Torvalds Backtracks: AI Won't Multiply Productivity by 10

What happened?

Linus Torvalds, creator of Linux and a key figure in free software, has publicly retracted his 2024 prediction about the impact of artificial intelligence on programmer productivity. During the Open Source Summit India 2026, Torvalds stated that his claim that large language models (LLMs) could increase productivity tenfold was “unscientific” and, in his own words, “a number pulled out of my ass, obviously”.

According to Slashdot citing ZDNet, Torvalds clarified his current stance: “we are at the point where we expect AI to create more productivity than it takes away, but we certainly saw more junk generated by LLMs than useful code until early this year”. He noted that AI-generated patches are often “pointless band-aids” that fix the symptom but not the root cause, leaving the bug latent.

Torvalds explained that AI-generated patches are often superficial: “they can fix the immediate problem, but the type of bug remains, and it's waiting in the hallway to hit you somewhere else.” Even now, most useful patches require more than just the LLM: “we've had to push quite hard… if you find a bug with an LLM, it's not enough to ask the LLM to make a bug report and then throw it over the fence to us. We want to see a suggested patch; we want to see the human who ran the LLM act as an intermediary.”

Why is this important?

Torvalds is no ordinary analyst: his judgment on software development is followed by millions of developers and companies that adopt Linux. His public retraction casts doubt on the extreme productivity promises sold by companies like GitHub Copilot, Cursor, or Codeium. Additionally, it warns about the increase of “noise” in repositories: automatic patches that overwhelm maintainers and consume human time to verify their validity.

The warning comes at a time when many generative AI code startups base their value proposition on accelerating development. If Torvalds himself says AI generates more junk than useful code, the sales pitch weakens. This echoes other tech bubbles, like CASE tools in the 1980s or the promise of automatic programming with COBOL, which also generated exaggerated expectations that later moderated.

The historical context is relevant: in 2024, Torvalds had stated that LLMs could multiply programmer productivity by ten, a statement many took as a strong endorsement of generative AI in development. Now, two years later, his view has drastically moderated, which could influence enterprise adoption and investments in the sector.

What consequences will it have?

In the short term, development teams are likely to be more cautious when integrating LLMs into their workflows, especially in critical projects like the Linux kernel. The requirement for a human to act as an intermediary (“back-and-forth”) between the model and the repository will become a best practice. This could slow the adoption of tools that promise full automation, but also reduce the maintenance burden caused by defective patches.

In the long term, the industry could pivot toward AI tools more specialized in bug detection (where Torvalds acknowledges “absolutely amazing” advances) and less in autonomous code generation. Torvalds noted that some bugs found by AI have been “absolutely, amazingly, I mean, interesting in a painful way,” especially security issues that “appear in the tech press two days later.” Despite the embarrassment, he said: “I'm not a person who kills the messenger. I think we are much better off with LLMs finding bugs, even when they are embarrassing and are things we probably should have found two decades ago.”

It could also slow investment in generic LLMs for programming and redirect it toward models specifically trained for safe patching. Companies like GitHub, which have bet heavily on Copilot, may be pressured to demonstrate real value beyond code generation. In the market, startups like Cursor and Codeium will also have to adjust their pitch, moving from “10x acceleration” to “intelligent assistance with human oversight.”

Torvalds also mentioned that for his personal projects, he uses LLMs as rapid prototypers, but insists the generated code is not directly usable. “I use them as a way to prototype things… often the code is not usable in that form, but it's a great way to test something,” he said. This distinction between prototyping and production will be key to avoiding disappointment and for developers to adopt a more realistic approach.

What should readers know?

• Don't blindly trust LLMs for patching critical software. As Torvalds says, automatic patches are often superficial and can hide deeper problems. Experience in the Linux kernel shows that human review remains indispensable.
• Human review remains indispensable. Even when AI finds bugs, a developer who understands the context and proposes a robust solution is needed. Torvalds demands that the human act as an intermediary between the LLM and the repository.
• AI is useful for prototyping and detecting vulnerabilities. Torvalds praises the ability of LLMs to find security flaws, even if sometimes embarrassing. In his own words, “we are much better off with LLMs finding bugs, even when they are embarrassing.”
• Productivity predictions from AI vendors should be taken with skepticism. If the creator of Linux retracts, any similar claim should be critically examined. History shows that promises of extreme productivity rarely come true without adjustments.
• The future of AI in development is not mass generation, but intelligent assistance. The trend points to models that collaborate with humans, not replace them. Torvalds sums it up: “there are many useful and less useful uses for AI.”

“There are many useful and less useful uses for AI. I think Godzilla is a great place to stop,” Torvalds joked, making clear that hype has limits. On his travels, he even uses AI to add Godzilla to photos he sends to his children, showing a playful side of technology.

In summary, Torvalds' retraction is a call to reality in a sector dominated by hype. Developers and companies will need to balance enthusiasm for AI with a critical evaluation of its actual capabilities, especially in critical software environments. Productivity is not measured only in lines of code generated, but in the quality and maintainability of the final result.