Rendered at 03:01:27 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
hung 1 days ago [-]
Shouldn't it be running at every tick of the game rather than just the junctions? Pac-Man should be able to change direction at any point.
felix089 17 hours ago [-]
Ideally yes but that would not work with the model latency. But it is already partly covered when a ghost enters its corridor it is asked again. Here's how the decisions work in detail, also below a copy https://github.com/opper-ai/jevman-benchmark/blob/main/READM...
How decisions work
When a character commits to a corridor its next junction is known, so the game asks jev about it straight away (one System One request per frame, one choice question per character, options = legal directions described with computed facts: distances to pellets, power pellets, fruit and ghosts, whether the nearest ghost is coming closer, and whether a ghost can reach the end of the corridor before Pac-Man). Up to three requests are in flight at once. If the character reaches the junction before the answer, it waits there (its panel card says "thinking…"). After 2 s, or on an error, it uses a greedy rule.
Pac-Man can also get a second question mid-corridor: when a dangerous ghost is in the corridor ahead, or can reach the junction at its end before he does, jev is asked whether to keep going or turn back right now (pacman_escape). Pac-Man keeps moving while it is open, and each situation is asked once.
felix089 11 hours ago [-]
Ideally yes but that would not work with the model latency. But it is already partly covered when a ghost enters pacmans corridor he is asked again. Here's how the decisions work in detail https://github.com/opper-ai/jevman-benchmark/blob/main/READM...
hyperhello 22 hours ago [-]
It may not matter if the ghosts' next move is predictable over that maximum distance.
theowaway213456 14 hours ago [-]
That requires much more thinking (because of calculation) though doesn't it? Which decision models are not optimized for.
hyperhello 9 hours ago [-]
Speculating, but it seems like needing to be able to change direction at any moment would be more calculating, not less.
felix089 11 hours ago [-]
Quick update, we doubled the models tested based on community contributions, and a new model took the #1 spot. also well over 1000 games have been played against decision model ghosts.
https://opper.ai/jevman-benchmark/#leaderboard
gsandahl 7 hours ago [-]
i guess a model trained on this game will end up taking the #1 at some point
NichoPaolucci 1 days ago [-]
This is neat. At least I'm still better than the models at pac-man.
Also love the idea of a shared pool for users to try things out. I was considering more of a crowdfunded approach for one of my toy projects, something like... Giving it $10 in credits to begin and somehow allowing users to feed a buck in if they wanted.
felix089 16 hours ago [-]
Great to hear, if it's about inference credits we actually built a thing called an AI wallet where users can just easily pay for their own inference on apps. if that's interesting lmk, id be happy to support this https://opper.ai/ai-wallet
Jev is useful to run locally overnight: it can classify the results while you sleep, ready for you to review in the morning.
fancyswimtime 20 hours ago [-]
Yes, try with toxic/toxichat and CFPB
soltanov 22 hours ago [-]
Running cloud API inference for Pac-Man is peak modern software engineering. A 2 MB local ONNX model on a CPU core beats network round-trip jitter every time.
ozozozd 1 days ago [-]
Super cool!
I wish the controls were a little easier on mobile.
felix089 18 hours ago [-]
Thank you and yea agreed fixing it now, try again in an hour from now :)
ismailmaj 16 hours ago [-]
it would be nice to have a regular LLM for reference
felix089 4 hours ago [-]
Someone actually ran the benchmark on Qwen 3.8 Flash Next NVFP4, which is a chat model and submitted it via PR. We merged and it's live on the leader board in the top 3.
How decisions work
When a character commits to a corridor its next junction is known, so the game asks jev about it straight away (one System One request per frame, one choice question per character, options = legal directions described with computed facts: distances to pellets, power pellets, fruit and ghosts, whether the nearest ghost is coming closer, and whether a ghost can reach the end of the corridor before Pac-Man). Up to three requests are in flight at once. If the character reaches the junction before the answer, it waits there (its panel card says "thinking…"). After 2 s, or on an error, it uses a greedy rule.
Pac-Man can also get a second question mid-corridor: when a dangerous ghost is in the corridor ahead, or can reach the junction at its end before he does, jev is asked whether to keep going or turn back right now (pacman_escape). Pac-Man keeps moving while it is open, and each situation is asked once.
Also love the idea of a shared pool for users to try things out. I was considering more of a crowdfunded approach for one of my toy projects, something like... Giving it $10 in credits to begin and somehow allowing users to feed a buck in if they wanted.
I trained some to do some interesting things, including playing doom: https://github.com/nicobrenner/jeffy
I’ll try training one for this benchmark, seems like fun
I wish the controls were a little easier on mobile.
What even is this reality.