Someone got an LLM running on a Commodore 64 from 1982, and it runs as well as you'd imagine

In the grand AI race, all sides are trying to get their LLMs working on the weakest, cheapest hardware possible. The weaker the hardware, the less power is required to run the LLMs. But how about getting an AI model running on hardware that came out in 1982?

That’s exactly what Maciej Witkowiak is trying to do with their new project. Titled “Llama2.c64,” this project aims to get a simple AI running on a Commodore 64, which came out long, long before LLMs were even a thing. And while it definitely won’t be beating GPT-4o anytime soon, it’s still an impressive feat.

Llama2.c64 brings the new into the old, albeit slowly

As first reported by Hackaday, Maciej Witkowiak showed off Llama2.c64 to the world. As you might imagine, you’re not going to get high performance out of this hardware, but getting anything running at all is really cool. Here’s how Witkowiak describes it:

Ported to C64 by Maciej ‘YTM/Elysium’ Witkowiak using oscar64

Llama2.c64 is a port of llama2.c to the Commodore C64 equipped with at least 2MB REU.

It runs the 260K tinystories model bringing Llama2’s capabilities to the unique C64 hardware environment.

This is not a chat model.

Rather, imagine prompting a 3-year-old child with the beginning of a story — they will continue it to the best of their vocabulary and abilities.

You can see an example of Llama2.c64 creating a story in the image above. Sure enough, it does sound like a child rambling on while telling a story, but given how little the AI was given to work with and how much it continued with the story, it’s really amazing for a 42-year-old computer.