In a new video, ProjectCD.Chronicles shows off something that sounds like science fiction for 1983: a working neural network running in under 250 lines of BASIC on a stock Commodore 64.
The program teaches the C64 to recognize user-drawn alphanumeric patterns, even when the input is messy or incomplete. It borrows the idea of neurons and synapses from biological brains but trims it down enough to fit inside the modest memory of the C64 — without setting the thing on fire.
Unlike typical computer memory, which stores neat, fixed values, this project imitates human-style associative memory — the kind that lets you recognize a badly drawn cat faster than a processor could add two numbers. Users train the network by feeding it patterns, and when they input a scrambled version later, the system recalls the closest match by adjusting connection weights across two simulated neuron arrays.
Of course, it’s no brainiac. The network can only juggle a few patterns before it gets confused — about the equivalent of a dog trying to remember more than three tricks. Still, that limit highlights just how much wiring real brains pack in.
Originally pitched to Commodore magazines and rejected as “too esoteric,” this project feels less like an oddball experiment now and more like an early spark of ideas that would later fuel real AI development. It’s a sharp reminder that even early home computers were ready for serious experiments — if you knew how to push them.
Original Paper: Neural Network on a Commodore 64
https://www.fourmilab.ch/documents/commodore/BrainSim/
