An AI wrote every weight of this network
No training, no examples, no code: the model was given only a description of the input and wrote the 1,210 numbers of a tiny 64-16-10 network by hand. Draw a digit and see what it answers.
Draw one digit (0-9)
8x8 blocks of ink, 64 numbers
The 16 hidden neurons: what each one looks for (yellow = fires on your digit)
Score on 1,797 real handwritten digits it never saw
How these networks were made
One prompt, sent once to each model, on Nov 3, 2026. The prompt described the 64 inputs (an 8x8 grid of ink counts) and how digits are usually drawn, and the exact format of the answer. It contained no example digits and no data. Tools were off and no tool was used: GPT-6 Astra ran at maximum reasoning effort through Codex (read-only sandbox, 10 min 17 s, 0 commands); Claude Opus 5.5 ran at high effort through Claude Code with every tool disabled (7 min 51 s). A second Opus 5.5 run at maximum effort was stopped after 65 minutes without an answer. The weights are published exactly as the models wrote them.
The two baselines are the same 64-16-10 network: random weights, and the classic way, trained with gradient descent on 1,000 of the test-set digits (so its score on all 1,797 includes digits it saw; on the 797 it never saw it gets 94.4%).
The exact prompt, the scripts that sent it and scored the answers, and the full answers are in the Ruocco Academy. More demos: ruocco.it/ai-demos.
Test digits: UCI "Optical Recognition of Handwritten Digits" (Alpaydin & Kaynak, CC BY 4.0), test set as shipped in scikit-learn. Experiment inspired by "Sparks of Weight-Space Programming" (MIT CSAIL, Nov 2026). Made for VitoMind.