We Got Claude to Build CUDA Kernels and teach open models!

A new tool called `upskill` enables the transfer of capabilities from large, state-of-the-art models to smaller, open-source ones. The process involves using a powerful model like Claude to perform a complex task, such as writing a CUDA kernel, and then capturing its successful process into a reusable "skill" file. This skill, containing instructions and code, can then be provided to a less capable model to enhance its performance on that specific task. The `upskill` tool also includes an evaluation framework to measure the impact of the skill on different models by comparing accuracy and token efficiency. This method allows for cost reduction and enables smaller, locally-run models to tackle more difficult and domain-specific problems.