Facebook says it’s made a ‘human-level’ AI board game
Facebook, or as we should now call them Meta, announced earlier today that its CICERO artificial intelligence has achieved “human-level performance” in the board game. diplomaticIt is worth noting that this is a game built on human interaction, not moves and manipulation (such as chess).
Here’s a rather harrowing intro:
If you have never played diplomatic, and so might be wondering what the big deal is, it’s a board game first released in the 1950s, mostly played by people just sitting around tables (or breaking into rooms). ) and negotiate stuff. No dice or cards affect play; everything is defined by human communication with other people.
So, for the creators of the AI to say that it is playing at “human level” in a game like this is quite a bold statement! One thing that Meta supports by saying that CICERO is actually operating on two different levels, one that handles the progress and state of the game, the other that tries to communicate with the human level in a way we can understand and interact with.
Meta enlisted “World Champion of Diplomacy” Andrew Goff to support their claim, who said that “A lot of human players will soften their approach or they will start motivated by revenge and CICERO never did it. It only plays the situation when it sees it. So it’s ruthless when it comes to executing its strategy, but it’s not cruel in a way that annoys or frustrates other players.”
That sounds optimal, but as Goff said, maybe also Optimal. That reflects that while CICERO is playing well enough to keep up with humans, it’s still far from perfect. As Meta said herself in a blog post, CICERO “sometimes creates inconsistent dialogue that can undermine its goals”, and my own criticism is every example they provide of how they communicate (like the example inside bottom) makes it look like a scared psychotic office worker that if they don’t end each sentence with!!! you will think they are a bad person.
Of course, the ultimate goal of this program is not to win board games. It simply uses diplomatic as a “sandbox” to “enhance human-AI interaction”:
While CICERO is only capable of playing Diplomacy, the technology behind the achievement is relevant to many real-world applications. For instance, controlling natural language generation through planning and RL can reduce communication barriers between humans and AI-powered agents. For instance, today’s AI assistants excel at simple question-answering tasks, such as telling you the weather, but what if they could sustain a long conversation with one another? goal to teach you a new skill? Alternatively, imagine a video game in which non-player characters (NPCs) can plan and chat just like people — understand your motives and tailor the conversation accordingly — to help you on your quest to storm the castle.
I may not be a billionaire running Facebook, but instead of spending all my time and money making AI assistants better, which no one outside of AI research and public spending is doing. You seem interested, could we… hire people I could talk to instead?