# ARM engineer uses AI to speed up Linux kernel compilation

> Lorenzo Stoakes, a principal engineer at ARM, has used AI to develop a series of patches to speed up kernel compilation.
> By José Maldonado · Artificial Intelligence · 2026-09-10

Artificial intelligence keeps creeping into the most unexpected corners of software development. The latest example comes from Lorenzo Stoakes, a principal engineer at ARM and maintainer of the Linux kernel's memory management subsystem, who has submitted a series of 23 patches to drastically speed up kernel compilation with the help of a large language model (LLM).

As detailed in the [**Linux mailing list**](https://lore.kernel.org/rust-for-linux/20260908-build-speedup-v1-0-5dc1ac01672d@kernel.org/), the result is striking. On the one hand, builds are 36% faster with all modules enabled (allmodconfig), up to 70% faster for incremental builds and 90% for "noop" builds (those with no changes to recompile). But the most interesting part is not just the figures, but how they were achieved.

## Single-threading remains the bottleneck

The Linux kernel compilation process has long carried several structural problems. But one of them is a huge bottleneck: several critical phases run on a single processor thread, which slows overall performance even on machines with dozens of cores.

Stoakes identified that the dead time in these serialized sections was the main obstacle to reducing build times. This is a widely known problem, but no one had tried to solve it because of its complexity; Stoakes is the first.

The patches address multiple components of the build infrastructure: Kbuild (the main build system), kallsyms (kernel symbol management), modpost (module post-processing), objtool (object validation), mksysmap (system map generation) and the Rust build system. The proposal, therefore, is not a one-off change, but a cross-cutting rewrite to parallelize tasks that previously blocked the flow.

[FreeBSD ports AMD's ROCm software for HPC and AI on GPUs](https://observatorioblockchain.com/en/technology/freebsd-integra-software-rocm-amd-hpc-ia/)

## AI as a "detective", not as a programmer

It sounds simple to do, but it is a real headache. However, Stoakes did not directly use an AI to magically solve the problem for him; instead, he used the model in two phases. First, to locate the bottlenecks by analyzing the code and build logs, and then to explore possible solutions.

Stoakes himself sums it up with unusual frankness in his message to the kernel mailing list (LKML):

> An LLM was used first to determine where the bottlenecks were and then to figure out how to improve them. It generated a lot of code, much of it horrible. I audited and extensively rewrote much of it, and heavily edited the commit messages, the cover letter and the comments.

In addition, the model also orchestrated test builds, debugging and analysis, acting almost like an automated CI agent. But the final validation — both the correctness of the code and the verification of the resulting kernels and performance improvements — was entirely manual.

## The "Assisted-by" tag

To record the use of AI, all commits include an Assisted-by tag. This detail responds to an open debate in the kernel community about transparency in the use of AI tools.

In January 2026, Linus Torvalds had already weighed in on the discussion about policy documentation for AI-assisted contributions, noting that **"people who use AI slop are not going to document their patches as such"**. Stoakes, for his part, has for months defended the need for a clear policy and for opt-in by maintainers, arguing that the review burden falls on them and they must be able to reject patches generated without real human oversight.

His position was set out in the August 2026 "AI slop" thread, where he responded to Jason Gunthorpe (NVIDIA) defending the responsible use of AI **"on a short leash"** and anticipating that the community will have to move toward a trust model where new contributors earn credibility with small changes before proposing larger patches.

**[5 ultra-light Linux distros to revive your old PC](https://observatorioblockchain.com/en/technology/5-distros-linux-ultraligeras-resucitar-viejo-pc/)**

### Lorenzo Stoakes' track record

And perhaps that position is the one that makes the most sense, especially given Stoakes' own profile. Lorenzo Stoakes is no amateur: he is co-maintainer of the Linux kernel's memory management (mm) subsystem, responsible for critical areas such as reverse mapping, Transparent Huge Pages (THP) and the kernel's memory APIs. He has more than 20 years in systems development: Oracle, WorldQuant, Squarepoint Capital, Imagination Technologies, and since July 2026 at ARM working on KVM and core memory.

He has spoken at the Linux Plumbers Conference and at the prestigious LSF/MM (Linux Storage, Filesystem, Memory Management & BPF), and he is the author of a 1,300-page reference book on memory management in Linux (No Starch Press).

A track record that makes clear his technical credentials are unquestionable, which gives weight to his approach. After all, Stoakes argues that **"AI as an exploration tool, never as a substitute for engineering judgment."**

### What's next for these patches?

The series is under review on the kernel list. There is still no guarantee that all of it will make it into mainline, but the methodological precedent is already there: an experienced maintainer uses AI to speed up problem discovery, rejects its "horrible" code, rewrites what is necessary, validates everything by hand and documents it transparently.

In an ecosystem where trust is earned commit by commit, Stoakes' case offers a template for how to integrate AI without breaking the verification rules that have sustained the kernel for three decades.

After all, with these changes, Linux compilation has just become faster, something most distros will welcome and which we will possibly see repeated in other developments (e.g. gcc, whose bootstrap can take hours on certain hardware). And that is something developers can applaud, or perhaps not.

**[This is Hegotá, Ethereum's major upgrade for the post-quantum era](https://observatorioblockchain.com/en/ethereum/hegota-actualizacion-poscuantica-ethereum/)**
