Linus Torvalds said his previous estimate that large-language models could increase programmer productivity tenfold was not based on scientific data.
As the creator of the Linux kernel, Torvalds' perspective influences thousands of developers and the broader open-source ecosystem. His shift in tone reflects the practical challenges of integrating generative AI into complex, high-stability software environments where a single error can have global repercussions.
Speaking Sunday at the Open Source Summit India 2026 in Bangalore, Torvalds addressed the actual impact of AI on the kernel development process [1], [2]. He specifically referred to his previous 10x productivity claim [1] as a rough guess. "That was pulled out of my ass number, obviously," Torvalds said [2].
Torvalds noted that the influx of AI-generated contributions has not yet yielded the efficiency gains some expected. He said the project saw more junk being generated by LLMs than useful code up until early this year [1]. This surge of low-quality patches creates additional work for human maintainers who must vet and reject flawed submissions, a process that can negate the time saved by automated coding.
Despite the current frustrations with "junk" patches, Torvalds expressed a long-term optimism for the technology. He said the goal is to reach a point where the tools create more productivity than they take away [1].
Torvalds and other maintainers have been working to finalize an official AI policy for Linux kernel developers to manage these submissions [2]. The policy aims to balance the potential for rapid iteration with the necessity of maintaining a rigorous, human-verified codebase.
“"That was pulled out of my ass number, obviously."”
The admission highlights a growing gap between the marketing promises of AI coding assistants and the reality of maintaining production-grade software. While LLMs can generate code quickly, the 'cost' of AI is shifted from the writer to the reviewer. For the Linux kernel, where security and stability are paramount, the high volume of low-quality AI patches represents a net loss in productivity unless the tools can significantly improve their accuracy.
