August 2, 2026
Old computer, new AI, instant chaos
Autoregressive Language Model on the 6502 Processor
This 1970s school computer just played AI—and the comments instantly got chaotic
TLDR: A developer got a tiny text-writing AI to run on a 1970s-style home computer with almost no memory, which is wild on its own. The community loved the stunt but instantly argued over whether it should’ve been written in lower-level code, turning admiration into delightful nerd drama.
A developer somehow squeezed a tiny text-generating AI onto a 1975-era 8-bit chip with barely enough memory to hold a modern selfie, then ran it on his dad’s old BBC Micro. The machine slowly produced a dreamy, slightly cursed fairy-tale sentence—part adorable, part haunted fridge poetry—and the community absolutely ate it up. For many readers, the real magic wasn’t the output. It was the sheer audacity: training the model on a MacBook, then loading it onto a vintage machine through a DIY audio cable pretending to be a tape drive. Yes, really. If that sounds like a tech fever dream, that’s because it is.
But the comments quickly split into classic internet factions. One camp was pure awe: people called it “super cool,” “amazing,” and even started dreaming about tiny offline AI in smart glasses. Another camp immediately turned into backseat retro-engineers, arguing the creator left performance on the table by using C instead of hand-written assembly. Translation for non-nerds: they loved it so much they started heckling the build like sports fans screaming at the coach. Then came the history buffs, who were obsessed with the idea that this kind of AI demo might have been technically possible half a century ago. And of course, someone had to ask the ultimate vintage-computing question: why not use extra memory tricks and stream the data in? In other words, even when the machine is from the 70s, the comment section is very 2020s: equal parts wonder, nitpicking, and glorious nerd drama.
Key Points
- •The project trained a tiny Mamba-based autoregressive language model and ran inference on a real BBC Micro using the 1975 MOS 6502 processor.
- •The final system fit within about 25KB of user-space memory, using 9KB for inference code and 13KB for model weights.
- •Because the 6502 lacks multiplication and only supports 8-bit integer operations, the implementation relied on BitNet ternary quantization to replace multiply-heavy operations with add/subtract steps.
- •The deployment workflow used cc65 for compiling C to 6502 code, PlayUEF plus a custom 3.5mm-to-tape cable for transfer, and sim65/jsbeeb for testing and validation.
- •To stay within the parameter budget, the model used a 27-token vocabulary and a hidden dimension of 56, with 4 BitNet parameters packed per byte for faster unpacking.