A model wrote the fastest megakernel anyone's submitted, and beat every other frontier model doing it

Jack Clark's Import AI 464 (around July 6) led with something I've been turning over all week. Claude Fable autonomously wrote what Clark calls "the first genuine (and fastest) megakernel" submitted to the KernelBench-Mega leaderboard. An 18.71x speedup in hand-written CUDA on an RTX PRO 6000 Blackwell versus an optimized PyTorch baseline (Import AI).

The detail that got me: torch.profiler showed exactly one cooperative kernel launch per decoded token. Rival models decomposed the same work into 4 to 14 launches. Fable fused the whole thing into a single launch. And it did it in raw CUDA while the competition wrote Triton and still lost. Opus 4.8 hit 14.4x. GLM-5.2 got 11.14x. GPT-5.5 managed 4.34x. Fable nearly doubled the best of them.

Writing a fused megakernel is deep systems work. You're hand-managing shared memory, warp coordination, and launch overhead, the stuff that separates people who "know CUDA" from people who ship production inference kernels. A model did it better than an optimized PyTorch baseline, unattended.

Two things make this bigger than a benchmark score. One, it's a direct hit on inference economics. If models can author hardware-specific kernels that beat hand-tuned baselines, the cost curve for serving them bends, and the people who can do that optimization by hand just got a very capable collaborator or competitor. Two, it's a concrete recursive-self-improvement data point. Not a think-piece about AI improving AI. An actual measured instance of a model making the hardware that runs models faster.

This connects to a thread running through the whole week. swyx says Anthropic's "ultracode" is "scarily good at burning tokens" but only pays off if you architect your repo so subagents parallelize, "subroutines but intelligent" (X/swyx). Redis creator antirez is writing a local DeepSeek 4 inference engine and says well-managed automatic programming now beats "decently developed" hand-written code, even in high-stakes C (antirez.com). Systems programming was supposed to be the last redoubt. It isn't holding.

What to do: if you self-host inference, this is your signal to try model-generated kernel optimization on your actual hot paths rather than assuming it's toy-grade. It clearly isn't anymore. And keep your skepticism calibrated by story 4, because a headline speedup and a saturated benchmark are two very different kinds of number.