Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

CICERO: An AI agent that negotiates, persuades, and cooperates with people

ai.facebook.com

Post reply on HN
Print view View original

261–270 of 300 posts

Re: CICERO: An AI agent that negotiates, persuades, and cooperates with people

#261

Post by vintermann »

dmurray wrote: post #168
thevinter wrote: post #119

Earlier quoted context omitted.

It depends on the players, but for how the game is structured you still have a good chance of "winning" (finishing alive with in a draw) even if you're way underpowered, as long as you're friendly and provide some value

It's also bad form to resign even if you have no hope and will be finished off in the next few turns, because it will inevitably favour one of your opponents not to have to worry about you at all.

"Bad form"? In diplomacy? Isn't weak players acting as kingmakers the core of what you should worry about as a player with a shot at victory?

Re: CICERO: An AI agent that negotiates, persuades, and cooperates with people

#262

Post by sanxiyn »

eightysixfour wrote: post #121
ekdinan wrote: post #62

Earlier quoted context omitted.

Indeed, consistency (with the lengthy dialogue histories, but also the game state as well as game plans) was a huge challenge for us. We spent a lot of time working on techniques for detecting and filtering these kinds of low quality messages. You can see the Supplementary Materials in the paper for full details, but TL;DR: we built a suite of classifiers for detecting common mistakes (training classifiers to discrim…

Do you think it is an easier challenge than than just “converse with a human” since there is a purpose/game state which is driving the interactions?

It is closed domain, so yes, certainly easier than open domain chatbot.

Re: CICERO: An AI agent that negotiates, persuades, and cooperates with people

#263

Post by sanxiyn »

dqpb wrote: post #185
ekdinan wrote: post #62

Earlier quoted context omitted.

Indeed, consistency (with the lengthy dialogue histories, but also the game state as well as game plans) was a huge challenge for us. We spent a lot of time working on techniques for detecting and filtering these kinds of low quality messages. You can see the Supplementary Materials in the paper for full details, but TL;DR: we built a suite of classifiers for detecting common mistakes (training classifiers to discrim…

Did you capture conversational "transactions" as structured data in the game state, or was the chat history itself the only storage for that aspect of the game? I would think you could avoid much of this issue by creating a more sophisticated structured game model and use the language model only for converting between structured and unstructured data.

They do have structured game model although it doesn't capture everything in chat. Language model still had lots of problems with consistency even with structured game model input.

Re: CICERO: An AI agent that negotiates, persuades, and cooperates with people

#264

Post by sanxiyn »

LucasBrandt wrote: post #58
modeless wrote: post #52

I plan to read the paper later, but anyone have a TL;DR for how they connected a language model to the game state? That seems like the real advance here. Language models are so prone to making stuff up and spouting nonsense, and controlling them is really hard.

> At a high level, Cicero combines a dialogue module with a strategic reasoning module, along with a filtering process to reject low-quality messages. The strategic reasoning module generates "intents" (a set of planned actions for the agent and its speaking partner) which are sent to the dialogue module. They trained the model to be able to generate messages from those intents using data from messages in real games…

No, intent labeling wasn't manual, it was done automatically from raw data.

Re: CICERO: An AI agent that negotiates, persuades, and cooperates with people

#265

Post by mach1ne »

ImHereToVote wrote: post #247
mach1ne wrote: post #245

Earlier quoted context omitted.

At some point they probably will be, but it seems unlikely to happen within the next 7 years at least. Deep Learning still has the same deep-seated issues it has had since its inception.

https://intelligence.org/2017/10/13/fire-alarm/

I don't think it's a question of incremental progress, with AGI behind some ghost line. I think the current progress is, despite all the excitement, actually orthogonal to true AGI development. We're developing ever more sophisticated narrow AI, not even taking baby steps to proper true AI. I believe that the progress to the latter would be evidenced by the development of artificial intelligence first with animal capabilities, then with child-level capabilities and ultimately adult human tier capabilities. But I could be wrong - perhaps our current navigation of the problem space will reveal some strange pathway to true AI.

Re: CICERO: An AI agent that negotiates, persuades, and cooperates with people

#266

Post by camkego »

This seems like a good foundation for more testing of the AI-box experiment.

"It is a thought experiment where we create an AI which is isolated but convinces a human into allowing it a digital connection to the outside internet, whereupon it escapes into the wild." @curiosimo/Reddit

https://towardsdatascience.com/the-ai-box-experiment-18b1398...

Re: CICERO: An AI agent that negotiates, persuades, and cooperates with people

#267

Post by renewiltord »

sanxiyn wrote: post #252
nopinsight wrote: post #235

Earlier quoted context omitted.

"Having read the paper & supplementary materials, watched narrated game & spoken to one of the human players I'm pretty concerned. The @ScienceMagazine paper centres 'human-AI cooperation' & the bot is not supposed to lie. However, videos clearly show deception/manipulation" "Screenshots of the stab below. The human player said: "The bot is supposed to never lie [...] I doubt this was the case here" "I was definitely…

As far as I can tell, as described in the paper, the bot in fact never lies, in this sense: there is a model that generates messages from moves, where messages should correspond to moves, and when the bot says any messages, at the time, they are generated from moves the bot truthfully intends to play. On the other hand, the bot has no concept whatsoever of keeping its words. After saying words, it is free to change i…

Amusing, it's described like the "buggers" in Ender's Game!

Re: CICERO: An AI agent that negotiates, persuades, and cooperates with people

#268

Post by dmurray »

vintermann wrote: post #261
dmurray wrote: post #168

Earlier quoted context omitted.

It's also bad form to resign even if you have no hope and will be finished off in the next few turns, because it will inevitably favour one of your opponents not to have to worry about you at all.

"Bad form"? In diplomacy? Isn't weak players acting as kingmakers the core of what you should worry about as a player with a shot at victory?

Yes, and it's totally kosher to say "if you stab me on this turn I will order 'all units hold' until the end of the game" - equivalent to resigning.

But it's not OK to irrevocably commit to that decision, by, say, leaving the room and driving home.

I think this is true even in groups that take quite a liberal approach to gamesmanship and what might be cheating in other games: intentionally submitting illegal orders, peeking at other players' orders, etc. I don't know how to reconcile this logically other than by saying the game only works when all players are trying to win. Some would go further and say the game only works when most or all players are trying for a solo victory, since if you can be certain several players are happy with a 3 or 4-way draw that will always be the outcome.

Re: CICERO: An AI agent that negotiates, persuades, and cooperates with people

#269

Post by julianbuse »

ndonnellan wrote: post #14
kneebonian wrote: post #5

For those hearing of this game for the first time and think it sounds fun. Be warned, it is considered to be a game that will end friendships that have lasted for years, and monopoly is considered to be a game that causes less bad feelings. An AI may be able to speak like a person but will never be able to hang onto that long burning simmering hatred from when Brad didn't support my army and instead flipped on me by…

If you want all these things and exponential growth, check out Neptune's pride: https://np.ironhelmet.com/ for basically "diplomacy in space with a few extra dimensions". I haven't played in a few years (it was great at the beginning of the pandemic), but it was a lot of "fun" in the same way that diplomacy is "fun". Except with a space theme and a lot of anxiety waking up at 2am hoping your fleet arrives at an empty…

IMO a better game is subterfuge, because it allows you to schedule commands into the future, which means you don't have to wake up at ungodly hours of the night to execute a command.

Re: CICERO: An AI agent that negotiates, persuades, and cooperates with people

#270

Post by WastingMyTime89 »

mach1ne wrote: post #265
ImHereToVote wrote: post #247

Earlier quoted context omitted.

https://intelligence.org/2017/10/13/fire-alarm/

I don't think it's a question of incremental progress, with AGI behind some ghost line. I think the current progress is, despite all the excitement, actually orthogonal to true AGI development. We're developing ever more sophisticated narrow AI, not even taking baby steps to proper true AI. I believe that the progress to the latter would be evidenced by the development of artificial intelligence first with animal cap…

> But I could be wrong - perhaps our current navigation of the problem space will reveal some strange pathway to true AI.

Most likely wrong if you compare to how most other development in engineering happened.

From my point of view, the most likely path towards some form of weakly general intelligence at this point is emergence. We keep working on these narrow problems and from the broadening networks at some point we end up with something indistinguishable from a general ai inadvertently.

Post reply on HN