4 ms·

https://www.science.org/doi/10.1126/science.ade9097 https://www.science.org/doi/10.1126/science.ade9097 Abstract: Despite much progress in training AI systems

https://www.science.org/doi/10.1126/science.ade9097 https://www.science.org/doi/10.1126/science.ade9097 Abstract: Despite much progress in training AI systems to imitate human language, building agents that use language to communicate intentionally with humans in interactive environments remains a major challenge. We introduce CICERO, the first AI agent to achieve human-level performance in Diplomacy, a strategy game involving both cooperation and competition that emphasizes natural language negotiation and tactical coordination between seven players. CICERO integrates a language model with planning and reinforcement learn- ing algorithms by inferring players’ beliefs and intentions from its conversations and generating dialogue in pursuit of its plans. Across 40 games of an anonymous online Diplomacy league, CICERO achieved more than double the average score of the human players and ranked in the top 10% of participants who played more than one game.

digdugdirk 4y ago
"... by inferring players’ beliefs and intentions from its conversations and generating dialogue in pursuit of its plans." ... Isn't this the bad path of AI research? An unbeatable and utterly convincing conversational AI that knows exactly what you want to achieve, and then completely nullifies your attempts at reaching your goals while simultaneously achieving it's own?
- stavros 4y ago
  You've read a lot into that sentence. It literally says: > ... by inferring players’ beliefs and intentions from its conversations and generating dialogue in pursuit of its plans. As opposed to what? Not inferring intentions, or generating dialogue against its own plans, or at random? It's just doing what a person does. It understands the other person, and says things to further its goals.
  - mtlmtlmtlmtl 4y ago
    Sure, people do that. But people who only do that are sociopaths.
    - stavros 4y ago
      I'm sure the AI hangs out at the pub with its friends after hours though.
      - smaudet 4y ago
        If only, like a sociopath it hangs out with no one...
- EGreg 4y ago
  Just wait until we have AI swarms. So far all this has been with one player, amid others, no collusion. You’re going to be surrounded very shortly by sleeper bots, including on HN. Relying on dang and others to root out bots will be futile. A swarm could easily collude to downvote people or get them ostrasized by their own friend group, as we have already seen when it came to crypto, metoo, BLM, lockdowns, vaccines and now Ukraine. It’s really not hard for a bot swarm to completely exploit society in these and many more ways, and we are not ready for it. By the time the botswarms arrive online, it’ll be too late to do anything. Update: LinkedIn already has a huge problem of fake profiles applying for jobs and offering jobs. And Twitter is overrun with bots. But soon, bot-written articles will be out-shared by your own friends rather than that “hack liberal/establishment/hasbeen paper” NYTimes.com The animatrix had a good sort of storyline on this, but it involved a lot of unrealistic violence
  - worldsayshi 4y ago
    There are ways to ensure that people are people on the internet though. Not good ways but there are ways. You could for example limit users to those that log in with electronic id's issued by a government or other organisation that you trust to assert that the user is human and then force real names or a single user name for that e-id.
    - EGreg 4y ago
      But when bots do a lot of things better than humans, including comment, answer questions, make jokes, say romantic things, etc. many would start to prefer them to humans in their online communities, and give them even more social capital. This already happened in other areas of life. Both fathers and mothers now neglect their own children and elderly parents so they can work for corporations. They often prefer this and find meaning in climbing the corporate ladder. Eventually, their own labor will be rendered obsolete, but for now they're in a race to the bottom to work harder and neglect their family even more. They even stick them in nursing homes. Also, you no longer want to ask people for directions, you use Google Maps. You no longer ask your parents, teachers or libraries when you can just look it up online with no judgment. Finally, look at industries like Wall Street trading. It used to be a bunch of guys in a pit. You'd call up your broker or whatever. Now everything is automated with bots. Everyone prefers bots. They make up the bulk of trading with real capital. These bots are are working for corporations, which employ less and less humans. So the present is already a bunch of corporations owning bots and bots creating content for other bots. In the finance industry. Now how different is a bunch of text generation online? I think the human contributions will be vanishingly small in most communities. The question is ... what is this all for? Dropping demand for human services is a byproduct of making things more and more efficient...
bee_rider 4y ago
> Cicero participated anonymously in 40 games of Diplomacy in a “blitz” league on webDiplomacy.net > Cicero ranked in the top 10% of participants who played more than one game and 2 nd out of 19 participants in the league that played 5 or more games. > As part of the league, Cicero participated in an 8-game tournament involving 21 participants, 6 of whom played at least 5 games. Participants could play a maximum of 6 games with their rank determined by the average of their best 3 games. Cicero placed 1st in this tournament. This bit seems a little more impressive I think. Being in the top 10% of people who’ve played at least two games might leave a lot of bad players to beat up on. Winning a tournament might (?) mean you have to beat at least a couple players who understand the thing. It is sort of funny to think about — anyone who gets really legitimately good at anything competitive goes through multiple rounds of being the best in their social group, and then moving on from that group to a new one that is comprised of people who were the top-tier of that previous level. It isn’t obvious to me where on that informal ladder this tournament was. But anyway, maybe the AI will follow the trajectory of chess AIs and quickly race away from human competition.
- treis 4y ago
  I'm a little bit suspicious of this. They're not explicit about the scoring but taking the average of top 3 results is a huge advantage to those that played more games. Diplomacy is a bit of a choose your own adventure game too. Like there's an objective criteria (average SCs at the agreed end of game) but the human tendency is to try and win individual games. Humans will often choose to play sub-optimal strategies for better entertainment value. I think the real accomplishment here is the ability to fool humans into thinking their not playing a bot. That's an impressive thing to do even these days.
- gcanyon 4y ago
  There's an interesting question of "how much do the literal best humans suck at this?" For example, in chess Magnus Carlsen might be able to beat Stockfish given a handicap of just a pawn or two. An even better computer player than Stockfish might give up three or more pawns, but even a perfect player would likely lose to Carlsen if giving up a rook. -- I'm making this up, I don't think anyone knows the real values, but as far as I know no one is remotely projecting that perfect play could overcome e.g. a queen handicap. Similarly, in Go it seems unlikely that perfect play could overcome a nine-stone handicap (again, I could be wrong, I'm not remotely a dan-level player). All to say, it seems likely that Diplomacy is a game where the difference between "the best human play" and "the best possible play" is much larger than either Go or Chess.

digdugdirk 4y ago

"... by inferring players’ beliefs and intentions from its conversations and generating dialogue in pursuit of its plans." ... Isn't this the bad path of AI research? An unbeatable and utterly convincing conversational AI that knows exactly what you want to achieve, and then completely nullifies your attempts at reaching your goals while simultaneously achieving it's own?

stavros 4y ago
You've read a lot into that sentence. It literally says: > ... by inferring players’ beliefs and intentions from its conversations and generating dialogue in pursuit of its plans. As opposed to what? Not inferring intentions, or generating dialogue against its own plans, or at random? It's just doing what a person does. It understands the other person, and says things to further its goals.
- mtlmtlmtlmtl 4y ago
  Sure, people do that. But people who only do that are sociopaths.
  - stavros 4y ago
    I'm sure the AI hangs out at the pub with its friends after hours though.
    - smaudet 4y ago
      If only, like a sociopath it hangs out with no one...
EGreg 4y ago
Just wait until we have AI swarms. So far all this has been with one player, amid others, no collusion. You’re going to be surrounded very shortly by sleeper bots, including on HN. Relying on dang and others to root out bots will be futile. A swarm could easily collude to downvote people or get them ostrasized by their own friend group, as we have already seen when it came to crypto, metoo, BLM, lockdowns, vaccines and now Ukraine. It’s really not hard for a bot swarm to completely exploit society in these and many more ways, and we are not ready for it. By the time the botswarms arrive online, it’ll be too late to do anything. Update: LinkedIn already has a huge problem of fake profiles applying for jobs and offering jobs. And Twitter is overrun with bots. But soon, bot-written articles will be out-shared by your own friends rather than that “hack liberal/establishment/hasbeen paper” NYTimes.com The animatrix had a good sort of storyline on this, but it involved a lot of unrealistic violence
- worldsayshi 4y ago
  There are ways to ensure that people are people on the internet though. Not good ways but there are ways. You could for example limit users to those that log in with electronic id's issued by a government or other organisation that you trust to assert that the user is human and then force real names or a single user name for that e-id.
  - EGreg 4y ago
    But when bots do a lot of things better than humans, including comment, answer questions, make jokes, say romantic things, etc. many would start to prefer them to humans in their online communities, and give them even more social capital. This already happened in other areas of life. Both fathers and mothers now neglect their own children and elderly parents so they can work for corporations. They often prefer this and find meaning in climbing the corporate ladder. Eventually, their own labor will be rendered obsolete, but for now they're in a race to the bottom to work harder and neglect their family even more. They even stick them in nursing homes. Also, you no longer want to ask people for directions, you use Google Maps. You no longer ask your parents, teachers or libraries when you can just look it up online with no judgment. Finally, look at industries like Wall Street trading. It used to be a bunch of guys in a pit. You'd call up your broker or whatever. Now everything is automated with bots. Everyone prefers bots. They make up the bulk of trading with real capital. These bots are are working for corporations, which employ less and less humans. So the present is already a bunch of corporations owning bots and bots creating content for other bots. In the finance industry. Now how different is a bunch of text generation online? I think the human contributions will be vanishingly small in most communities. The question is ... what is this all for? Dropping demand for human services is a byproduct of making things more and more efficient...

bee_rider 4y ago

> Cicero participated anonymously in 40 games of Diplomacy in a “blitz” league on webDiplomacy.net > Cicero ranked in the top 10% of participants who played more than one game and 2 nd out of 19 participants in the league that played 5 or more games. > As part of the league, Cicero participated in an 8-game tournament involving 21 participants, 6 of whom played at least 5 games. Participants could play a maximum of 6 games with their rank determined by the average of their best 3 games. Cicero placed 1st in this tournament. This bit seems a little more impressive I think. Being in the top 10% of people who’ve played at least two games might leave a lot of bad players to beat up on. Winning a tournament might (?) mean you have to beat at least a couple players who understand the thing. It is sort of funny to think about — anyone who gets really legitimately good at anything competitive goes through multiple rounds of being the best in their social group, and then moving on from that group to a new one that is comprised of people who were the top-tier of that previous level. It isn’t obvious to me where on that informal ladder this tournament was. But anyway, maybe the AI will follow the trajectory of chess AIs and quickly race away from human competition.

treis 4y ago
I'm a little bit suspicious of this. They're not explicit about the scoring but taking the average of top 3 results is a huge advantage to those that played more games. Diplomacy is a bit of a choose your own adventure game too. Like there's an objective criteria (average SCs at the agreed end of game) but the human tendency is to try and win individual games. Humans will often choose to play sub-optimal strategies for better entertainment value. I think the real accomplishment here is the ability to fool humans into thinking their not playing a bot. That's an impressive thing to do even these days.
gcanyon 4y ago
There's an interesting question of "how much do the literal best humans suck at this?" For example, in chess Magnus Carlsen might be able to beat Stockfish given a handicap of just a pawn or two. An even better computer player than Stockfish might give up three or more pawns, but even a perfect player would likely lose to Carlsen if giving up a rook. -- I'm making this up, I don't think anyone knows the real values, but as far as I know no one is remotely projecting that perfect play could overcome e.g. a queen handicap. Similarly, in Go it seems unlikely that perfect play could overcome a nine-stone handicap (again, I could be wrong, I'm not remotely a dan-level player). All to say, it seems likely that Diplomacy is a game where the difference between "the best human play" and "the best possible play" is much larger than either Go or Chess.