Linus Torvalds has lost the plot

Whether AI tooling is useful for programming is no longer a question. Nobody is questioning the man's judgement when it comes to using these tools responsibly, I don't think anyone believes Linus is 'vibecoding'. The problem with Linux kernel development embracing AI is far more profound than most people realize. The problem in question has been a fundamental issue for software long before these AI coding tools existed, only now it's being obscured even further.

Linus has chosen to ignore a dire warning given to us by one of the forefathers of the Unix operating system, Ken Thompson. The 'Trusting Trust' problem Thompson demonstrated in his 1984 Turing award lecture proved that a compiler can be modified to insert backdoors into everything it compiles, including future copies of the compiler itself, completely undetectable in the source code. This problem has never been solved. Wheeler's method for Diverse Double Compiling only solves it in theory and relies on the assumption of an already trusted 2nd compiler (Paul Jakma's 2010 paper critiquing DDC goes into more detail). Self-reproducing builds don't even come close because the core issue of the build tools isn't addressed at all. Bootstrapping also falls short because it relies on an existing binary (seed) to advance to the next version of the compiler, the chain of compromise has not been eliminated, the fundamental problem is never confronted in any meaningful way.


Whether you're using Claude Code from the terminal, an IDE with Cursor's models embedded within it or GitHub Copilot; the AI models you're using are upstream from the compiler, with greater privilege. No matter how much control you think you have, the tool you're using has become the developer, the reviewer & the build system. The difference between Ken Thompson's infected compiler and a compromised AI coding assistant, is that the coding tool doesn't need a malicious self-producing binary mechanism. It has access to the .env file, can make changes to a CI/CD pipeline and alter test fixtures.

What really amazes me is the wilful ignorance towards distillation; training on model on the outputs of another. Private companies with proprietary AI models simultaneously poisoning their data in an effort to stop other AI models from scraping data while depending on AI tooling that trains on the same ecosystem.

Go ahead and reverse engineer the binary you just compiled to see if there are any discrepancies, but did you compile the software that's helping you review the binary yourself ? What about the compiler you used to compile that software ?



https://redd.it/1uykaa9
@r_linux