Research

• KernelChecker: telling whether a GPU kernel written by a language model is actually correct, not just close enough on a few test inputs.
• OmniVLA on the edge: shrinking a large vision-language-action model until it fits on the robot it's meant to drive.

More on GitHub.