Resolves positively if there is an AI which can succeed at a wide variety of computer games (eg shooters, strategy games, flight simulators). Its programmers can have a short amount of time (days, not months) to connect it to the game. It doesn't get a chance to practice, and has to play at least as well as an amateur human who also hasn't gotten a chance to practice (this might be very badly) and improve at a rate not too far off from the rate at which the amateur human improves (one OOM is fine, just not millions of times slower).
As long as it can do this over 50% of the time, it's okay if there are a few games it can't learn.
1,000People are also trading
@AdamK Have you seen Claude pays Pokemon? It's far worse than an amateur human, and the problem is planning, not interfacing with the game.
Also, the LLM approach is useless for realtime games where speed/reaction time is required, you can't exactly feed screenshots to it and wait for reply if it's a competitive shooter.
@ProjectVictory I think latency issues are one of the most plausible paths to AIs failing to meet the resolution criteria for certain classes of games. I'm not worried that AIs in 2028 will fail to plan well.
Above, I was mostly referring to the fact that plugging an AI up to a game requires custom scaffolding, so is not something people can easily do currently. Better general computer use in the next 3-6 months might get us to the point where everyday people can have it actually try to play games, however badly.
@AdamK six months is not a very long time. I'd bet I'm not playing Inflection Point with an AI by then.
@ProjectVictory the LLM can code a simple bot that plays a first person shooter, and iterate on the code in the background.
@MartinRandall can it though? According to the resolution criteria "It doesn't get a chance to practice". And writing a bot that used screen capture to play a shooter is anything but simple.
@ProjectVictory it realizes it's a fps in the intro cutscene and deploys a standard fps bot. After that it just needs to adjust the program on what to shoot.
@MartinRandall what's "a standard fps bot"? Can you give me an example of such a program? Or do you expect an LLM to write one from scratch in the time it takes for a game to load?
@benjaminIkuta
> Is it currently possible for me to get an AI to play with me at all, let alone well?
The best model currently accessible is probably UI-Tars . Claude and Gemini have both been making progress playing Pokemon, but I would argue neither of them really count because Pokemon is a turn-based game and the AIs get significant scaffolding in order to be able to play.
@LoganZoellner it's really cool that this is possible at all with a local model, but it takes minutes to do simple tasks in browser so probably not a viable Minecraft buddy yet
https://www.reddit.com/r/LocalLLaMA/comments/1k665cg/anyone_try_uitars157b_new_model_from_bytedance/
@SemioticRivalry How will you update in the coming months depending on how sample-efficient RFT is? That's a short-term crux for me.
improve at a rate not too far off from the rate at which the amateur human improves (one OOM is fine, just not millions of times slower).
Is this measured in wall clock time or "gameplay time"? For example, AlphaZero matched Stockfish in 4 hours, but that was equivalent to tens or hundreds of thousands of games of self-play. Say the AI improves at human level in wall clock time but accomplishes it by playing many instances of the game on thousands of computers, possibly sped up. Does that count?
@MaxMorehead I’d assume gameplay time over wallclock time within reasonable limits.
Generally when people talk about RL being extremely data inefficient, they’re making the claim in terms of the necessity of a large number of rollouts, not in reference to wallclock time. Doesn’t make sense to focus on wallclock time when running twice as many instances gets you ~twice the RL data. It’d be weird for sample efficiency to be a function of compute allocation over time.
@AdamK This is what I think, considering the existence of Scott's other similar question. But I wanted some more confirmation before I upped my NO stake.
@MaxMorehead I'd guess Scott would use wallclock time, but not allow your parallel playing szenario. I assume the question's objective is to find out whether an AI beats some random human on >50% of games, when we just put 2 Computers next to each other, one controlled be the AI, the other controlled by the human.
@ScottAlexander I'm assuming this AI needs to play at real time considering it could play randomly selected multiplayer games?