KubeCon China - at 33-and-a-third, Linux is a long player. So, why does Linus Torvalds hate AI?

Friday morning, Hong Kong time, a packed hall of maintainers and developers did something unusual in Chinese culture - they whooped and roared their approval at the sight of an urbane Finnish American onstage.

Thus, Hong Kong welcomed the father of Linux, Linus Torvalds, to the KubeCon and CloudNativeCon stage – though he acknowledged he was here for the third-billed conference in this package event: the Open Source Summit. “I didn’t know that,” he joked when welcomed to KubeCon specifically – his first visit to Hong Kong, though not his first to China.

Now in his mid-50s, Torvalds has spent well over half of his life on open-source operating system Linux, which is one-third of a century old this year – plus a fair chunk of it on version control system Git, which he also launched.

When asked how many of the audience were younger than Linux, roughly one-third put their hands up: so, to them, Torvalds must be a legendary figure: a prime mover of the industry they love. (So much so that China is now the world’s number two open-source contributor, as my previous report explained. He said:

This Sunday, it will be 33 years since I sent the first email [about Linux]. It wasn't the release itself, which was a little later, but it was the first email where I said, ‘Hey, I've been working on this thing since April, and it's almost ready now.’ And here we are literally a third of a century later – 33 years plus four months – and it's still almost ready! We'll see if we ever get there.

Indeed, the Linux kernel is currently at release 6.11 RC4. Torvalds said:

What’s interesting to me is we’ve been doing this same [iterative 9-10 week] release process now for almost 20 years. And you'd think that all the basics would have been fixed long ago! But that's not actually true: a lot of the discussions we have going on right now, both privately and on the main list, are still about core kernel operations.

For example, we have long discussions about memory management, which is not like some new hardware thing. I mean, new hardware is where a lot of the lines of development go, and that's why drivers tend to be half of every new kernel release. But I find it interesting that we're still discussing really core issues that I would have thought would have been solved ages ago.

New behavior patterns end up meaning that we still need to tweak these core things. But I like that. I am not complaining at all. This is why I do operating systems.

Candor

In a world of Big Tech pizazz and endless talk of ‘innovation’ – which is generally designed to give share prices a boost – Torvald’s candor and pragmatism are refreshing: a culture in which ‘move collaboratively and fix things’ trumps moving fast and breaking them. Has ‘the vision thing’ ever been important to him? He said:

I don't really have this policy; it's the details that really matter. We have big-picture ideas of where we want to go in the long range, but most of the real development is about getting all the details right. You don't look five years ahead for that. You look one or two releases ahead at most.

Later this year, we will have the 20th anniversary of the Real-Time Linux project […] and the people involved are finally at the point where they feel like it is done! They are still tweaking the last things, so that they will be completely merged into the upstream part.

So, when people think that we do kernel development very quickly and releases every two months, it's true, but that’s partly because there's all this parallel development going on. Some features happen quickly, in a couple of months, but a lot of features have years, and in some cases decades, of work behind them before they get fully merged.

The current policy of rolling updates and releases arose out of a process that was originally much more chaotic, he explained: 

Good code is good code, but I keep very strict rules on release management. So, if your code is not ready by the time the merge window opens, I will not take it, because I don't want the kind of pain that we had 20-plus years ago. So, today we have a very reliable release process, even if the individual features may not always be released when we would like them to be.

He added:

Bugs will happen, and any bug can be a security bug, if somebody is clever enough to figure out how to misuse it. So, being a kernel, security is obviously one of the most critical things we must keep in mind. But at the same time, we need to move forward. We need to support new hardware. We need to support all these new ways of doing development or using hardware, and that means that we can't make security the only priority.

So, what of the growing importance of programming language Rust in the Linux kernel – something about which there has been intense debate online? Torvalds said:

The very slowly increased footprint of Rust has been a bit frustrating. I was expecting uptake to be faster, but part of it – a large part of it, admittedly – has been a lot of old-time kernel developers are so used to C and really don't know Rust, so they're not excited about having to learn a whole new language that is, in some respects, fairly different. So, there's been some pushback for that reason.

Another reason has been the Rust infrastructure itself has not been super stable. So, in the last release I made, we finally got to the point where the Rust compiler that we can use for the kernel is the standard upstream Rust compiler, so we don't need to have extra version checks and things like that.

I'm hoping that we're over some of the initial problems, but it has taken us one or two years and we're not there yet.

Interest

However, with KubeCon, CloudNativeCon, AI_dev, and the Open Source Summit being presented as a single, integrated event in Hong Kong – with different content streams and linking keynotes – Torvalds was clear that his own focus is very specific: he is just not interested in most of it, he said.

Indeed, he travelled thousands of miles to make his point:

One of the things that makes me enjoy open source so much is obviously just the community. Being involved and having lots of different people to communicate with, that's what really keeps me going.  But the thing that makes it all practical is that people specialize in what they're interested in. So, when I say no, it's because it's just not my interest area. So, when it comes to things like Cloud, Kubernetes, or AI, to pick the hot topic of the day, I see myself as a kernel person who wants to support that, but I don't see myself as a person who is at all interested in the end result!

So, when AI people came in, that was wonderful, because it meant somebody at NVIDIA had got much more involved on the kernel side, and NVIDIA went from being on my list of companies who are not good to my list of people who are doing really good work. But that doesn't mean that I, personally, end up being interested in AI. I am just interested in what we need to do in the kernel to support AI. I still see myself as a core kernel person. But I think it's a good thing that people specialize. So, if people ask me about Linux using the cloud, I'm like, ‘I know Linux. But I don't know cloud.’

But might AI have a role to play in the development of Linux itself? Torvalds said:

I have been talking to people who are looking at this, at making AI understand kernel code, and I am hopeful. I mean, I still don't like AI. Not in the sense that it's this horrible hype thing that everybody talks about. But it keeps on coming, and at some point, you just want to tune it out! But at the same time, I'm hoping that in five years or maybe sooner, we will be in the situation where we take AI more for granted, and we actually have these everyday tools that aren't just writing.

I realize that you can use AI today to write JavaScript or Python and things like that, but we're not at the point where AI is yet helping us find bad patterns in the kernel source code. But there are people working on that, and I'm actually optimistic about it. I'm not so much interested in AI writing code. I'm much more interested in it finding bugs proactively and doing code review, helping maintainers and developers run better code. And I think we will get there, but we're not there yet.

Could he be more specific about the role that AI might play in the kernel? Torvalds said:

What I've been hoping for – and I've been talking to a couple of people from big companies that I won't name – is not the traditional LLM that just predicts what you're doing, but something that hopefully takes the kernel source code history and other projects into account and learns what good code patterns are, and red flags things that it says look suspicious. However, right now, most of the tools are at the ‘fairly obvious’ stage, and I think AI can do better. But it's probably not commercially the number one priority for AI companies. So, we'll just have to see what happens.

Then he added:

[With AI it’s] what's the big vision? And things like that. But I've never had a big vision. And I don't want to have them! I associate visions with drugs and mental issues. I see myself as a plugging engineer, and I'm proud of that. I don't like that much vision for the future of open source!

My take

In this day and age where hysterical levels of hype cloud business decisions based on precious little evidence, we should treasure figures who are all about doing a single job brilliantly, collaboratively, and predictably.