Linus Torvalds has walked back his earlier optimism about large language models boosting programmer productivity tenfold. The Linux kernel maintainer now argues that LLMs are generating more low-quality code and unhelpful reports than genuine contributions to the project, forcing developers to spend time filtering out errors and false information rather than advancing actual work.
The reality, according to Torvalds, is messier than the hype suggested. While LLMs can produce volume, much of it creates overhead. Developers waste cycles catching mistakes and chasing dead ends introduced by AI-generated suggestions—a friction that erodes rather than multiplies productivity gains.
That said, Torvalds isn't dismissing AI entirely. He acknowledges that the technology has already proven useful for identifying real security flaws, including historical vulnerabilities that might have otherwise stayed hidden. The practical value exists, but it's narrower and more selective than the broad 10x multiplier he once entertained.
