Sutro problems

A collection of small, self-contained problems used as benchmarks for the Sutro Group's energy-efficient learning research.

• matmul/ — 4x4 and 16x16 matmul
• sparse-parity/ — approximate sparse parity: recover the k secret bit positions at the lowest energy
• mnist/ — learn from labeled images and predict test digits
• mnist-a100/ — learn MNIST from 10,000 labels as fast as possible on an A100: mnist.score(my_method, difficulty=1), five difficulties

main holds only the problems and their record submissions. Open a pull request against main only for a result that qualifies for a leaderboard table. Experiments, prototypes, harness proposals, and unfinished or non-record results belong on other branches; push them there and link them from an issue or a pull request against that branch instead of merging them into main.