Pyre Code ships browser ML exercises

This is a genuinely useful learning tool because it turns “read the paper, then code it” into a tight feedback loop instead of a notebook exercise. The local grading setup makes the practice feel closer to real engineering work, with hidden tests catching edge cases instead of rewarding hand-wavy answers. The problem set is unusually concrete for ML education: attention variants, inference kernels, alignment methods, diffusion, distributed training, and more. Browser-based Monaco editing removes setup friction while keeping the environment self-hosted and private by default. Optional AI help is a pragmatic addition, but the core value is still the implementation drill, not the chatbot layer. This looks strongest as an interview-prep and onboarding tool for teams that want deeper model literacy.