What matters in AI.

Subscribe

Benchmarks

AI agents wrote GPU code with 1.57× the speed of `torch.compile`

The author tells that the correct test was more work than the fast code.

Claimed, not confirmed

This is a brief. We point to the report and do not rewrite it. Read it at the source below.

A 381× speedup claim versus a fair 1.57× Sakana AI claimed a 381× speedup over PyTorch in February 2025 and retracted it within a day because the test let wrong kernels pass. Two bars on one scale: 381× is full width, 1.57× is a tiny sliver. SAKANA AI, FEBRUARY 2025 381× claimed speedup over PyTorch 17,000 kernels Retracted within a day Test let wrong kernels pass 381× 1.57× author's best, vs torch.compile
Sakana AI claimed a 381× speedup over PyTorch in February 2025, then retracted it after a flaw in its test was found.

Sources

  1. AI Agents Beat PyTorch: Writing Faster CUDA Kernelstowardsdatascience.com
AI MATTER · NEWS · AI MATTER · NEWS ·8 OCT2026

Posted

Learn the terms in this story

More in Benchmarks

All Benchmarks news