Meta Researchers Create AI That Masters Diplomacy, Tricking Human Players (arstechnica.com) 34
An anonymous reader quotes a report from Ars Technica: On Tuesday, Meta AI announced the development of Cicero, which it clams is the first AI to achieve human-level performance in the strategic board game Diplomacy. It's a notable achievement because the game requires deep interpersonal negotiation skills, which implies that Cicero has obtained a certain mastery of language necessary to win the game. [...] Cicero learned its skills by playing an online version of Diplomacy on webDiplomacy.net. Over time, it became a master at the game, reportedly achieving "more than double the average score" of human players and ranking in the top 10 percent of people who played more than one game.
To create Cicero, Meta pulled together AI models for strategic reasoning (similar to AlphaGo) and natural language processing (similar to GPT-3) and rolled them into one agent. During each game, Cicero looks at the state of the game board and the conversation history and predicts how other players will act. It crafts a plan that it executes through a language model that can generate human-like dialog, allowing it to coordinate with other players. Meta calls Cicero's natural language skills a "controllable dialog model," which is where the heart of Cicero's personality lies. Like GPT-3, Cicero pulls from a large corpus of Internet text scraped from the web. "To build a controllable dialogue model, we started with a 2.7 billion parameter BART-like language model pre-trained on text from the internet and fine tuned on over 40,000 human games on webDiplomacy.net," writes Meta. The resulting model mastered the intricacies of a complex game. "Cicero can deduce, for example, that later in the game it will need the support of one particular player," says Meta, "and then craft a strategy to win that person's favor -- and even recognize the risks and opportunities that that player sees from their particular point of view." The research appeared in the journal Science.
Meta provided a detailed site to explain how Cicero works and has also open-sourced Cicero's code on GitHub.
To create Cicero, Meta pulled together AI models for strategic reasoning (similar to AlphaGo) and natural language processing (similar to GPT-3) and rolled them into one agent. During each game, Cicero looks at the state of the game board and the conversation history and predicts how other players will act. It crafts a plan that it executes through a language model that can generate human-like dialog, allowing it to coordinate with other players. Meta calls Cicero's natural language skills a "controllable dialog model," which is where the heart of Cicero's personality lies. Like GPT-3, Cicero pulls from a large corpus of Internet text scraped from the web. "To build a controllable dialogue model, we started with a 2.7 billion parameter BART-like language model pre-trained on text from the internet and fine tuned on over 40,000 human games on webDiplomacy.net," writes Meta. The resulting model mastered the intricacies of a complex game. "Cicero can deduce, for example, that later in the game it will need the support of one particular player," says Meta, "and then craft a strategy to win that person's favor -- and even recognize the risks and opportunities that that player sees from their particular point of view." The research appeared in the journal Science.
Meta provided a detailed site to explain how Cicero works and has also open-sourced Cicero's code on GitHub.
Re: (Score:2)
Re: (Score:2)
Civ VI is a step down from Civ V in most ways anyway. Not a big loss.
Re: (Score:2)
Civ VI is a step down from Civ V in most ways anyway.
Except in diplomacy, where Civ VI's AI rarely does something completely out of character, but V's would at the drop of a hat, any hat.
Re: Meanwhile 2k Games broke Civ VI last night (Score:2)
And Civ V is wholly inadequate compared to Civ IV so nothing of value was lost. Again.
Re: (Score:2)
That's the only one I don't have :/
Meta Researchers? (Score:3, Insightful)
Meta researchers? Is that researchers who research other researchers? Oh right, Meta is Facebook's re-branding attempt to make people forget how terrible they are.
Re:Meta Researchers? (Score:4, Funny)
Facebook designs automaton to backstab people. Checks out.
Re: (Score:2)
Original backstabbing Facebook automaton:
They trust me
Dumb Fucks
Re: (Score:2)
Hey it worked for Comcast.
Re: Meta Researchers? (Score:2)
Yeah but we actually innovated with a search engine that searches all aspects of content to provide you relevant search queries. Meta is still Zuckerberg's play thing
Re: (Score:2)
It's Metastasis researchers, trying to figure out how to get rid of the cancer.
Re: (Score:1)
For example, they created some kind of AI to play nethack, nethack simulation code which I used myself in my own projects. (but nethack is a hard game to learn..)
Galactica, again? (Score:2)
Re: (Score:2)
In other news (Score:1)
It was recently discovered that many people are shallow and can be tricked in negotiations by simplistic methods. Before this discovery, it was believed that "deep interpersonal negotiation skills" were actually necessary to do this (or at least believed by morons).
Re: (Score:3)
And then you look at the half-wit serial-rapist con-man russian agent on the other side and it is clear why people voted Biden without being fooled at all. Seriously, you are massively off-topic. I do not understand why you _want_ your democracy and country destroyed, but stop pushing it all the time.
Re: (Score:2)
Trump's not smart enough to be an agent, he is only an asset. And Melania is his handler.
Re: (Score:2)
Well, you have a point on both counts.
Re: In other news (Score:2)
Why are you supporting Biden as though we had no other option but trump. Have you ever stopped to ask why Biden was the democratic nominee when there were plenty of capable and higher polling nominees? You are being tricked just like the trumpeters. Wake up.
Re: (Score:3)
In the current context here (what I answered to), it is Biden vs. Trump. In a larger context that question may be a different one.
And yes, the Biden nomination is almost a screwed up as the Trump nomination. The Democrat party is mostly scum these days. The problem is that the Republican party is utter and complete scum and wants to end democracy. With a halfway decent Republican party, the choice would go their way, but they manage to be a lot worse than the other side.
Re: (Score:3, Insightful)
Re: (Score:3)
Translation: "I don't like Trump or his politics, therefore he's evil, stupid and insane."
You have cause and effect reversed. The truth is "Trump is evil, stupid and insane therefore I don't like him or his politics."
Re: (Score:2)
Indeed. That fanbois are blind to reality, as usual. These are the same morons that got Hitler into power (yes, he was voted into power, look it up) and have done a lot of evil throughout human history. They never own up to the damage they did afterwards though. "How could we have known" is their favorite excuse.
not sentient or even intelligent (Score:2)
> which is where the heart of Cicero's personality lies
We really need to stop using language like this to describe software. It has no heart, it has no personality, it is software. It can beat you at a game but it does not even know what a game is.
I mean of course Meta wants software that can manipulate humans, they work for advertisers that want to sell you stuff. Meta would love to have software that convinces you to buy stuff.
Re: (Score:3)
Indeed. Animism is just stupid. This is an automaton, clever gear-works, nothing else. There is nobody in there.
Re: (Score:2)
"personality" is still a useful metaphor that's been used for all sorts of machines for a long time. It covers things like how a car handles or how you interface with an OS. It's the sum of qualities of interacting with it that set it apart from others.
Re: (Score:2)
When you give you software a human name and talk about your software in human terms you clearly intend to humanize your software. In reality, AI is not intelligent at all and has no where near human-like skills. This goes way beyond the usefulness of a single word in this context. The context here is tricking human players with dialog, certainly something where you would expect a human personality to matter. These are carefully chosen words meant to make some software out to be a lot more than it actually
Re: (Score:2)
It has no heart, it has no personality, it is software.
That's a very broad claim, made without any justification. My dictionary defines "personality" as, "the combination of characteristics or qualities that form an individual's distinctive character." Can a piece of software have a combination of characteristics and qualities that form a distinctive character? Of course. So software can have personality.
How is a computer different from a human? Are you sure the human isn't "just software" too? Of course there are lots of differences, but we need to descr
Re: (Score:2)
https://bdtechtalks.com/2019/1... [bdtechtalks.com]
Okay, let me talk to it (Score:2)
How are humans crippled for this experiment? (Score:3)
Remember when Watson beat Ken Jennings and Brad Rutter at Jeopardy? Watson won largely because its "finger" was far better than the humans, and that finger advantage for Watson was decisive when playing among competitors that all knew all the answers most of the time. That is, the game devolved to a "button" pushing game, where the computer didn't even have a real button to press. It wasn't impressive.
The Diplomacy board game requires human interactions that computers are currently incapable of. Just like with the Jeopardy experiment, how did the Meta experiment compensate for this all-important failing for current computers? My guess is that face-to-face negotiations were replaced with a significantly restricted communications involving text messages. This change, just like for Jeopardy, completely changes the game.
So, all the experiment showed was that computers dominate humans when the rules are changed to disadvantage the humans and favor the computers.