Linux: Torvalds Integrates AI into the Kernel

In Prague, Linus Torvalds praised AI as a tool that brings joy back to coding and helps uncover defects, while calling for caution whenever the code involves the kernel. He adopts a pragmatic stance: integrating AI into processes without exhausting maintainers, staying true to a philosophy of incremental improvements that has proven effective for two decades.

Torvalds' Priority: Avoiding Maintainer Burnout

Torvalds reported that during the Linux Kernel Maintainer Summit, three-quarters of the discussions the previous day focused on how to make AI generation and review less stressful and more useful. According to him, the main obstacle is not obtaining fixes, but rather getting genuinely useful work through the review process without exhausting the maintainers. He noted that AI-generated fixes and bug reports can overwhelm maintainers. Despite this, automated verification is becoming a permanent fixture in the process, which has relied on a stable model for about 20 years, with release cycles now set around nine to ten weeks.

Real Gains: Security and Reviving Forgotten Drivers

Torvalds acknowledged tangible benefits: AI tools identify real security issues and highlight defects in old, underused drivers. He believes that the corrections from these tools add value while emphasizing their cost to the organization. Review systems like Sashiko publish their analyses on the Linux Kernel Mailing List, and some subsystem maintainers now expect a fix to be reviewed before acceptance. Not all pull requests undergo AI review, but some already require it, and it seems likely that this requirement will become more widespread in the future.

Red Lines: Caution on "Real" Code and Issues in Mumbai

Torvalds stressed the need for great caution whenever dealing with "real and important" code. He reminded attendees in Mumbai that convincing but fabricated bug reports can require significant effort to refute, and overly targeted fixes may only address a symptom without tackling the root cause. He described some submissions as "band-aid fixes" lacking thoughtful consideration. This caution reflects his distinction between recreational use of AI and its responsible application at the core of the Linux kernel.

Making Programming Accessible: AI, Beginners, and Personal Projects

Torvalds sees AI as a lever that makes programming more enjoyable and enables beginners to accomplish tasks that would otherwise be difficult for them. He referred to "AI as a gateway drug," mentioning a guitar pedal project whose C interface, on a microcontroller and small screen, seemed dated from the 80s, even though he does not program in Java. Having started coding around 1981, he observes that today, the rising skill level makes it harder for novices to perceive the value of their own efforts. In Prague, during a conversation with Dirk Hohndel, he reiterated his enjoyment of using AI in this spirit.

From BitKeeper to Git: A Continuous Adaptation of Practices

On the occasion of Linux's 35th anniversary, Torvalds reflected on the evolution of the project's organization. In its early days, without a real infrastructure, he tracked changes via tarballs, with daily internal patches and at least weekly releases; this mode persisted until 2002 as the community expanded. He had resisted centralized source control tools, favoring developer independence and their local environments, but the manual integration of contributions became a bottleneck. The adoption of BitKeeper facilitated the merging of subsystems like networking and ARM, despite a proprietary license that divided the community; however, Torvalds still considers it very successful and not mandatory for all maintainers. After a licensing conflict in 2005, he designed Git, with a functional first version emerging in 11 days featuring a new object model, although the early versions confused CVS users. The subsequent success still surprises him, with a 2022 survey attributing Git usage to nearly 94% of respondents and almost 97% of professional developers; he sometimes feels he receives too much credit and handed over maintenance about six months later. Git today, he emphasizes, is much better than in 2005. Meanwhile, Linux has transitioned to frequent releases with merge windows and release candidates: annual releases had begun to slip over years, leading to a proposal for a shorter cycle, initially five weeks, and stabilized between nine and ten weeks for nearly 20 years. This trajectory aligns with his preference for incremental changes rather than spectacular launches. In this environment, where part of the open-source world remains wary of AI and where players like System76 prohibit it in certain components, Torvalds advocates for measured integration, guided by the utility and robustness of processes.