I didn’t really expect it to be a lecture with dice. I didn’t really consider the randomness involved in the setup.
The game has a lot of fantasy fun, in which the players get to temporarily roleplay as some of the most influential people in the world: AI CEOs. There is also the challenge of trying to get the highest share of the world you can, while other people are trying to do the same thing. So it’s about negotiation as well as strategy on how you hedge between safety research and capabilities (dynamics that arose from the rules of a 10 coin budget), with some competitive fellowship, and discovering the incentives behind why AI companies rush to build capabilities (produces the dynamics of players spending lots of coins on capability research to stay ahead of others).
It was quite interesting to see the difference between what people say and what they actually do. For example, S. said they were going to put 7 into capabilities but then put in 0. It illustrates that it is very important to make sure that rules are enforceable. In this game, promises didn’t have much of a reputational cost; it’s kind of expected that you deceive and try to get the most in this specific game. I feel like it’s somewhat different for politicians, or maybe not, because I’ve heard that US politicians don’t really think about it that way: there’s only a vague chance they get reelected, and it’s not really in their best interest to maintain the country’s reputation when someone else is going to take control of it later on.
Playtesters called it “a gambling game”, and that’s not what I intended. I was referring more to risk tolerance and uncertainty: that you should hedge, rather than just gamble, like, “Oh, there’s a very small chance, so I should do it.” If it is considered a gambling game, that means it’s like slot machines, where you can’t control the outcome. So what’s the point of doing anything about it? You can reduce the chances, though, so you do have autonomy over it.
Playtest 1 showed that it was pretty hard to stay quiet while I was playtesting. Now I realize that it’s best that I don’t explain, because in actual games I won’t be there every time someone plays, and the game has to be self-explanatory in and of itself. If it’s confusing and I explain too much, I might not get the feedback that it’s confusing.
A playtester said that anyone this game convinces is “already convinced.” Perhaps I agree, because the game assumes capability spending is dangerous (someone may think models just do what you tell them to do). The reason I’m going to the specification-gaming game is that it convinces people that it is very difficult to specify to large language models like ChatGPT to be good. What do you mean by good? It is sometimes best to go the unintended way to maximize a metric. For example, there is a famous story about British colonial officials in India who wanted to lower the number of cobras, so they put a bounty on their heads. But people started breeding cobras to claim the bounty, so once the bounty stopped, everyone released the cobras, and there ended up being more cobras at the end.
For P2, I will try to do more playtests where I am quiet. It feels natural to just go straight to the game: “Oh, I had this idea of a game.” I had it vaguely before the class: what if we have fictional AI companies in the AI race? It teaches people that currently the incentives are that AI companies will spend a lot on capabilities, even though they might want to spend on safety. I will try to start P2 from first principles: what is the purpose I want from creating the game? Then I’ll go from there into what elements I should have to convince people to go into AI safety.
