alt.hn

4/8/2026 at 12:32:56 AM

Show HN: Can an AI model fit on a single pixel?

 https://github.com/dvelton/ai-pixel

by deevelton

4/8/2026 at 3:00:30 AM

This was just a fun thought experiment.

Feed it two inputs (e.g. chance of rain and wind speed; this is one of the examples in the demo) and it learns to answer a yes/no question like "bring an umbrella?" It's a one-neuron binary classifier with three learned parameters: two weights and a bias. Those three numbers map directly to Red, Green, and Blue. Save the model: you get a 1x1 PNG. Load the pixel: you get your classifier back. The color is the model.

by deevelton

4/8/2026 at 6:22:48 AM

It is fun. The trainer seems to get stuck in local minima though.

by mememememememo

4/8/2026 at 6:41:07 AM

Thanks; does this fix it? https://github.com/dvelton/ai-pixel/commit/ba04e1c649e0ec99b...

by deevelton

4/8/2026 at 7:42:52 AM

Tried again and no. To reproduce create a corridor e.g.

    X    O
    
     X     O

      X   O
       X     O

       X   O
Line can get jammed in non vertical positon.

by mememememememo

4/8/2026 at 12:52:49 PM

Increased default epochs to 2000 instead of 500, I think that fixed it but let me know if you still run into this.

by deevelton

4/8/2026 at 9:02:10 AM

This is super fun! Very elegant idea

Has anyone done this for larger neural nets? Is there a way to extract some kind of pattern or is the image just noise no matter how you construct it? I'd be curious to see something like that

by crackalamoo

4/8/2026 at 1:01:37 PM

Not sure what happens if you start scaling this, but curious too. Would love to see someone who actually knows ML try it.

by deevelton