Show HN: Pac-Bench – How well can models one-shot a Pac-Man game?

This reminds me of a couple decades ago when some friends and I set ourselves a challenge to, individually, each create as much of a Pac-Man clone as possible in 10 hours. None of us had any experience in games programming or graphics coding. It was great fun and we all learned a lot.

No-one ended up with a complete clone but I loved how we all ended up focusing on different things, like pixel-perfect graphics versus accuracy in gameplay, and how we all brought our existing skills to the challenge despite not really knowing what we were doing.

I expect if we had AI models available it would have ruined the pleasure of figuring it out for ourselves. I feel kind of sad for the next generation of developers who won't have that experience.

Interesting, recently I am working on my own clone of Pac-Man. LLM implementations lose lots of details. They are not 1:1 replication of the original game. For example, the behavior of the ghost is not the same as the original. If I have not implemented the game myself, I can not tell the differences. What LLM produced look like the original game, but they are not.

I tried this with DOOM. Fable 5 did a pretty shit job. Astra made pretty crazy animated sprites and was pretty good considering.

The fact that these are at all playable and 100x my programming skill level is pretty depressing from a certain pov. The ThreeJS dude posted ab how demotivated he was to continue his work, and while I was never a dev that did much with webgl, I commiserate.

This prompt is a good way to test how well models fill in missing context because it's so nondescript. They're definitely improving.

Remember when people considered you a genius for prompting with "You are a skilled writer....".

This would be a much better benchmark if it contained some twist which was not in the training set (e.g. pelicans on bicycles were not common before simonw).

Make a pacman game where pacman can always eat ghosts, but ghosts drop pellets.

I don't think it makes sense to have the prompt be that short. This is basically a bench.ark of how models interpret an overly vague prompt. It should at least be "Create a pacman clone in a single html page. Make it faithful to the original" if that's what we're scoring it on.

Here's GLM-5.2: http://show.ahnfelt.net/demos/glm5.2-pacman.html

Interesting the `high` effort on the Claude models. Did you find that going above that is unnecessary?

Curious if models grasp the ghost patterns or just react. Pac-Man's more complex than it seems for one-shot learning.

Recent and related:

https://news.ycombinator.com/item?id=49882889

Gonna be rude and say I don't think this is an interesting or useful benchmark tbh.

Clearly some new RLAAS/dataset/env is being used for this now (it doesn't even seem that complicated, you have one LLM judge whether gameplay is recognizable as the original game or not and another trying to implement a logically/semantically identical version of the game). It's why the performance improvement on this workload has been so dramatic.

Everything is going to go from 0->1 on this benchmark in short order because of that.

Pretty underwhelming. Surely they’ve ingested some existing code on the web and regurgitating that.

Clearly the high-tech plagiarism machine will successfully plagiarise one of the most plagiarised games.

Always assumed Pac-Man was an easy solve for modern AI. Guess those ghost patterns are trickier than they look for one-shot learning.

Tried opus 5.5 for the same?

Opus 5-5 seemed like a perfect clone, with others displaying flaws in an initial look. Astra notably created a bunch of surrounding ugly crap to look at.

I thought this was about playing a pacman game. I feel like we already established that LLMs can create simple games and "local" apps at this point

I'm curious; can you explain why you picked Pac-Man, in particular?

Someday we’re gonna miss the times where models create jank.

These results get better and better, but the half-baked pac mans games and pelicans are just funny.

You cannot prompt that level of brokenness. Same for those slop-posters you see everywhere.

It would be great to do the models tests but with different harnesses too to see how they impact on the same models and efforts.

coolio thanks

[dead]