LLM agents play Pokemon showdown
olivergrabner.comAs a tangent, there has recently been drama with AI and Pokemon Showdown because someone trained a bot (using standard reinforcement learning + a rewrite of the Pokemon battle engine to Rust for max self-play parallelism) that is apparently good enough to get to the top of the leaderboard.
That could be good news.
We may are approaching the point that online games become just playing against bots. I would love to see arcades and gaming cafes back. Real human connection instead of playing against bots. Ironically. For many years. The internet was the place to play against humans. AI, decompilation, and other tech may turn that around.
Yeah I saw that too which was partly the inspo for this experiment. I feel like its kind of getting to a place like in chess where AI is getting better and while theres alot of drama right now, I feel like it'll be something like stockfish in the future
Oh, I haven't been keeping up with Showdown news lately... which format? I assume either OU or randbats?
> Pokémon doubles is a game-theory environment without full information. The opponents' sets are hidden[...]
I'd be curious to see how a model fares in VGC rules with open team sheets, since that's where you swap out information-hiding games for good old-fashioned mind games :P
> [...] and damage is random [...]
Not to nitpick, but since damage isn't truly random, did it take possible damage rolls into effect? Being able to tell whether your opponent has a damage-boosting item/ability matters a lot!
Did you try out different reasoning levels? Wondering if messaging is basically a proxy for additional reasoning
what was the selection for the teams? random per round? how does it play against people in ladder?
Yeah! So I did teams from a fixed, random seeded pool and then would rerun it with the switched teams. It also doesn't play against ladder yet but I played against it and it's not horrible. It plays better than an amateur but I was able to beat it a good amount
That makes sense -- these are details that would be good to have in the methodology :)
Would be interested on the results from running a similar experiment with more powerful models, though that's more of a curiosity thing than anything lol
Pokémon Showdown isn't a game so much as it is competitive Wiki reading.
Not going to lie I agree lol but still interesting to see
Jev will give better results on this kind of thing.
For sure there are many better models for this use case[0]
At the same time, it’s interesting seeing how generally capable language models/LLMs/agents/harnesses are getting
We’ll probably see a lot more experiments and benchmarks like these in the near future
0: simple classifiers can play quite a few game, including Doom https://playground.jeffyclassify.com/#doom
Yeah Jev can play it but Jev cant communicate with its teammate in the 2v2 + it cant reason and is just a decision model
I agree with you
> but Jev cant communicate
However, it can chat https://news.ycombinator.com/item?id=49778162
Not very well of course :)