Developer ports Karpathy's GPT-2 code from CUDA to SYCL using AI agents

On November 7, 2026, a post on DEV Community's AI section described a developer project built for Meta's Global AI Developer Hackathon. The author created Muse Port, which uses Meta's Muse model together with an AI agent to move a GPT-2 training codebase originally written in pure CUDA onto SYCL, and adds a separate verification mechanism to make sure the ported code produces exactly the same output as the original.

The motivation is portability. A lot of AI software is written in CUDA and therefore only runs on NVIDIA GPUs, whereas SYCL, an open standard from Khronos (the group behind OpenGL and Vulkan), lets the same C++ code run on NVIDIA, AMD, Intel, and plain CPUs. Shifting code from CUDA to SYCL has traditionally been slow, careful work. According to the author, AI coding agents are now capable enough to take on much of it. The remaining problem is trust: an agent saying it is done is not the same as the code actually being correct, which is why Muse Port pairs the porting work with independent validation.

The material only contains this one first-person account, so details such as the verification setup, the resulting performance, and any subsequent progress are not reported. As of the post, the project is presented as a hackathon entry built on Meta's Muse model.