Sources
Latent Space Features Vlad Feinberg: Kernel-Level Performance Work Is the 'Most Direct Path' Into Frontier AI Labs
Latent Space's May 19 newsletter highlights a blog post by Google Distinguished Engineer Vlad Feinberg (published May 10) arguing that kernel-level performance optimization — making abstract ML operations practical to run — is the biggest bottleneck in LLM work and 'the most direct path into the labs.' He recommends learning JAX, Pallas kernel development, and Chinchilla scaling laws. DeepMind researcher Aidan Clark endorsed the core skill advice but pushed back publicly on X, stating he 'cannot for the life of me endorse give up all your weekends and become a work machine.'
↳ Follow the thread