Wrote a blog post about AI-generated kernels for Apple devices to make models run faster. We saw improvements of ~1.87X on average, but it was necessary to use a swarm of agents - the swarm dramatically outperformed the individual agents.