Linus Torvalds uses AI to debug an Intel GPU driver bug
10 points by Yogthos
10 points by Yogthos
The debugging using AI is great, as this sounds indeed like a horrible issue to debug and it's great that the fix exists now and is in the kernel.
However.
"My kernel has LLM output in it" is a ship-has-sailed situation, but it seems my kernel now has slop in it, and slop from Linus' hand. That makes me unhappy.
I have a hook which bans Claude from writing comments. I get two benefits from it: my coworkers are no longer subjected to Claude's verbal diarrhea, and it slows me down in a way that I've found helpful. I get to spend a lot of time in the code itself, which has been invaluable at catching bugs or suboptimal designs, both in what it generates as well as pre-existing ones. (A recent example was that the LLM's first cut used a second associated type in Rust when it should have actually forced the two associated types to go together via a trait. The second associated type provided a degree of freedom that should not have existed.)
For me a lot of engineering discipline with LLMs has been deliberately introducing roadblocks like this. It works well — I think the code I publish now, both the LLM-assisted and the hand-written gnarliest bits, is the most careful code I've ever written.
Out of curiosity, does feedback from the hook seem to have any effect on stopping the model from burning tokens to produce the undesired prose comments in future work? It would seem more useful if it worked that way as opposed to working as a filter on the output.
It does, yeah, for the rest of the session. Usually the LLM will spit something overly verbose out, get chided by the hook, then write much shorter comments that conform to the hook. Though I don't think there's a huge marginal difference in the number of tokens.
It's possible that Linus is still in the honeymoon phase with AI output. It really does feel quite magical, at least until you start noticing the same repeating patterns everywhere and it begins to feel gross.