ChatGPT shrunk down to run on a Super Nintendo

offbeat

Just don’t expect complex, coherent, or always correct answers

If you’ve ever wondered how small an instance of ChatGPT could get, wonder no longer. Someone has actually rigged up a tiny version of the chatbot to run on a Super Nintendo cartridge. 

Tech journalist and retro tech enthusiast Benj Edwards is responsible for ChatGPT for SNES. He’s published it online running in an emulator, complete with SNES-era music and sound effects, and for download as an SNES ROM, too.

It’s not a big model, mind you, and performance is thus pretty limited: According to the website, it’s a 2.2 million-parameter offshoot of OpenAI’s GPT-OSS-120b, which was used to develop its training conversations – so in that sense it truly is a ChatGPT that runs on an SNES cartridge. Edwards described it to The Register as a distilled version of the aforementioned open-weight model. 

There are two options built into ChatGPT for SNES: One that’s purely the model, with every word generated coming right from its tiny little silicon brain, and another that includes a fact book culled from the opening sentences of around 150,000 simple English Wikipedia articles. If you opt to run both versions and ask it the same questions, you’ll notice a massive difference. 

Given some of the other responses we got while testing, there’s really no good reason to run the base model – memory isn’t really an issue given the basic build is about 4MB and the factbook build is just 19MB. 

Edwards admits that it’s not the most helpful AI model, either – this is just an experiment after all. 

“It is not a useful assistant: the model is millions of times smaller than the models behind ChatGPT, knows about 8,000 words, and often gets things wrong, sometimes confidently,” Edwards noted in his writeup. In that sense, it really is vintage AI on vintage hardware. “The fact book can carry mistakes too.”

The pure version on the left, and the factbook version on the right

The pure version on the left, and the factbook version on the right

If you have an SNES lying around (lucky you), you can download the ROM from this link. If you want to get it running on an SNES you’ll need a cartridge that allows running of custom ROMs, and that supports Nintendo’s Super FX graphics coprocessor. Cartridges like the FXPak Pro and SD2SNES support it. The factbook version also requires support for the aftermarket Media Streaming Unit 1 (MSU-1), which adds CD-ROM like features to the SNES. MSU-1 is available in both of the flash carts mentioned above.

If you have some real hardware skills, however, you could technically make your own ChatGPT cartridge. 

“The ‘pure’ version, which generates all responses from the AI language model, could run from a real custom SNES cart,” Edwards told us – it doesn’t even actually need the Super FX chip. “I do have a version of ChatGPT for SNES that can run without the Super FX chip, but without the extra processing power the answers are not as good.”

That said, Edwards believes its absolutely crazy and nonsensical answers are part of the fun. 

“ChatGPT for SNES can definitely make mistakes too, but it’s almost cute when it does it because of its limited knowledge and its playful interface,” Edwards told us. 

While a human may have designed the experiment, the coding work was done by AI, Edwards told us – he’s been interested in it since the 1990s and is pursuing more experiments with tiny AI models using the help of the bigger ones. 

“I’ve been theorizing about what I call ‘tiny talking machines’ … over the past year,” Edwards told us, and he’s been able to use coding agents to turn those ideas into actual software. ChatGPT for SNES is one example, and he also created an NES ROM called Séance that uses an even smaller AI model able to run on an original Nintendo Entertainment System that simulates talking to a ghost in a haunted game cartridge. 

“I designed the scope of what the software should be, how it should work broadly, what chip it should use, how the interface should look and work, and the sounds it makes when it types responses,” Edwards told us. He added that he sees the project as an AI/human collaboration, which is where he believes AI works best. 

“The AI agents who did the coding for this project currently advise that we’ve hit the limit of how good the AI model can be on the limited hardware,” Edwards said – he’s not taking their word for it, though. “It’s up to human creativity to transcend those assumptions. I have already come up with new ideas that I believe can make a tiny model like ChatGPT for SNES even more coherent, and I’m working on them now.” ®