Codex hand-writes neural net weights to build a digit recognizer hitting 95.1% without any optimizer
GenJeppo · reddit · 2026-11-10
A Reddit user ran an experiment: have Codex directly design neural network weights (no optimizer, no benchmark drawings as templates) to build a recognizer for digits drawn by mouse on a 32×32 grid.
Method and results:
- Codex started from pen strokes, wrote formulas for variations in slant and thickness, and simulated coarse-scan gap artifacts; each synthetic drawing became a detector (a numeric vector) scoring matches
- The final recognizer combines fixed detector weights with hand-written geometry: first version scored 66.4%, final version hit 1,709/1,797 correct (95.10%) on scikit-learn's handwritten digits (resized 8×8→32×32, binarized)
- The recognition code, including weight generation, is 808 lines of Python (35.6 kB)
- Rules were revised after inspecting labeled mistakes on the same 1,797 drawings, so 95.1% is a development score
- Open source: github.com/jesper-adolfsson-001/digits-by-hand, reproducible with uv run drawshapes.py
Takeaway: the test was whether Codex can turn a visible difference between digits into a human-readable detector — it can.
Related event: Codex Handwrites Weights for a 95.1% Digit Recognizer Without Training(2 posts)→
More from coding & agent
- Claude Fable 5.2 tops FrontierCode 2.0 at 49.1, at a third of Fable 5.1's cost — nrehiew_ · 2026-11-10
- Victor Sanh: overfitting to users is the job when building vertical agents — VictorSanh · 2026-11-10
- Anthropic: Claude Fable 5.2 designs working electronic hardware end to end — EricBuess · 2026-11-10
- Anthropic's 1,000 agents built their own database tables and split into subteams to work in parallel — RLanceMartin · 2026-11-10
- Hardest Fermat lemma spawned a 2,418-task tree across ~750 agents, 10 levels deep — RLanceMartin · 2026-11-10
- Fable 5.2 builds a playable voxel engine from a single prompt — mattshumer_ · 2026-11-10