Language-Free 'New Mind' Writes Working GPU Kernels

A 'new mind' system trained without natural language still generates GPU kernels that pass all tests about 28% of the time, with speedups of up to 66x over Triton kernels.

2026-09-27 ~ 2026-09-27 · 2 related posts

• Model trained without natural language writes kernels passing 28% of tests, up to 66x faster than Triton — cephaloform · 2026-09-27
• A 'new type of mind' trained without natural language writes kernels at up to 66x Triton speed — repligate · 2026-09-27