Linus Torvalds ships his first LLM-assisted kernel patch series
Torvalds used Claude Code to write a kernel patch series — and the kernel's own reviewers found a race in it within a day. Also: Anduril prices a $1.3 billion employee tender, and the Baker lab publishes two Nature papers on AI-designed enzymes.
Linus Torvalds has posted his first patch series written with a large language model — and he says so in the commit message. The series, aimed at a long-standing page-refcount scalability problem, makes folio lookup notice when folios are hot, delays releasing them until after an RCU period, and then does the actual folio data reads inside that RCU window so the fast path needs no refcount at all. Torvalds attributes the work plainly: "The code, the test programs, the review behind the list above and the first draft of this text were done with an LLM coding assistant (Claude Code) under my direction." He says the change produced a 23x performance increase on his Ryzen Threadripper, lifting a threaded read benchmark from 13.3 million to 304 million reads per second, after profiling showed contended refcount cache lines accounting for 99.92% of measured core load. The honest framing is narrower than the headline number: on an ordinary laptop the gain is closer to threefold, and Torvalds himself calls the 23x figure unrepresentative of real workloads.
Why it matters: this is the kernel's top maintainer demonstrating the workflow he spent the last month arguing about — AI as a tool for people who already know the codebase, not a shortcut past it. The series is not merged and not in mainline; it lives on a branch, and a reviewer has already reported a race in the new batched read path where data can be read that was never written there. That is the actual signal here. An LLM-assisted patch series reached the mailing list, got a serious technical review, and found a real bug in a day — which is roughly what good kernel review is supposed to do, and a much better advertisement for the tools than a clean merge would have been.
Anduril has opened a tender offer of up to roughly $1.3 billion for employee shares at a valuation near $90 billion, according to Bloomberg. The defense-tech company is buying back stock at $94.80 a share, and for the first time the offer extends beyond the United States, United Kingdom and Australia to staff in 16 countries. It opened Tuesday and closes November 9, and both current and former employees are eligible. Anduril declined to comment. The mark is a steep step from the $61 billion valuation it carried in May, when a $5 billion round led by Thrive Capital and Andreessen Horowitz roughly doubled the $30.5 billion it was worth in June 2025 — and it puts the company's price above most of the AI infrastructure names it shares a capital market with.
The number to watch is not the valuation but the revenue behind it. Anduril has disclosed about $2.2 billion in 2025 revenue, roughly double the prior year, and is projecting around $4.3 billion for 2026, with a Pentagon enterprise agreement whose ceiling runs to $20 billion over ten years. Tender offers at this scale are how private companies let early staff cash out without a listing, and doing it across 16 countries says something about how far the hiring base has spread.
David Baker's lab at the University of Washington has published two Nature papers showing that AI design tools can build working proteases — enzymes that cut other proteins at a chosen sequence. The team kept only the catalytic core of two natural enzymes, a zinc protease from a crayfish digestive system and a cysteine protease from yeast, and let the design pipeline rebuild everything else. RFdiffusion2 generated new backbones under constraints that force the catalytic residues together and hold the target sequence nearby, ProteinMPNN filled in the amino acids, and AlphaFold3 checked that the results did not simply match known structures. The cysteine work screened 12 backbones with 69 sequence sets and got 13 active designs — a 19% hit rate. The zinc work tested 135 designs; the best beat the uncatalysed reaction by more than 100 million-fold, and swapping four amino acids near the active site pushed that past 10 billion-fold.
The clinical angle is what makes it more than a methods paper: the team designed zinc proteases against TDP-43, amyloid-beta and serum amyloid A, the proteins implicated in frontotemporal dementia, Alzheimer's and AA amyloidosis. The limitation is equally clear. The tools are excellent at structure and indifferent to activity — they will happily build a protein that folds correctly and cleaves slowly, which is why the best variants came from random mutation around the designed core rather than from the design itself. We covered the delivery side of AI protein design earlier — an AI-designed protein delivers RNA better than lipid nanoparticles — and this is the harder half of the problem: not getting a protein where it needs to go, but making it do chemistry once it arrives.
What to watch: whether the folio series gets its race fixed and lands in mainline, and whether a 16-country tender becomes the template for defense-tech liquidity.
Should a first patch series written with an AI assistant get held to a different review standard than a human's — or exactly the same one? Tell us in the comments.
