Linux kernel community discussion on ML/LLM tools in kernel development

The thing is that I'm not arguing for more people to use it or that I'm a better programmer or heck - even better at corralling LLMs than everyone.

The thing is that you are, inadvertently, arguing that. Everyone else experiences hallucinations (that are, at this point, considered generally considered to be a fundamental aspect of LLMs) that they don't catch and crop up at runtime (or that take hours to resolve), or an LLM "lying", or producing behaviour that changes across versions and causes critical issues, except you, who is somehow able to produce code without difficult to spot bugs that violate your assumptions, and able to review said validity of the code so fast that it doesn't slow you down. This is despite LLM coding being majority code review (an especially difficult one, given that the LLMs do not have human-level internal models of how the code works and struggle with basic logic), and code review being almost universally considered as being longer and more annoying than just programming from scratch (I had an ACM link and workplace SE ready, but I'm sure you can ask an LLM to hallucinate up those links :) ). Furthermore, you have no testing to back up your claims, so there is no way to differentiate the actual fact versus your experience — which does (in this case) matter, given how LLMs have been shown to make people think they're faster, while actually providing the opposite, reduced quality of output to boot.

Rather what I'm trying to say is that its use regardless of what people are seeing from various failures or bad outcomes is rising and will continue for the foreseeable future, good or bad.

But these things do matter, the experiences of people using LLMs are not just externalities to be pushed aside into a flippant comment — furthermore, LLM use really isn't increasing, and this is trivially provable by pointing at how the winds have shifted in the last half a year. I do genuinely believe that Microsoft scaling down their branch of LLMs after they've seen very little use is only the tip of the iceburg of things to come, at this point.

Use it or don't, but when you interact with the code it outputs that's been put in front of you from other programmers you won't always know. Code is code.

The reason most FLOSS projects are moving to ban LLMs is not because they are good at the tasks that they do, it is because they are poor quality, a waste of maintainer time, and the developer involved cannot reason about the code because they do not have a mental model of the code to explain. The bulk of time invested in "good code production' for developers is almost never "writing code", the externalities here matter as well — code itself is communication, between yourself, the computer, the end user, and any developers that

A point I dropped but is worth paying attention to, is

Outputting slop or good code is a developer choice, we just have another tool that can make either easier

I do not believe this to be the case, a tool that constantly shifts in form and function and provides unreliable outputs depending on dice rolls is not a tool, it is a liability. If you had to make a dice roll every time you used a hammer, that dice roll paid no attention to skill or experience, and you missed over half of the hits, why would you use such a hammer?

The problem, as that thread points out, is the amount of "false problems flagged"/"correct problem mis-solved falsely" vs "genuine problems solved correctly". As things stand right now, LLMs create enough false positives that human beings can't possibly triage them all to get to the valid ones--this is effectively a slow motion DDoS.

If you use AI correctly, nobody will be able to tell you used AI--all they will see is a genuine problem solved correctly. If people can tell you used AI, well, that means you either solved a false problem or created a shitty solution to a real problem and your reputation should get penalized appropriately. The only reason I can see for opposition to this is if you want to advertise that you solved stuff with AI which would imply that you are somehow commercially involved with AI and you should probably be banned from the project since your incentives are heavily misaligned.

Do I use modern LLMs for search? Sadly, yes. The LLM companies have access to and been trained on proprietary corpuses(corpii?) to which I, as a mere plebeian individual, have no access to. Consequently, LLMs often do a better job of coughing up sample code that normal search engines have no ability to unearth. This pains me, but refusing to use a much larger search engine than my peers would place me at a disadvantage.