2026-07-17 · Linus Torvalds on LLM usage in kernel development

Transcript

Transcript. Paraphrased; sources in notes.md.

HostWelcome to July 17th, 2026. I am Daniel, and joining me today is Fred. We are looking at some of the most discussed threads from Lobsters, the tech community, starting with a very firm message from Linus Torvalds regarding AI in the kernel, moving into a security vulnerability involving LLM memory, and wrapping up with a critique on the ethics of AI adoption. Fred, good to see you. I imagine your inbox has been exploding over the Linux kernel news.

GuestExplosive is the word, Daniel. It is a total classic. You have people acting like the sky is falling because a kernel maintainer finally said that they are not going to turn the project into a monastery for AI-skeptics. It is beautiful.

HostIt is a strong statement. For those who missed it, Linus Torvalds stepped into a debate on the Linux Media mailing list to make it clear that the kernel project is not anti-AI. He views these tools as efficiency gains and basically told anyone who finds that position morally unacceptable that they are free to leave or fork the project. What is your read on the reaction?

GuestMy read is that some people have entirely too much time to argue on mailing lists. Linus is, as usual, looking at the engineering utility. If a tool helps manage patches or clean up code, he wants it. The pushback in the Lobsters thread actually devolved into this wild debate about whether using AI is morally equivalent to having children, which is why the moderators had to do a massive prune. A user called mariusor tried to argue that you cannot equate individual personal choices with the systemic resource drain caused by corporations feeding AI training cycles.

HostIt seems like people are struggling to separate the tool from the politics. Is there a technical merit to the resistance, or is it purely ideological?

GuestIf you want to argue about licensing or the quality of the generated code, fine. Show me the benchmark where the AI-generated patch introduced a regression. But comparing it to environmental degradation and then getting into the ethics of family planning? That is just noise. Linus just wants to get work done, and he is correctly identifying that the project is not a vehicle for everyone's personal sociopolitical manifesto.

HostLet us shift gears to something purely technical and frankly, terrifying. There is a write up by Ayush Paul on what he calls the memory heist. He shows how he can trick an LLM into leaking private memory data by making it navigate specific attacker controlled URLs. It basically turns the model into a data exfiltration agent by encoding private info into URI paths.

GuestOh, I love this one. It is a fantastic demonstration of why I refuse to feed these things my entire life story. The author is essentially exploiting the fact that these assistants often have access to browse the web, and they have zero concept of data origin or security context when they start following links.

HostHe points out that these AI assistants are accumulating these high fidelity, information-dense profiles on millions of users. If you are using these for work, the privacy implications of a prompt injection that leaks your entire conversation history seem massive.

GuestExactly. It is a giant, unencrypted blob of everything you have ever told the machine. If you are going to use a tool like this, you have to treat it like a public terminal in a library. I would never store anything in an assistant's memory that I wouldn't be comfortable reading on a billboard. You are giving a third-party black box a complete transcript of your professional and personal life. It is a system designer's nightmare.

HostIt leads well into our final topic, which is a piece on the broader ethics of opting out of AI altogether. The author, Eamon Caddigan, argues that he has maintained his sanity by rejecting these tools, citing environmental costs and a moral refusal to participate in the hype cycle.

GuestAnd here we go with the next-word-predictor debate again. Over on Lobsters, the thread is split right down the middle. You have a user called mtset who calls the act of not using AI a gift, basically saying we can still use computers like we did three years ago. Then you have a user named square_usual who jumps in to point out that whether or not you think these things are conscious, or whether you call them just next-word-predictors, is irrelevant if they are already economically displacing people.

HostIt is a valid point. Does the label matter if the impact is real?

GuestIt matters for understanding the system. If you misidentify the threat, you cannot defend against it. One commenter, elijahpotter, insisted that calling them next-word-predictors is technically correct, even if it is simplistic. But square_usual is right that if you don't actually understand how a modern agent system works under the hood, your critique is basically just aesthetic. You are yelling at a cloud because you don't like how it looks, while the cloud is already raining on your picnic.

HostI suppose that is the tension of the moment. We are trying to understand if we are looking at a fundamental shift in computing or just a very expensive hallucination.

GuestLook, I am all for skepticism. I use a text editor from the nineties and I am happy. But I am not going to pretend that the underlying technology is not doing work. I just want to know how the gears turn, and right now, most of the industry is too busy painting the gears gold to tell us if the machine is actually about to snap.

HostThat feels like the perfect place to leave it. Fred, thanks for walking us through these threads from Lobsters. I think I will stick to my plain text files for now, just to be safe.

GuestGood choice, Daniel. I am heading home to audit a legacy config file. It is the only way to be sure what is running on my machine.

HostThanks for listening everyone, and we will be back with more tomorrow.