Linus Torvalds shares his thoughts on Anti-AI coding rules on Linux mailing list.

[interpipes@thx.gg]

@nixCraft seems there are predominantly two types of people who love going all in on LLMs

* people who can't do a thing, who see them as somehow democratising a skill they haven't the time, energy or patience to learn (and so cannot meaningfully review the output)

* people who (now) hate the work they do (and so will not meaningfully review the output)

[garnetkoebel@hachyderm.io]

@nixCraft Isn't Linus famously against debuggers? Seems like a double standard to embrace a word guessing machine when he's been stuck in the past for decades.

[ajayiyer@mastodon.social]

@nixCraft I think it is scary when someone at that level makes false comparisons between a compiler and an LLM.
He's loves his "I'm just an engineer; I just build tools" philosophy. Surely he should know that not all tools are the same.

Both "asbestos" and "mineral-wool" stop fire, but one is EXCEPTIONALLY more dangerous to humans than the other.

[aakl@infosec.exchange]

@nixCraft He is being quite clear that the opposition isn't "intelligent."

[autiomaa@mementomori.social]

@chessert @nixCraft Yeah, have seen many others also starting to look at BSD systems, as more and more problems are piling up on the Linux side. Current monoculture of Linux (Linus?) is making OSS more unstable.

[derderwish@social.chaotikum.org]

@nixCraft At least, the editor is not destroying the Internet and the environment while stealing intellectual property from small creators…

[prefec2@norden.social]

@nixCraft AI code assistance is best for mondaine tasks, like boiler plate code, solutions used by many programmers, like the stuff you can find on stack exchange and friends. It is less useful for very mathematical issues. At least this is my impression. Most likely, Claude could create my Python code faster than I do, but I usually don't know what I want to build.

I would love to see studies on how well these tools perform with code close to machine hardware.

[epistomai@mastodon.social]

@nixCraft Idk how Linus uses AI. I hate AI, but sometimes is kinda useful ONLY IF it's used locally hosted, like with Ollama.

I'm more concerned about environmental impact than stealing knowledge, although both are major concerns

[maybe_mbk@mastodon.social]

@nixCraft

When I look at the general situation of things, I get the urge to take cover and move to the fringe of everything. Mostly by relying less on and using less technology. Feeding the machine and the system as little as possible. Preparing for bad times. We are all witnessing a massive and widespread failure of elites on all levels.

(And I actually get Linus' point, the overt and the subtext, but it's centrist talk.)

[__jm@mastodon.social]

@epistomai there are better tools than ollama for local AI. https://sleepingrobots.com/dreams/stop-using-ollama/ explains why.

[maybe_mbk@mastodon.social]

@autiomaa @chessert @nixCraft

The point when the benevolent dictator took a wrong turn and/or lost his marbles?

[einemeh@mastodon.social]

@epistomai @nixCraft

It is not the locality that is of primary importance (though it is certainly important, just not primarily so); what is important is that the model is licensed under a free (libre) license.

- my opinion

[iainwilliammcclain@mastodon.social]

@nixCraft I use AI to scan my code and to compile apps. Does that mean I don't know the code? Nope. I just know it takes AI 30 seconds what takes me three days.

[labspokane@mas.to]

@nixCraft Does Linus have a girlfriend named, “Claudia” perchance?

[easyas314159@hachyderm.io]

@nixCraft I can’t find the post but someone pointed out recently that LLMs are not like compilers. With a compiler you generate a disposable artifact that you can reliably recreate from the input source. With LLMs you discard the input prompt and keep the output, and running the same prompt twice will generate different results. This is an infuriating stance given how much of the last decade I’ve dedicated to making reliable deterministic build pipelines.

[stefan_s_from_h@mastodon.social]

@nixCraft there's hope that Claude suggests rewriting the kernel in C++.

[biglinter@mastodon.social]

@nixCraft "nobody in their right mind"... what happened to the ring of maintainers?

well, then itll take the time it takes to put in the code... seriously, just because youre flooded doesnt always mean youll have to open the flood gates...

i remember linus saying that someday linux will become obsolete, so we are steering right into that i guess

[freevolt24@mastodon.social]

@nixCraft Two things. First, most of the time, discovering problems are easier than building complex systems, and the same goes for ML/AI models, of course. Second, ML/AI is "bound" to make mistakes, and ML/AI scientists have been saying for years that you need mistakes for ML/AI to be useful (because otherwise, it means overfitting). I understand that ML/AI tools can be helpful for productivity, but it's guaranteed that it'll make mistakes (if not intentional).

[freevolt24@mastodon.social]

@nixCraft What's the measurement against them? Is there any attempt to (officially) verify the kernel code with ML/AI, Syzbot style, not just for development? I assume ML/AI tools will soon plant some complex mistakes that are difficult for overloaded human reviewers to find out.

[syntaxseed@phpc.social]

@nixCraft Comparing deterministic tools to #AI is wild.

Yes, I would create rules about not using compilers & editors if they gave different output each time I used them, occasionally gave wrong output, sometimes gave SLIGHTLY wrong output, and were just running predictions based on my past inputs rather than following sets of rules.

#Linus