Claude Fable 5.2 hand-writes neural network weights and rebuilds a 1990s vision pipeline

Weight-Space Programming builds neural networks like a 1990s researcher

Claude Fable 5.2 hand-writes neural network weights and rebuilds a 1990s vision pipeline

Weight-Space Programming has frontier LLMs write a network's weights directly, with no gradient descent. The author gives Claude Fable 5.2 a fixed CNN and asks it to set all 5,994 MNIST weights by hand, allowing only formulas and strokes. Across three runs it rediscovers template matching, naive-Bayes log-odds scoring, and hand-built loop and corner detectors, then rebuilds its classifier from its own measurements, reaching about 91% test accuracy, close to Adam after 50 gradient steps.

Claude starts like a 1990s computer vision engineer. Its first layer is a set of line detectors at four angles, written from formulas.

More from this day

2026-11-20