The setup will be pretty simple:
Once the game starts, neither me nor the AI can pause the game.
We spawn at least 1000 squares away from each other.
We both spawn with a turret and 100 rounds of ammo in addition to the default spawn items (this is to prevent early-game handgun scrimmages from deciding the contest).
Whoever kills the other player first wins.
This match will happen sometime in December 2025, whenever it's most convenient for me. There will only be one match.
A stalemate isn't possible, because I will have to resign eventually (unlike the AI I'll need to sleep at some point if the game drags on indefinitely). Me resigning will count as me getting killed.
Either player dying from environmental factors in the game (e.g. biters, getting hit by a train) counts as a kill for the other player.
My skill level in Factorio is mediocre, but I have managed to complete the "vanilla" campaign (before the Space Age expansion came out).
The requirement for the AI is that it uses only screen pixel data as input and uses only mouse/keyboard for control. I will play against whatever AI of this sort that I believe is most likely to beat me.
Other details:
I will record the match.
I won't bet in this market.
Inspired by /VerySeriousPoster/will-computerusing-ai-agents-play-b
1,000People are also trading
For anyone who's curious about what Factorio PvP is like with people who somewhat know what they're doing, this video is a good showcase:
In this PvP mode at least, the players' starting base is in a box surrounded by turrets, so they have to get a setup that can break through the walls and defenses before they can kill the other players.
@Multicore Thanks!
One thing to point out with team vs team PvP is that some people can attack while the rest of their team is building or defending. That isn't possible with 1v1. Factorio doesn't have RTS-style units, like StarCraft or AoE2 (I think the spidertron might be an exception, but that's very late-game).
@singer BTW, I've just found that tanks can now be operated remotely.
If one can multitask a bunch of those effectively, they can screw the opposing base pretty hard.
And then try again.
@ProjectVictory not even close to same. This one is "find some cheese to shoot me on the first try" while the other is "learn the whole thing on the go".
@ICRainbow Not sure how serious you're being, but the third point in the setup of this market is for stopping "find some cheese to shoot me on the first try".
@singer I'm pretty sure if it ends up winning it would do it by catching you off guard with some stupid trick, and not by fair gun play.
@ICRainbow Factorio isn't a shooting game to begin with. The military aspect centers on attrition and automating base defenses. If an AI actually snuck into my base like James Bond and shot me while I had forgotten my own gun somewhere that would be more impressive than winning the long-distance war of attrition. Realistically if I lose it will be because of physical exhaustion after 12 hours of playing (I don't think that will happen either though.)
@singer Definitely. I'm not making a point that it'll do cheap tricks. But long stalemate seems unlikely to me.
@TimothyJohnson5c16 I would say so. But playing StarCraft by watching the screen and controlling a mouse and keyboard is very challenging even for modern AI. The successful neural network bots I'm aware of that can play fast-paced games all get to tap into the game engine somehow (which this market prohibits).
@singer 1. Have an AI or just a really in-depth script which plays the game extremely well given (possibly partial) information about the current or very recent state of the game.
2. Put most of the effort into extracting information from the pixels quickly and accurately. For example, moving the mouse one tile every frame and looking at the tool-tip on the right side of the screen which instantly displays exactly what is on that tile, in a nice standardized format that OCR or similar can easily read.
3. Given a constant stream of updated game information (from only pixels), and using that to figure out the next best game state (or a decent enough next game state), convert that to keyboard and mouse actions for the next few frames.
4. Repeat.
@EmilyThomas You could get that working after lots of trial and error. But I don't expect someone that talented to be devoting time to develop a system just to fit into the resolution criteria of this market (whether it was for StarCraft or for Factorio).
It's interesting to me that this is priced above 10% (currently at 43%) since it requires AGI at the minimum. Meanwhile, @VerySeriousPoster's Balatro market is at 35%, which is about a turn-based card game.
@singer I don't know much about Factorio but an awful lot of things have been said to require AGI and then been accomplished without it.
@SamuelMillerick Yes, that's a good prior to have. You might be imagining a highly specialized Factorio-specific agent, without vision or language capabilities, that was trained from scratch on millions of games of self-play (kind of like the OpenAI 5 bot). It's definitely technically possible for an agent like that to hand me my ass in Factorio, while still lacking general intelligence. But there's some reasons I'm certain that won't happen next year:
To build an agent like that, a major AI lab would need to spend lots of money and research time, like OAI did with their Dota 2 bot. It would have to be a new project specifically about Factorio for some reason, which is an obscure game. AI labs in this decade don't have much to prove by creating projects like that.
To qualify for this market, the agent would need to use just screen pixels and the mouse/keyboard. That's very atypical for an agent project built specifically for a single game.
Labs are moving to full-generalist agents that control GUIs by watching the screen and using the mouse/keyboard. If there's an AI which can resolve this market YES next year, it will almost certainly be one of those types of agents: a generalist with strong vision and language capabilities, and that had never played Factorio before. It would be an amazing agent to succeed, especially being able to run in realtime and mentally keep track of thousands of off-screen details, all in a game it wasn't built to play.
Therefore, my reason for thinking that a YES resolution would imply AGI has more to do with the priorities and research approaches of AI labs, rather than with Factorio itself. A generalist AI that can beat me at Factorio is almost certainly an AGI, and the only type of AI capable of that which could realistically exist by the end of 2025 would be a generalist.
@LiamZ The main environmental hazard is the alien bugs that attack you if your factory produces too much pollution, but the amount of production you'd do just to produce the supplies to fight another player in PvP doesn't seem likely to trigger any attacks.
@BrendanJackman That technology already exists. It just would lose to me within a few minutes. See: /singer/what-gui-tasks-will-claude-be-able
@singer can AI use vision and dispatch commands to bots or it would have to click the actions by itself?
@ICRainbow It has to use the exact same interface that I'm using, so I think the answer would be no it cannot dispatch commands. Certainly not to "bots" in the plural, it can only control one player.