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.
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.
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.
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.
>I wonder if we need better programming abstractions/ languages that can make programming easier for people.
That was nodejs, and it turns out people unwilling and or unable to improve things through systematic skill refinement just degrades into chaos.
Successful ecosystems are just pedantically lame enough to keep silly folks from going YOLO, but empowering enough to allow people to still have fun.
>which model has the best training code that was Pac-Man.
You mean which model is more cautious about copyright bleed-through of the $9Tn in FOSS and user code they misappropriated though isomorphic plagiarism. =3
My hope is that this drives a new generation of hyper creative content from those who don't give up. I mean, of course it does pacman well. Pacman and its clones have been done to the death over decades. But what if we ask AI to write finnegan's wake 2?
I’m quite optimistic about how AI tooling will raise the floor for creative work.
I personally have been working on a Three.js project with Opus 5 and 5.5 that I never would have continued with had I needed to dive into documentation by hand.
Seeing immediate results is incredibly motivating.
I'm finding new things to demotivate me. It's good and quick at doing everything I ask but sometimes as soon as it comes together I realize I didn't really want this or it's just not fun.
But maybe iterating quicker and finding that out sooner is still good. I don't know.
Try coming up with something that doesn't exist yet on the web, one of the good models and a harness will be able to do it. Yes it's hard to believe but you can't deny the current reality.
The game mechanics are shockingly hard for them to get right. You also reveal some of their personality traits. The GPT‑6 and Astra models tend to include a lot of cringe copy as well as their own personal graphics style.
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.
Yes, Opus 5.5 is the first one that I don't have any notes on. It's just as impressive on other tasks I've given it over the last week. A clear step change.
I was just coming here to say this. I looked at all of them, and Opus 5.5 at high effort seems significantly better than every other clone.
Buffered controls, different pathfinding AI for each ghost, level transitions, etc.
Gpt-5.6 sol high technically completed the assignment as well, but the gaps within the pipes used to construct the maze, the lack of pause when eating a ghost, etc made it feel far less polished.
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.
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.
Remember when people considered you a genius for prompting with "You are a skilled writer....".
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.
It seems like a lot of the programming is cookie cutter type stuff (where these ai programs shine) and should be easier.
They can be amazing, but promoting “make a pac man game” seems like an exercise in which model has the best training code that was Pac-Man.
That was nodejs, and it turns out people unwilling and or unable to improve things through systematic skill refinement just degrades into chaos.
Successful ecosystems are just pedantically lame enough to keep silly folks from going YOLO, but empowering enough to allow people to still have fun.
>which model has the best training code that was Pac-Man.
You mean which model is more cautious about copyright bleed-through of the $9Tn in FOSS and user code they misappropriated though isomorphic plagiarism. =3
Nodejs wasn't easier for people. JS is a horrible language where simple stuff like comparisons, array access, and member access are broken.
I personally have been working on a Three.js project with Opus 5 and 5.5 that I never would have continued with had I needed to dive into documentation by hand.
Seeing immediate results is incredibly motivating.
https://news.ycombinator.com/item?id=49882889
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 has the advantage of taking good screenshots and letting me know if it’s decent within five seconds of playing.
Buffered controls, different pathfinding AI for each ghost, level transitions, etc.
Gpt-5.6 sol high technically completed the assignment as well, but the gaps within the pipes used to construct the maze, the lack of pause when eating a ghost, etc made it feel far less polished.