Hugging Face (Twitter)

RT @ben_burtenshaw: We got Claude to teach open models how to write CUDA kernels.

This blog post walks you through transferring hard capabilities (like kernel writing) between models with agents skills. Here's the process:

- get a powerful model (like Claude Opus 4.5 or OpenAI GPT-5.2) to solve a hard problem
- convert that trace into an agent skill
- transfer it to open-source, cheaper, or local model
- measure if it actually helps

We tested this on a gnarly task: writing CUDA kernels for diffusers. The results? Some open models saw +45% accuracy improvements with the right skill.

But the skill didn't help every model equally. Some even degraded performance, or used way more tokens. If you're transferring skills, you should evaluate.

We used upskill, a new tool for generating and evaluating agent skills. It works like this:

uvx upskill generate "write nvidia kernels" --from ./trace.md