• Claude Opus 4.5 used as teacher to generate CUDA kernel skill files. • UpSkill tool automates generating, evaluating, and deploying agent skills effectively. • Skill files enable smaller, laptop‑friendly models to tackle hard GPU tasks. • Benchmark shows mixed performance: some models improve, others degrade. • Approach cuts costs and expands domain‑specific model capabilities. • Demonstrates practical workflow for upskilling agents on specialized problems.

Article Summaries:

  • A new workflow demonstrates how Claude Opus 4.5 can train smaller, open‑source models to write CUDA kernels. Using the “upskill” tool, the team first had Claude generate a kernel interactively, then convert the interaction trace into a reusable skill file. The skill is automatically tested against the original model and benchmarked on diffusers‑related CUDA tasks. Results show that the skill can match or improve performance while reducing token usage, offering a cost‑effective way to deploy lightweight models on domain‑specific, hard problems. The approach highlights agent skills as a practical medium for sharing capabilities across models.

Sources: