Resolves positively if there is an AI can learn to play randomly selected computer games (shooters, strategy games, flight simulators, etc) at the level of an amateur but not completely incompetent human player, given only a small amount of time (days, not years) for its programmers to connect it properly, and the opportunity to practice for arbitrary (but achievable) amounts of time.
I will resolve this positively if the AI succeeds more than half the time. It's okay if it also has a few games it just can't learn.
This could potentially resolve YES in a few months:
@benshindel "A few months" is seems very unlikely to me. Note that this paper only has the AI doing simple tasks that take ~10 seconds to complete, and only when instructed. Having it able to do the kinds of long-term tasks most games require seems like a pretty big step - one that would take years, not months, of progress.
1) “Human level” in most games is not as high of a bar as you think. It doesn’t need to “beat all humans” at the game.
2) Once this tool is released to game devs to use, it will become rapidly more clear whether it has the potential to be at the versatility and the level required for this market.
That’s what I meant by “in a few months”: once we see how this is used by the public or by a wider range of developers
Hoping this can play Inflection Point, because there's currently no single player option, and it's hard to get people to play with me.
.
@MatthewLeong markets like this should resolve NO if there is no information to make it resolve YES. Scott Alexander is a trustworthy market creator, so there shouldn't be any shenanigans.
@FlorisvanDoorn - If that's the case, then I feel like this won't be a specific goal of an AI research team, and so won't happen by default, despite potentially being possible.
I think we're in the realm of having LLMs with plugins, or things of that nature, and no longer focussing on conquering games as we have in the past.
If the AI is accessing the state of the game in a way a human cannot, will that count? I would say that if it has direct read and write access to game state, rather than HDMI-in and HID-out access to the game, it’s not playing the game at all.
Regardless, I don’t think that a couple of days will be sufficient to learn how to play randomly selected game from scratch even with total read access to the game state and HID-out access to control it.
@Gigacasting deepmind was able to do this with atari games
so this resolving to yes is just a matter of incremental progress. We'd have to suffer another winter imo for this to resolve to no.
@MichaelDickens Alternatively, "off-the-shelf" just means it's a general-purpose game-playing AI, rather than custom-built for a specific game
@AaronKaufman I don't think so. The prompt is giving it a few days to connect and practice, and a random game on average has only a few inputs you can perform, as well as several limiting rules. If the random game is Minecraft the AI might be screwed, but if it's Celeste maybe it is getting close already?
@AaronKaufman It does involve re-training of weights. How much can the model learn with the given set of weights such that it is not only able to perform well at the tasks its optimized its weights for but also learn to do new tasks from very very few examples
@vmjusto Anything that can be played with an Xbox controller has the same number of inputs, even if it ignores some of them.