A model runs thirty-five hours without a human

Alibaba’s Qwen team reported that Qwen3.7-Max worked a kernel-optimization task for 35 hours with no human intervention, running 432 kernel evaluations across 1,158 tool calls and finishing with a 10x geometric-mean speedup over the Triton reference (X post). The newsletter files it under the line that recursive self-improvement is no longer theoretical: the model is optimizing the accelerator kernels that models run on, here an Extend Attention kernel for Alibaba’s own T-Head PPU, hardware the model had not encountered in training, the same loop as GPT-5.5 topping KernelBench three weeks earlier and Opus 4.6’s 34x training speedup in February. On the autonomy axis it lands two weeks after METR’s suite topped out at sixteen hours and six weeks after UNC’s seventy-two-hour unattended run; the next code-speedup result in the corpus is the 52x speedup where a skilled human reaches 4x in June, and the kernel thread itself resumes with a 232-fold kernel speedup found by an auto-research loop in August.