Hazy Research: agents are replacing abstractions — CUDA DSLs are heading to retirement

ricklamers · x · 2026-09-22

Stanford Hazy Research (Stuart Sul, Chris Ré) argues in 'Retire the Abstractions' that CUDA DSLs are headed for retirement because agents are taking over the job of cognitive offloading. Last year the team needed a hand-built abstraction layer — and months fighting race conditions and deadlocks — to write megakernels; this year, for an MoE megakernel, they deleted the abstraction entirely and let agents work through the complexity directly, carrying rough, incomplete abstractions in the prompt instead of C++ templates. Tasks already doable without abstractions (optimized GEMM kernels) are now nearly automated with the right prompt, and previously unmanageable complexity became manageable since agents act as a 'compiler' from vague instructions to code. Rick Lamers, who shared the piece, notes his Astra agent agrees; Martin Casado adds that for OS-level work you can just feed the AI the paper and a setup and let it choose the implementation path.

Original post →

More from coding & agent

coding & agent channel →