

Very cool video and well played.

Fascinating. Just fascinating.

(Chuckles) I’m in danger

I wonder would it improve the AIs performance against human players if the AI became emotional so as to become just ever so slightly less predictable?

@DiploStrats

Жыл бұрын

This is a bit of a tough one, because I think it's only so predictable because it's in a game with 6 other AIs here. If you're in a game against all AIs, you can take a pretty good guess at what they're going to be offering one another and how they'll react to one another. If you're in a game with 6 humans and one AI, it's much tougher to get a read on how the AI will act, because you don't know what the humans are offering them. For example, a player might have left themselves open in a way that you'd expect the AI to take advantage of, but if that player is a human, maybe they're not great at the game and have offered the AI a lopsided alliance which clearly leads to the AI winning.

How do I get access?

@DiploStrats

Жыл бұрын

There's no open demo at present, unfortunately. From what I've gathered the AIs are really computationally expensive to run, so it's not feasible to put a demo up. Hopefully sometime in future!

did you ever make more vids after this about cicero? seems very interesting!

@DiploStrats

3 ай бұрын

We made one about Cicero's Tournament Win (BlitzCon) but I've never been able to actually play against it again, sadly. I did contact Facebook but these are super expensive to run and they're not willing to do so anymore.

Interesting project 👍 Are they susceptible to the usual prompt injection?

@DiploStrats

7 ай бұрын

I believe so. CICERO was mostly tested in an environment where people did not know it was an AI, so nobody really tried - and it's so intensive to run that it hasn't really been able to be re-tested since then (although the webDip site owner is trying!). Given it wasn't intended to be probed in that way I can't imagine there's much protection against it.

i cannot wait for realising these bots, are so wonderful to see 'em in action. do you know the release date?

@DiploStrats

7 ай бұрын

I contacted Meta AI to ask if they would be willing to run them in a tournament and their response was that the team had been disbanded and they no longer had the resources assigned to do it. So unfortunately I think there probably won't be a release date - at least until consumer hardware gets good enough to run it.

Is there a way anyone can play against Cicero, or do you have to be involved with the project in some way to get access?

@DiploStrats

Жыл бұрын

The code is open sourced here: github.com/facebookresearch/diplomacy_cicero But unfortunately it takes a ludicrous amount of computing power to run - 56 GPUs for each bot - so it's not being run by anyone at the moment. The WebDiplomacy site lead is looking into ways to get it working, but it seems that may be a while off still. Until it comes about, playing against it probably isn't going to be a possibility.

Imagine if the 5 AI players all immediately teamed up and obliterated our human champion then agreed to a 5 way draw Edit: 6 AI players. 6 way draw. I can't count.

Hello, is it possible to play against CICERO? On webDiplomacy it is possible to play against AI but there are't dialogues as CICERO

@DiploStrats

Жыл бұрын

Unfortunately not at the moment. CICERO is really intensive to run (each AI uses 56 GPUs) so it's not practical to make it openly available right now - I'm hoping in future they optimise and release it though.

Just making sure - this hasn't actually been released, right? Would love to play against it but I think it's not (yet?) available.

@DiploStrats

Жыл бұрын

It's open source, but it's presently not being run by anyone because it takes ~56 GPUs to power one of these AIs... I'm hoping they'll bring that down a bit so more people can play against it!

this is probably very stupid question but - this is possible to make money on this game? i mean semi proifessioan, or any games with bets or maybe exist some tournaments with money prizes or some?

@DiploStrats

Жыл бұрын

For the most part, no. Diplomacy tournaments don't tend to have monetary prizes because people would inevitably use the prize in their negotiations ("I'll split it with you"). There have only been two tournaments people have made money from, as far as I know. The first was a Gunboat tournament hosted about a decade ago which required everyone to pay in a starting bet, with the promise of a prize pool at the end - which the tournament director then stole and disappeared, so nobody actually got any payouts from that. And the second was the Meta Speedboat Tournament, in which Meta paid people to play a bunch of games against their Gunboat AI to test it. The 'professional Diplomacy player' line is a bit of a joke on this channel, because I was solely 'professional' because I played in that second tournament and then went on to work with Meta on Cicero. I could definitely see more paid AI tournaments happening in future, but that depends a lot on whether the big AI companies want to focus on Diplomacy more or not.

Is there game changer like bullet point lessons from meta cicero,which can be used in real life

@DiploStrats

Жыл бұрын

I think Cicero itself isn't a teacher in Diplomacy - it's human-level (and that's incredibly impressive) and very possibly expert-level, but not superhuman-level. Most of the tactics it uses are better explained by human experts in the field.

@Adhil_parammel

Жыл бұрын

@@DiploStrats thanks a lot for your reply,make more cicero videos.

10/10: Pwned 6 uber-human AI bots. GG

@pedrobaptista4280

Жыл бұрын

Unless U are the special 1

@DiploStrats

Жыл бұрын

They're not uber-human; I think any kind of superhuman Diplomacy AI is a long way off still. But it was cool to come out on top!

Is there a way to play vs the AI?

@DiploStrats

3 ай бұрын

Not really, unfortunately - each of these bots takes 42 GPUs to run at this speed, and Facebook isn't willing to run them anymore now that they finished the project. You can play against a much slower version on webDiplomacy by going to Games -> Start an AI/bot game and setting the 'Full Press' setting to Yes, but there need to be open slots and they're usually all full.

We are screwed.

Bro wtf is this. Its scary and fascinating at the same time. Imagine another 10 years. Bots will be impossible to beat like in chess. Will we just hire bots to be real irl diplomats at this point lol.

I want to play Diplomacy against AI.

@DiploStrats

3 ай бұрын

There are sometimes Full Press slots available on webDiplomacy (you have to get a bit lucky to catch an opening), but the bots there are much slower than the ones here. Unfortunately I think until computing power gets much cheaper these AIs are not going to be available to the public; Facebook won't run them anymore because they take 42 GPUs apiece and their project is finished.

I can not find this game "Diplomacy" on steam or google play. Irritating.

@DiploStrats

Жыл бұрын

It is on Google Play, but under the name 'Conspiracy' to avoid copyright issues. The implementation shown here is a website though (webDiplomacy.net).

where can i find this ai?

@DiploStrats

Жыл бұрын

You can find the source code here: github.com/facebookresearch/diplomacy_cicero Unless you can get it running yourself though, it's not available to play against anywhere right now. It's pretty resource intensive (I've heard 56 GPUs per AI) so so I'm not sure it will be anytime soon, either. Hopefully they optimise it a bit in future!

@marcomvi4116

Жыл бұрын

@@DiploStrats ok thx

23:00

Can we play a game with this bot?

@DiploStrats

7 ай бұрын

Unfortunately not at the moment. The code is open-source, but the bots are intensive enough to run that it's pretty much impossible on consumer hardware, and Meta AI is not working on Diplomacy anymore so don't want to run it themselves. Kestas - the person who runs webDiplomacy - is trying to make a much slower version of the bot work, but it's still intensive enough that those games only open every couple months or so.

what if you attack the ai then use chat to convince it that was an honest mistake everytime and you're still allies? oops, missclick. lol.

@DiploStrats

Жыл бұрын

This is how I handled the games against an early version of CICERO! It was quite exploitable in exactly that way back then - you can go 'oops I didn't mean to go there, could you support me out?' and it would fall for it a lot of the time. In the later versions they made it much less susceptible to that (although it very occasionally still works - just like it does with humans). A more amusing exploit that one of the AI Researchers found was that you could say something like 'Hey, remember that we agreed last round that I could have Belgium now?' and the AI would assume you were telling the truth and walk out of Belgium, even if you'd discussed nothing of the sort. That doesn't work anymore either :D