Garry Tan urges US open‑weight AI labs to also produce distilled versions of frontier models.
He argues that distillation lowers compute costs and broadens access, aligning with the open‑weight mission. By providing smaller, efficient models, labs can enable more developers to experiment without needing massive resources. The suggestion targets the emerging competitive landscape of AI research.




