Machine discovery

Machine learning was one level; the next level is machine discovery. At that level, machines find and generate the data needed to test hypotheses. Google’s discovery of fast matrix multiplication is a good example. I see one clear problem that could serve as a benchmark "optimization": we can assemble hundreds of training tasks as benchmarks, have code-generation models produce PyTorch optimizers, and evaluate them on speed and final performance. Once the setup is complete, little human involvement is required—though contributors can still add new benchmarks.