AI writes OpenAI accelerator kernels in 48 hours, outperforming experts

OpenAI engineer Chris Leary (formerly the creator of the XLA compiler at Google) published the article “Compilers 2.0: AI as stochastic optimizer” on September 1, 2026, showing that AI can derive optimized kernels for OpenAI accelerators with the same semantics in 48 hours, often outperforming manually tuned expert versions.

Compiled by Sergey KostenchukPublished 2026-09-02Updated 2026-09-18

2026-09-02 Coding OpenAI

💻 AI writes OpenAI accelerator kernels in 48 hours

OpenAI engineer Chris Leary (formerly the author of XLA at Google) published the article “Compilers 2.0: AI as stochastic optimizer.” Example from HotChips: from code almost indistinguishable from numpy, AI derived an optimized MLA kernel Jalapeño with the same semantics, often outperforming expert versions.

🌍 In OpenAI’s pipeline, the LLM replaces manual emitters and heuristics: correctness is guaranteed not by reading the code, but by checking semantic equivalence through strict mathematical contracts. Hardware performance depends on the number of attempts by the AI optimizer.

👤 AI does not “understand” assembly — it searches for moves in the program space, like STOKE, but at industrial scale. If the operation is mathematical and there is a strict equivalence contract, the contract should be checked, not the code.

Source 1: https://x.com/cdleary/status/2094878051238887834 Source 2: https://news.ycombinator.com/item?id=49530143

Sources