I had an AI agent port Karpathy's GPT-2 trainer from CUDA to SYCL, and made it prove every step
Muse Port plans and ports real CUDA projects to the open SYCL standard. Nothing counts as done until a check the model never sees says the output matches.