‘A lot worse than expected’: AI Pac-Man clones, reviewed
Can anyone, armed with just a dream and an AI chatbot, create a rudimentary version of a classic arcade game? I asked several enthusiasts who took a stab at it, with very mixed results
Surprising, considering arcade game clones are already in github. These models apparently cannot be trusted to even copy and paste a fully working project if that's basically what you're asking for.
It's probably mashing them all together and it's not smart enough to fix the integration hell it creates for itself 😅
No LLM can do this properly. Diffusion models might. Huge asterisk on "might."
Not even a sci-fi ultrabrain is gonna bang out a game from a three-word description, character by character, in order. There's a hierarchy of structure that has to get filled-in. LLMs are frankly witchcraft for how often they can guess they'll satisfy a right-brace, three thousand words in the future, and then actually do it.
Diffusion models can go 'and here's a function that does mumble mumble' and then get around to it on another loop.
... damn, these results are shockingly good, for a robot being given several English sentences. For zero human effort being put in - these models reproduced all the basic functionality of the game in question? Any of these would pass for clones on calculators or Win95 shareware. People have lost lawsuits over Pac-Man knockoffs this close. I'm sure they play like crap, but so did many human efforts, in the decade after it came out.
Isn’t computers writing code how Terminator starts?
No, it's flying autonomous killing machines. So what Palmer Luckey's been doing since Facebook bought out Oculus. Different branch of this particular capitalist hellscape.
... damn, these results are shockingly good, for a robot being given several English sentences. For zero human effort being put in - these models reproduced all the basic functionality of the game in question?
Obviously not. The models were trained online, and that includes the code to hundreds of pacman clones on GitHub, all accompanies by the info in the prompt used.
Considering they had perfect training data, the output is shockingly bad
