Linus Torvalds and Dirk Hohndel on AI Coding, Bugs and Hype

I want to examine two interviews about AI and coding featuring Linus Torvalds, creator of the Linux kernel and Git. He remains deeply involved in Linux development. The person interviewing him is Dirk Hohndel; distinguishing their voices matters because they do not express identical views.

In the first conversation, Hohndel describes language models as a kind of enhanced autocorrect and questions whether next-token prediction is intelligent. That is Hohndel’s characterization, not Torvalds’ own opening statement. Berman notes the debate over what language models can do beyond prediction; the analogy is a viewpoint, not a settled measure of intelligence.

Asked whether AI-written code will be submitted for Linux, Torvalds says he expects it and suspects some assistance is already happening. He sees automation as a continuation of programming history, from lower-level code and assembly toward languages and tools that make development easier. Berman notes the growing use of AI-assisted editors in 2024.

Torvalds is cautious about describing the change as revolutionary simply because it dominates the news. Berman interprets the transition through layers of abstraction: higher-level programming tools translate human intentions into instructions a machine can execute. Assembly and low-level code remain important in some domains, even if many developers do not write them routinely.

I think natural language may become another layer in that progression: a person describes the desired behavior and an AI system generates code. I also speculate that future models could use representations unfamiliar to humans. These are forecasts, not established outcomes, and imprecise requirements can still create incorrect software.

Hohndel asks whether models might eventually help with review and maintenance. Torvalds hopes they can catch the simple mistakes that developers overlook. They might also notice when code deviates from an established pattern and ask whether the difference was intentional.

Berman recounts a Microsoft developer’s opinion that AI code review was not broadly useful in his experience, but sometimes caught obvious errors such as misspellings or off-by-one mistakes. That anecdote does not establish performance across teams or tools. Subtle logic and business requirements can be harder to check, which is why human review remains valuable.

Compilers and linters already flag many syntactic and rule-based issues. The proposed opportunity is to supplement them with a model that spots less obvious inconsistencies and asks a developer to verify intent. Such suggestions can themselves be wrong, so a warning is a prompt for review, not a substitute for it.

Torvalds pushes back on Hohndel’s dismissive autocorrect analogy, saying these tools may be more than that and could help people work better. Berman compares the idea with human pattern recognition, while recognizing it as an analogy rather than proof that humans and language models reason in the same way.

The pair joke about optimism and humility. Torvalds observes that it took optimism to begin building a kernel decades earlier; Hohndel appreciates that optimism but says he does not necessarily share it. Their exchange is a useful reminder that openness to a tool and caution about its risks can coexist.

Hohndel worries that language models make things up, and that allowing them to act without a human catching errors could produce bugs affecting people and code. Torvalds responds that humans already produce plenty of bugs. Berman agrees that both sources of errors call for testing, review and safeguards, especially in mission-critical systems.

In the later interview, Hohndel and Torvalds turn to the hype surrounding AI, including companies adding AI to their pitches. Torvalds finds the marketing excessive but says the technology is interesting. He recalls earlier forms of AI, including expert systems, and resists confident predictions that jobs will disappear on a fixed timetable.

Torvalds points to a concrete benefit he sees: NVIDIA has become more involved in Linux memory-management work as GPU-driven computing has grown. He suggests waiting to see what AI becomes rather than accepting sweeping claims about the future. The remarks describe his perspective at the time, not a comprehensive assessment of NVIDIA’s contributions.

Berman disagrees about the pace of change. I speculate that many application developers may eventually be needed less because people could ask AI directly for an outcome, such as a reminder, rather than opening a to-do app. This is my prediction, not Torvalds’ claim; the example still relies on software and services to store data and deliver the reminder.

Hohndel says AI tools are already producing lasting changes in everyday work and again uses his autocorrect analogy. Torvalds says he is optimistic about tools that find bugs. He compares smarter tooling with the move from writing assembly to using compilers, while rejecting both catastrophic and utopian certainties.

Torvalds says technology hype cycles often contain a real idea beneath exaggerated claims, mentioning past waves such as crypto and cloud. Berman pushes back, noting that cloud infrastructure and substantial businesses are real. A technology can have important uses while particular marketing or predictions about it remain inflated.

Hohndel points to a long line of development aids, including Sparse for finding certain problems and Coccinelle for code transformation. Torvalds says existing kernel-rewriting tools can be hard to use because their rules must be specified in detail; AI might help make such workflows easier. Berman sees pattern-based model assistance as a complement to explicit scripts.

Hohndel then raises the importance of open data for training. Torvalds says that question is not his personal focus, while acknowledging that others in the open-source community care deeply about it. Berman argues that accessible data deserves attention alongside open algorithms; his claim that all open data has already been used should not be treated as established fact.