← sources

Codex built a handwritten digit recognizer without training it. It scored 95.1% using fixed weights and hand-written rules. Not sure how impressive this is?

source: reddit (reddit:computervision)published 2026-11-09 18:18ingested 2026-11-10 02:37

retained text

I recently read that LLMs can write neural-network weights directly, and I wanted to see what that meant in practice. So I gave Codex (GPT6-Astra-Medium) a challenge : recognize the digits 0–9 as I drew them with my mouse on a 32×32 grid. One condition: no training.

Codex wrote formulas for pen strokes and varied their slant, curvature, and thickness. It even simulated what happens when a coarse scan makes a narrow gap disappear. Each synthetic drawing was converted into a weight vector, and a dot product measured how closely my digit matched it.

The recognizer ended up combining **fixed weights with hand-written geometry**. 

The first version scored **66.4%**. The final version got **1,709 of 1,797 digits correct (95.10%)**. The recognition code, including weight generation, came to **808 lines of Python, about 35.6 kB**. No optimizer or backpropagation was used.

One caveat: the benchmark was scikit-learn’s 8×8 handwritten digits, resized to 32×32 and converted to black and white. We looked at mistakes on those same drawings as we refined the rules, so **95.1% is a development score**. This took about 2h to do. Codex wrote the code and ran the evaluations. I set the challenge and tried it with my own handwriting. To me this is super impressive, what do you experts think?

document fetched 2026-11-10 00:23 · canonical https://reddit.com/r/computervision/comments/1xrz914/codex_built_a_handwritten_digit_recognizer