From reddit:
Saw this clip from Linus Torvalds at a recent Linux Foundation talk and it’s been stuck in my head.
He’s not anti-AI. He actually says AI is a great tool and that he uses it himself, including pushing people toward local/self-hosted models instead of relying on big AI companies long-term. So this isn’t a “old man yells at cloud” take.
But the part that’s actually generating discussion is this comparison:
- AI increases your productivity by roughly 10x, in his estimate.
- Compilers increased productivity by roughly 1000x.
So by his own framing, the thing we’re currently calling “revolutionary” isn’t even close to the biggest leap programming has already been through. We’ve just forgotten because most of us never wrote machine code by hand.
Then he goes after the “99% of my code is written by AI” crowd directly says it makes him genuinely angry, because those same people never say “100% of my code is written by a compiler,” even though it’s the exact same outsourcing of execution. The difference, in his view, isn’t the tool it’s whether the person understands what’s actually happening underneath.
His actual warning on vibe coding:
good developers write good code,
bad developers write code that breaks,
and exactly the same will be true of whoever’s prompting the AI.
People who understand systems will prompt well.
People who don’t, will ship broken systems faster than ever.
He draws a line between throwaway “vibe coded” projects (totally fine, that’s what AI is great for) versus anything you intend to actually maintain for years where you have to understand the generated output, not just the prompt, or you won’t be able to maintain it long-term.
He also flagged a less talked-about side effect:
open source maintainers (not just kernel-level, but the thousands of smaller solo/3-person projects) are getting hit with floods of low-effort AI-generated bug reports where the reporter vanishes the moment you ask a clarifying question. That’s actual burnout, happening right now, not theoretical.
Genuinely curious what people here think is “vibe coding” actually a legitimate skill, or is it just a new name for not understanding your own systems? And is the AI-vs-compiler productivity comparison fair, or is he underselling what AI is actually capable of long-term?
Clip here if anyone wants the full context: https://youtu.be/3NSSGt9bZag
Personal notes:
I think that’s a very nuanced/balanced take by Torvalds, this comment from that youtube video linked in the reddit post really stuck out to me:
I think everyone agrees that programmers need to understand all of the code. But what people don’t talk about as much is how to achieve that. In my 30 years of professional programming, I have learned the only way to have that understanding and to have the skill to read code is to write code constantly. I don’t think AI should be used for anything but prototyping and research.
I myself have often used LLMs for seeing if a problem is “solvable” or not, or if something is “possible” or not - that’s my main use for LLM assisted coding if I ever do it. For ex. I recently vibe coded this bash script for my Nanobot instance because it was a throwaway project and I couldn’t care less about the actual script for this particular project. I just wanted to see if it could be done and wanted that to be done as soon as possible - that’s where LLMs come in for me.
From Jonathan Blow:
Most of the good programmers I know don’t really use AI all that much, certainly not for programming. (…) This is the biggest hype cycle in the entire history of computers. (…) Computers can play chess better than humans, but humans still enjoy it. Don’t buy into the hype, don’t freak out. (…) Good decisions (like deciding whether you want to change careers) are not based on the world’s biggest hype cycle.
He also mentions that AI is good for searching (way better than Google search he says) - but only as long as you verify the results.
Also interesting was Geohot’s opinion on LLMs and hype. In summary, he too agrees that it’s mostly hype. That’s where I got the Linus Torvalds quote from!
AI is something that’s happening mostly due to Moore’s law and general progress in computing, not something that they are doing. Of course they have a strong incentive against you finding this out, because then you might not want to give them billions of dollars.
All in all, most cool programmers agree that LLMs are good for menial tasks (ex. Geohot says, “But models are useful just like find replace, stack overflow, or all the regexes I never learned how to write and now never will!”) but for serious programming work, writing code that you understand, or thinking in terms of systems is way more valuable than being able to navigate some new AI tool. See the value of writing code by hand in Nobody Cracks Open a Programming Book Anymore.
This, by the way, is also the reason I left Twitter and don’t want to go back to the platform again. It was full of people my age or older flaunting their newest AI assisted application or some LLM related tool - it just felt like a giant echo chamber at one point where everybody was saying the same thing. The same words, “Claude”, “GPT”, etc. echoed in my feed every time I opened the app. So one day I just quit and never went back again, it felt pointless and the signal to noise ratio was very low on Twitter. I could fine tune my feed to match exactly what I wanted, but you can’t escape the echo chamber while choosing to be in the echo chamber at the same time. Anyways, I am content with my agent assisted newsletter about the latest AI developments or even hcakernews for that matter - both feel better than twitter.