In September, AI generated code has made up 17.25% of all Linux Kernel patches

If that makes you feel better, the guy who vibecoded an app in one sitting will probably have 0 clue on how the code works, the app itself will look like the average claude eyesore and the docs will be unreadable by humans without stimulants

All you need to us actually look around, has software quality dramatically increased over the last year? Have useful features been coming out at an unprecedented rate?

Because after all this talk about being left behind and I see nothing to show for it.

AI-assisted coding helps me a lot to churn through boilerplate, straightforward implementations, test writing, exploration of unfamiliar codebases, and other ancillary tasks in coding. It does not replace tacit knowledge/domain knowledge required to design a good solution, no matter how much of the architecture is described as documentation it cannot predict the future to have judgement about a proposed design.

In general AI is an assistance, it doesn't replace the intellectual work that is needed for a good implementation.

In my opinion, approach it like you'd approach any other tool: there are good use-cases for it while being definitely not the right tool for other cases.

The code itself indeed has become cheap but the question still is out whether it solves the right problem. I've shifted my work significantly toward designing more high-level portions where multiple things would interact in somewhat unpredictable manner (read as failure modes) that's where I found the models being quite poor, they do suggest many things to solve these or that problems but in a unity these suggestions miss second order effects. I was quite surprised that models can't hold all of the connected pieces simultaneously to reason about. But once I have a complete design after a week of deliberations and various attempts, the models can produce lots of correct code to cover that. And then another pass starts because they still deviate in implementation of various pieces and once you start questioning those the models start self-correcting. Judging the output is still way faster than writing the code yourself and indeed models still show some cases that slipped once mind. It's a definitive increase in terms of quality and value of the end code no doubt. Whether it can one-shot a complex problem, not really.

Please don’t do this if your livelihood depends on it. You’re just going to make your life more painful.

As a professional programmer, you’re selling a service. The people buying your service have fully bought into AIs and accepted their shortcomings. You can’t fight this.

You can still scratch your itch with hobby apps, or contributing to human-first projects like Zig.

The reality is that you will be left behind, you need to decide if thats a bad thing or not. If you want software development to be your career it probably is, if you enjoy it as a hobby than probably not. You're the 2020s version of the 1990s developer sticking with ASM.

I think if you review the code and keep a good mental model of the codebase, the costs are pretty minimal but the value is huge. I'm not going back...

Code is merely the means to an end. It does not matter who writes it as long as it does what you want. Correctly.

Stands to reason though that someone who knows programming AND has domain knowledge can get LLMs to produce much better output compared to someone who does not.

I never used to have time to make all the stuff I needed or was interested in. With LLMs, I can.

This "shipped a whole app while being in the gym" does not work for me though. It takes me a couple of days to a week to produce solid, functional software (~10KLOC). Simple tools (3-400LOC)? Yeah, those you can produce in 30-60 minutes.

As is with any engineering… a key skill is knowing which tool to use when and for what.

There will always be people who push the extreme limits of either way. “I will not use AI for anything!” Vs “I will vibe code this while I’m at the gym”.

The winner is the one that knows when and how to use AI as a _tool_ and focus on the better outcome.

Building a business on vibe coded software feels like building your house on sand.

Sooner or later you'll end up troubleshooting something you just don't understand.

And the tricky part about AI is that it only has to make one bad decision confidently, to then base all future decisions on that one bad one.

I code because the software i need is not there yet and these tools still fail at doing those.

Not sure why the next xy (already done) app has any economic value outside being a cheaper clone.

Still there is nothing wrong with having an AI point out problems which i assume is the case here and someone still is responsible and knowledgeable about the patched (!) code.

The same feeling, do AI knows the hardware? Does it have the context of the changes? Did it read Linus decisions why something should be the way he thinks is correct? I would prefer such approach where AI knows so much context as Linus and the most important persons in the kernel team

400 locs is a very exhausting day for me, I think I rarely did so much work a day in my entire 15 years career

Why would you code without the help of AI, genuinely asking? You can have quite strict "guardrails" that would maintain code consistency/quality that can be close to guaranteed nowadays.

It's not only about the pace, but also about the quality, most softwares have proven that not getting AI-reviewed is risky in its own, at least in the field of security.

> But what is the cost?

The cost is similar to doing any other task in a rushed, careless manner, with little regard for the consequences.

> sometimes I feel like I am being left behind.

And why does that bother you? “Forward” does not equal “good” or “desirable”. If a van is driving towards a cliff at high speed, would you not rather be left behind? “Being left behind” is a talking point to cause you FOMO, don’t fall for it.

You are left behind. That's like saying: "I am using punch cards for programming and I feel left behind as others are using real-time IDEs" You don't need to vibe code (and even this is only a matter of time as llms become better and better), but not using AI at all will definitely leave you behind.

in your version of the conspiracy theory, the developers are pressured to use AI for their patches in Linux from the company. not that they want to use it because they want to? Like everyone else?

My hypothesis:

Linux kernel is a huge project - with a lot of people volunteering. You’d be burning through a crazy amount of input tokens to gain enough context to get any work done. I doubt many volunteers pay for their own top-tier subscription plans that would be able to do so cost-effectively.

Embedded code and systems code tends to have a lot of extra constraints the model doesn't know about and aren't necessarily written down where it can see. This produces a higher rate of "superficially plausible but doesn't actually work" code.

(There are probably wider benchmarks, but this was a definite finding in our company wide evaluation of effectiveness. LLMs aren't great at RTL either)

I would assume that linux devs are relatively older. Furthermore they are probably very knowledgeable in that area, and one thing I've noted on here is people who are knowledgeable not liking the quality of code output from LLMs, if they have stringent quality requirements it might be easier for them to write the code themselves.

Small changes can have a huge impact.. far beyond your context window ;-)

How big is your circle? What industries does it span? Maybe the people writing 0% code manually are outliers and your circle is relatively small?

The Linux kernel is essentially the nightmare scenario for a LLM.

No tests, not much documentation to work with, a very high acceptability bar, multiple ways to do the smallest change and every single word on your commit messages is very important.

You can start with a LLM but there going to be some manual modifications you will have to make, even with the best model available

Just thinking loudly here, I guess Co-Author in commits is a hint. PRs where LLMs have been involved is probably another measurement. Github Copoilot has its own account too.

They require disclosure of what tool was used if submitting AI-helped code.

(Which probably means this is slightly underreported from people not disclosing.)