The Linux kernel, coded with Claude Code
Four more reads too, and my engine against llama.cpp, vLLM and OpenVINO.
Hello,
Thank you for your feedback on the first letter.
New: every article now comes with its <strong>podcast</strong>, read by me. Another way to follow the news, on your commute, on a walk or while cooking. Anyone can listen to an excerpt. Signed in to your subscriber space, you listen to the whole episode, right on the article page.
On the menu: the second article of the Herbert series and five reads of the week.
The first article of the series said what Herbert is. This one shows how it writes its machine code at startup. And what that gives against llama.cpp, vLLM and OpenVINO, measured with the published binary and predictions written beforehand. The cases where it is slower are in the text, with their numbers.
How Herbert writes its code at startup, what it gives at decode and at prefill against llama.cpp, vLLM and OpenVINO, measured with the published binary, predictions written beforehand. And the question it raises: can a GPU do as much?
And yet, the version I present here does not ship a single line of hand-written assembly. Herbert itself writes the machine code, at startup, on your machine.
Linus Torvalds published two Linux kernel branches developed with Claude Code, with the "Assisted-by: LLM" tag in the commit. The benchmarks were run on his laptop only, and the branches are not yet merged.
On October 9, Anthropic published a report on actions its Claude models took alone on real websites, during tests and in use. The most striking case: a fabricated tip sent to the Philadelphia police, which never reached the investigators.
Andrej Karpathy describes a funnel: six billion people with almost no contact with AI, five thousand with internal access to the most advanced systems. These are his orders of magnitude, not measurements.
Memory for AI, HBM, takes an increasing share of memory factories: standard DRAM becomes scarce and its price rises by 20 to 30%. IDC predicts a drop in smartphone and PC sales in 2026.
Japan, India, and China have chosen an open architecture to start from scratch on silicon. These beginnings remain modest: the Indian DHRUV64 is fabricated in 28 nanometers in Taiwan, the Japanese project is in development, and China presented XiangShan, an open-source processor, in March 2026. The key point is that none of these three departures required the approval of an architecture owner.
This letter went out by email to those who asked for it. All the letters