Linus Torvalds On Why AI Works For Him 100
An anonymous reader quotes a report from ZDNet: Many open-source developers still dislike AI. The latest example is System76 banning AI-generated content from its COSMIC desktop. At Open Source Summit Europe, Linux creator Linus Torvalds offered a different take: "I actually really like using AI," Torvald told Dirk Hohndel, his good friend and head of Ericsson Software Technology, in their latest discussion. [...] Torvalds said, "I think that you need to be very careful using AI when you're doing something real and important." Torvalds explained he hadn't always felt that way about AI programming. "I used to think it was just not very good -- but now I'm at the point where I actually really like using AI."
"To be fair," he continued, "I don't do a lot of kernel programming. I'm a maintainer. The Linux kernel is a project where other people do the real work, and I act as the collection point. But I still enjoy programming immensely. I think AI is a wonderful tool if you use it correctly and treat it as a tool, and I find it makes programming much more enjoyable." Torvalds recalled, "I started programming in 1981 or so, and back then computers were much simpler, and you could understand what they were doing, and also the programs you compared your own programs to were much simpler."
"It's a different story today," Torvalds said. "The bar in software engineering has grown so high that it's hard to see your own small efforts as worth anything, because you're used to all these polished, professional programs. [...] A lot of people make fun of vibe coding, but I think it's a wonderful way to find joy in programming. It makes you feel like you're doing something relevant. It allows you, as a new programmer, to do things that you would otherwise have a really hard time with."
He continued: "I love the concept of AI as a gateway drug because I was working on my own toy project, the guitar pedal, and I knew what I was doing. I had a user interface that I had written myself in C that was running on a microcontroller that was running on this small screen, and I was like, "This is just stupid. It looks like something from the '80s; I wanted it to look like something from this century. But I don't do Java. I do kernels." "So, I use AI to do the things that I'm bad at," Torvalds went on, "I'd seen all the same YouTube videos you have. Give AI a prompt, and it will do it. So I gave it a prompt, and it did it, and it was horrendous. Then I said, 'Hey, look! I already did this in C. I did it the right way. It's just ugly." Torvalds acknowledged that AI tools are helping identify security vulnerabilities and improve the Linux kernel, but at the cost of overwhelming maintainers with questionable bug reports and superficial fixes. He called some submissions "mindless band-aid kind of patches."
That said, despite AI's growing influence, Torvalds believes Linux's success comes from steady, incremental development rather than sweeping changes. "I'm a big believer in the whole incremental small changes that, over time, turn into big features," he said. The development goal remains "incremental improvement and steady progress all the time."
"To be fair," he continued, "I don't do a lot of kernel programming. I'm a maintainer. The Linux kernel is a project where other people do the real work, and I act as the collection point. But I still enjoy programming immensely. I think AI is a wonderful tool if you use it correctly and treat it as a tool, and I find it makes programming much more enjoyable." Torvalds recalled, "I started programming in 1981 or so, and back then computers were much simpler, and you could understand what they were doing, and also the programs you compared your own programs to were much simpler."
"It's a different story today," Torvalds said. "The bar in software engineering has grown so high that it's hard to see your own small efforts as worth anything, because you're used to all these polished, professional programs. [...] A lot of people make fun of vibe coding, but I think it's a wonderful way to find joy in programming. It makes you feel like you're doing something relevant. It allows you, as a new programmer, to do things that you would otherwise have a really hard time with."
He continued: "I love the concept of AI as a gateway drug because I was working on my own toy project, the guitar pedal, and I knew what I was doing. I had a user interface that I had written myself in C that was running on a microcontroller that was running on this small screen, and I was like, "This is just stupid. It looks like something from the '80s; I wanted it to look like something from this century. But I don't do Java. I do kernels." "So, I use AI to do the things that I'm bad at," Torvalds went on, "I'd seen all the same YouTube videos you have. Give AI a prompt, and it will do it. So I gave it a prompt, and it did it, and it was horrendous. Then I said, 'Hey, look! I already did this in C. I did it the right way. It's just ugly." Torvalds acknowledged that AI tools are helping identify security vulnerabilities and improve the Linux kernel, but at the cost of overwhelming maintainers with questionable bug reports and superficial fixes. He called some submissions "mindless band-aid kind of patches."
That said, despite AI's growing influence, Torvalds believes Linux's success comes from steady, incremental development rather than sweeping changes. "I'm a big believer in the whole incremental small changes that, over time, turn into big features," he said. The development goal remains "incremental improvement and steady progress all the time."
I'm sure AI is great (Score:5, Interesting)
...in the hands of a competent, careful and passionate programmer like Linux Torvalds.
What I worry about is when Bob in the junior software dev team, who does his job to put food on the table, and just good enough to keep his job, uses it to vibe-code company data as his day job.
Re: I'm sure AI is great (Score:3)
Re: (Score:2)
Re: (Score:3)
Re: (Score:1)
At this point, I'm willing to trust the code Claude produces more than the code that Bob produces. It's far from perfect, but it's better than many a junior dev at this point. I suspect with the next generation of agents, it'll be as good as seniors tbh.
Re:I'm sure AI is great (Score:5, Interesting)
Not likely. Bob will just make bigger messes with AI. That's what is being repeatedly stated in these articles - Only the skilled can get results that aren't a total bodge.
And of course advocates keeps saying, after each big event when each new super super super AI is announced ... and ends as just more of the same, "It'll be better the next gen."
Re: (Score:2)
Define the difference between mess and works. AI makes messy code, but compared to some of the rubbish that low quality human coding farms put out, I'm not actually confident I can say which is worse.
Re:I'm sure AI is great (Score:5, Interesting)
Define the difference between mess and works. AI makes messy code, but compared to some of the rubbish that low quality human coding farms put out, I'm not actually confident I can say which is worse.
I agree. And it can make decent code as well.
I use an SDR radio, (Flex 6600) a high performance server with an RF front end. The company producing the radio makes a Windows only program to run it - SmartSDR. There is a version for MacOS that is similar but better, built on the APIs available. No Linux version. That's the background.
Along comes a group making an open source application called AetherSDR. Built with Claude. Uses the API from FlexRadio as well, and runs on Windows, MacOS, and Linux https://www.aethersdr.com/ [aethersdr.com]
And it works pretty darn well. There are humans in the development loop, but still some impressive stuff based on Claude AI There are bug fixes, but the other two have bug fixes as well. I haven't done the tests yet, but the claims are that AetherSDR has better noise reduction than the SmartSDR products.
Re: (Score:2)
It's far from perfect, but it's better than many a junior dev at this point.
People who talk about "junior devs" don't know much about the software industry.
Re: (Score:2)
I would interpret what he said as: It's great as a toy.
Re: (Score:2)
I'm sure every comparison can be made. Lead was a wonder building material for the Romans too. And don't forget asbestos.
Re: (Score:2)
My point exactly. You're no valid argument yourself.
Did they consider it toy really ? (Score:3)
I love slashdot for sending me off on interesting distractions.
I believe you are referring to the Hero engine, nicked by Rome from the Greeks and possibly by the Greeks from the Egyptians.
If my AI Summary is right.
It would be wrong to refer to AI as a toy as it would be to refer it to it as the bringer of an age of abundance and utopia as it is presently marketed.
Similarly it's wrong to call it intelligence, it is at best a simulcra of a sub set of features of intelligence. But even to call it that we'd ne
Re: (Score:3)
Claude is far from a toy. It does incredible work if you're reverse engineering something.
Re: (Score:3)
There is also a lot of other work that it can do really well.
I was asked to create a little one page process flow chart for some stuff my team maintains. We had all the documentation for everything as handfuls of markdown files in each code repository. The thing is not complex. I could have opened Visio and created a nice looking chart in probably 1/2 an hour to an hour.
Instead I told Claude to do it. It was done in seconds, took be me about five min to look it over and conclude there were no major error
Re: (Score:2)
Another thing that I have found Claude is really good at doing is taking an error from a log and doing a deep dive analysis of what happened. I give it read-only database access, access to www logs, access to application logs, access to a checked out source tree, etc. I paste in the error message (or in some cases paste in a tech support email from a user) and have it write up an analysis, a proposed response to the customer, and any recommendations (if there is a bug or actual problem).
Re: (Score:2)
Re: (Score:2)
Claude is far from a toy. It does incredible work if you're reverse engineering something.
Exactly. I have a SDR program written by Claude named AetherSDR. It is good, like better than all the other offerings, Like SmartSDR and SmartSDR Mac. And this program runs on Linux as well.
Re: (Score:3)
Exactly this. I have experimented with using AI, but absolutely forbid it to actually check in code. I just let it show me it's proposal. It can be useful at that, but often needs to be prodded about criteria it left out and frequently the final result needs a re-write for maintainability at least. It has a tendency to use really convoluted logic which can be greatly simplified without changing the result.
It does much better as a search tool. It's pretty good at finding exactly where some feature is impleme
Re: (Score:2)
People who are just making little apps and scripts to improve productivity could benefit from AI. I've used it to do a few little PowerShell and Python scripts. I'm not very familiar with either language and don't have time to learn, so some little automation scripts that I otherwise wouldn't have had the inclination to make myself are valuable time savers.
For example I have a PowerShell script that takes unstapled magazines that have been scanned in a sheet feed scanner and splits them into correctly numbe
Re: (Score:2)
Yes very much so. On some ways modern AI coding agents are a bit like the VB6 of our modern era. While they are much maligned by people who think they can code better than them, these agents have really empowered a lot of people with little to no coding experience to build all kinds of vertical applications for their businesses that integrate data from a variety of sources (silos really). If they have an idea for an app that would benefit their business, after some trial and error Claude is able to deliv
Re: (Score:2)
Why do you turn off the tests and checks that would apply when Bob tries to check in 100% handwritten code just because Bob is trying to check in AI-generated code?
Or is the reality that you were never testing the check in and you've just gotten lucky so far?
Re: (Score:2)
Or make that Eve, the basically incompetent attacker, that can suddenly attack 100s of badly secured web-shops for cheap.
Linus and anybody basically on his level does very careful review of the results when they matter. And he understands when they matter and when they do not. That means bad AI code will not get in and most AI code is bad. But put a coding LLM in the hands of somebody with significant less skill and you slowly create a catastrophe. It will still take a few years for that to become impossibl
Re: (Score:2)
Re: (Score:2)
He's some guy who works for Tim Apple I think.
Re: I'm sure AI is great (Score:2)
At my job, AI is used by my boss to second guess the work engineering is doing. He burns a lot of our time having us check things and explain things. Almost entirely false positives, but he sure is smug when AI does catch something.
Re: (Score:2)
Nah. The problem here is that the studies I've seen suggest skills atrophy when people become reliant on AI. So the likelihood is that Torvalds will, ultimately, deteriorate to that junior software dev you intensely dislike.\
The good news though is that if you're a C programmer, then even if you're the kind of C programmer who thinks Rust isn't necessary because only stupid C programmers make mistakes, unlike you, you're perfect (etc, you've seen the type, and would not touch a line of their code), you will
I'm not good at lovely documents (Score:1)
And that's been my experience with AI so far. If I really screw around with it I can get it to do what I wanted to do but I could have just done it myself in the first place. And since I have to check all the work it's doing it's just not worth the effort.
I've be
Re: (Score:2)
“So, I use AI to do the things that I’m bad at,” Torvalds went on,
I guess the key is to use AI to do something you're bad at. Apparently you are good at making lovely documents, so that wasn't effective.
Re: (Score:2)
The fact that I can make something barely presentable in entirely too much time doesn't make me good at it.
I still just don't understand what Torvalds is saying here. It doesn't sound like a ringing endorsement of AI. It sounds like he tried to use it and it didn't work. Maybe he just thinks it's a cool new thing and he doesn't fear it taking his job or his water and electricity since he's pr
Re: (Score:2)
I still just don't understand what Torvalds is saying here. It doesn't sound like a ringing endorsement of AI.
You are right. Maybe Linux Foundation members [linuxfoundation.org] have been pressuring him to use it for something in the kernel. They pressured him successfully in the past.
Re: (Score:2)
I guess it takes people who can't do a thing and makes them feel like they could do the thing
That's a great and concise way to put it, and it applies as much to Linus as to Bob the junior intern.
Re: (Score:2)
Well darn-it, I was hoping I could use AI for creating documents since I am in the same boat when it comes to creating documentation. Back to slugging my way through .md files.
Re: I'm not good at lovely documents (Score:2)
I have had some luck using Claude to build Typst templates (typesetting system). It is especially useful if I hate the way a document looks and I want to rip it apart and redo it. I usually just put up with it, but the AI will mindless work on a few variations that I can preview.
Re: (Score:2)
Critical thinking means you think for yourself instead of agreeing to whatever Torvalds (or anybody else) says, because Torvalds is a big name in that domain. So, AC might not necessarily prefer to blindly listen to some other slashdot commenters, but their view might happen to align. Let's not be proxies/repeaters for "important people" mouths.
The code generators work because they save work. (Score:2)
"Smart" code generators that can replicate code with some variability based on probabilities derived from their training sets are useful because they can save a lot of menial work. Code generation where you can give clear instructions and/or pseudocode and rely on the code generator to build some of the infrastructural code parts also works well. Ditto for finding bugs and adding minor features postfactum.
Since that's approximately 95% of all software today, yes, it is a huge relief. Where it fails and has
Re: (Score:2)
This only works with really competent review. Which, unfortunately, most coders cannot do. Hence what will happen in most cases is a lot of tech-debt, vulnerabilities, complexity and tech-debt slowly and quietly creeping in. That stuff becomes exponentially harder to fix and at some time, which may not be obvious in advance, all you can do is throw everything away and start over. No, that will not happen for the Linux kernel. But the kernel comes with very hard and very capable quality guards. Most software
Re: (Score:2)
This only works with really competent review.
True of any work whatsoever.
Re: (Score:2)
Yes. But in established engineering fields, that review is mandatory and if you do not do it, you may even go to prison.
Re: (Score:2)
Re: (Score:2)
Yes. If the process is half-assed. The reason for that competent review is quality assurance not only for the code but also for the coder. And a second person that has real responsibility for the quality.
Re: (Score:2)
And no one schedules enough time for a full review.
Re: (Score:2)
[AI] has a tendency to use really convoluted logic which can be greatly simplified without changing the result.
AI rewards strict APIs [dri.es]
Re: (Score:2)
"AI" doesn't understand.
So, it "rewards" whatever you provide as input, as its generation is magically tied to it.
Yeah No (Score:5, Interesting)
At my employer, like at other s/w shops, AI is being pushed hard. And we recon: let's learn it while the tokens are subsidized and the going is good.
Anyhow, the results it comes up with can be either amazing, or utterly wrongheaded. It takes experience to know which, though.
Recently, some of our non-programmers (e.g. scrum masters) have started to use AI to help with our dev efforts, e.g. fixing failing integration tests or increasing code coverage. They have no idea about code, just that it works or doesn't. Of course, there are still PRs that come our way for approval. Some of the stuff is simply hard to explain why it is sub-standard code that will be a maintenance nightmare in future (when a human needs to read and understand what's going on). When we are able to write a reasonable critique in the PR rejection, it goes through another round through the agent, which looks at the comment, spins the usual "Oh you are absolutely right" then comes up with something closer to suitable. But those cycles don't really save time, just shift it somewhere else. Sometimes it would have been quicker to just write the damn code. The LLM neither understands the bigger picture of business rules and data structure of what it is writing code on (which in many cases is living in experience in people's brains, not explicitly written somewhere, because it is deemed obvious), nor does it follow rules of thumb (e.g. Clean Code) that exist to make code more readable to humans. (Programming languages exists precisely to make code readable to humans in addition to the computer - else we could just go back to machine language.)
I guess it is good at an individual level like Linus' hobby pedal, but when a corporation is struggling financially and sees AI as some vague answer, without knowing exactly how it should help, you get a lot of running around like headless chickens that throw all sort of things at a wall to see what sticks. At the moment, quite exhausting.
Re: (Score:2)
I feel you brother. The assumption is that it will not be necessary to understand what's going on. As in "understand the code". Any problems can be just thrown back at AI and it will come up with the fix. Says my PHB.
Just like the old joke: if it doesn't solve your problems, you are not using enough of it.
Basically as we moved from pushing buttons for 0 and 1 to assembly, then to C, then to Java/Pyth
Re: (Score:2)
I feel you brother. The assumption is that it will not be necessary to understand what's going on. As in "understand the code". Any problems can be just thrown back at AI and it will come up with the fix. Says my PHB.
Yes, good luck with that, for example when it is a new vulnerability the AI cannot understand (and attackers _will_ go after those exactly for that reason). One of my students partially finances himself with bug-bounties. Quite a few intricate bugs are completely out of reach of the newest and best models and he has access to all of them. He did show me some examples. Now if you have that situation, and the code is hard or impossible to understand for humans, you are screwed.
Re:Yeah No (Score:4, Interesting)
Disclaimer: my employer is heavily into AI, so it's their way or the highway for me ;-)
I've found AI to be immensely liberating, actually. I can achieve huge things in just minutes. Then again, I'm an actual engineer, not a code monkey, so I know to put in debug or logging or whatever, and I can (at least) eyeball the code intelligently. Honestly, getting something to work isn't hard - it never really was. Getting it to work, and fail gracefully or properly is much harder, and getting it to say why it failed and what you should do about it is harder still. Nursing AI through those stages does take some work, and honestly, it's hard to know if its really quicker than "just write the code". What it definitely does do though, is lets me write a solution in a bunch of different languages without me having to stop and (re?)learn the language before I can start. That gets me across a lot more systems, so makes me more broadly capable. My team isn't necessarily any smaller, but we all do everything (rather than having that one guy that does all the java because no one else can be bothered to learn it, for example).
What AI is really good at though is 'big stuff' you don't want to do. I AI'ed up a big change, it was over something like 76 files (lots of files moving around, new files, etc). The human reviewer didn't like it, so I got Cursor to break it down into three stacked PRs. The first is just moving files around, the second is adding a load of new ones, and the third does the actual plumbing into the existing stuff. It took all of 10 minutes to break my big PR into 3 smaller ones - there's no way a human could do that.
In summary, I'd say AI is a bit of a "mixed bag". It's not all things to all tasks like the AI bro's like to pretend, it's not going to take everyone's job, but it is going to take over the way most clerical work gets done. The wonderous glory of beautiful code might well be a casualty, but in return, some people at least will be able to get across a lot of different things much easier. Maybe, just maybe they'll be a bit more productive as a whole too - we'll see.
Re: (Score:2)
Do you ask to AI agent thoroughly document the architecture, APIs, protocols, and data flow as you go? Do you use planning mode in the coding agents?
Personally I have good success with coding agents when I spent significant time with the agent building out a precise specification document that describes in detail what I want to do, how it should do it, how the data is formatted, protocols to use, etc. And then I ask it to establish concrete step-by-step development plans. I'm only working on these project
Re: (Score:2)
Non programmer writing unit tests with AI is dangerous. AI has a habit of writing unit tests that match the code, so the new tests will pass even if there are bugs in the code. I have seen AI doing this so many times that I must assume that it is normal for the AI to do so. 100% coverage, 0% value.
Despite this, AI is not useless for writing unit tests. You can ask it to find an error from the code and then ask it to write unit test that triggers that error. This way you can usually verify that the finding i
Open source has fragmented because of AI (Score:1)
Errrmm ... No.(?!?) (Score:5, Insightful)
Dude, FOSS has been fragmented since the dawn of time. It's what we lament about every day. "Why can't there be one distro?" "Why so many JS frameworks?" "Oh, yeah, yet another ORM and yet another PHP CMS, just what we needed!" etc.
FOSS certainly doesn't need AI to be the noisy mess that it is. If anything I expect AI to accelerate development for teams who want to get well-rounded user-centric stuff done and know what they're doing.
Re: (Score:2)
On the plus side, FOSS still has competition and an evolutionary process going on. That has stopped a while ago with the commercial crap.
Re: (Score:2)
Dude, FOSS has been fragmented since the dawn of time. It's what we lament about every day. "Why can't there be one distro?" "Why so many JS frameworks?"
And why can't there be one version of Windows? That's not actually the same thing I know, but a lot of us don't care. I want my computer to run my software, and not have the operating system get in teh way. And all the different distros being a source of triggers to annoyance for people is pretty silly. Like why would anyone care if I tech my students on Mint because it looks a good bit like the operating system they are leaving? But OMFG - someone is using MATE, or Zorin! Get the pitchforks and torches ou
Re: (Score:2)
Meanwhile people are silently reinstalling Windows or buying Macs due to the constant flamewars.
We don't need the weakest members of the herd anyway.
The amount of dead distros on distrowatch from protest forks that last a few months is shocking
No, what's shocking is the number of distros period. What a spectacular amount of wasted effort.
110% agree with Linus! (Score:3)
"So, I use AI to do the things that I'm bad at."
Absolutely. I totally agree and his reasoning makes total sense to me, it's the same conclusion I would draw.
Re:110% agree with Linus! (Score:5, Interesting)
Using it to do things that you're bad at means you won't get any better (fine, a lot of the time) but also you can't evaluate the output properly as you don't understand it, and of course you get impressed by observing (sometimes just apparent) competency. Using it to do things that you're good at, erodes your skills.
Re: (Score:3)
Not necessarily. I suck at writing UI code, but I certainly can evaluate and test it just fine. I also suck at writing documentation but I can certainly evaluate the AI generated documentation for accuracy and clarity. Those are two areas Claude has just really helped me.
Re: (Score:2)
I suppose I was focussing on code wrt evaluation. With UI, I disagree. You can somewhat evaluate the UI, sure, but if you don't understand the code (that's what makes one suck at it), you don't know if it does things you never intended. Your manual evaluation is a small window in the space of all UI interactions (assuming a complex enough UI). And patches like "well Claude make sure the UI does or does not do that" will quickly create mayhem in the code, as genAIs tend to be additive. And more code means it
Re: (Score:2)
I disagree. It depends on how you use it. You can either ask it to write the app for you, or you can ask it questions like "How do I write if-else using this language?" or "How do I store data into database in this environment" or "What does this code do?" "Explain this part me". AI is like a personal teacher, if you use it that way.
For things that you are good at, you can use it so that you first write or think of the solution yourself, then you ask AI to write a solution and compare them. Did the AI do so
Re: (Score:2)
I disagree. It depends on how you use it. You can either ask it to write the app for you, or you can ask it questions like "How do I write if-else using this language?" or "How do I store data into database in this environment" or "What does this code do?" "Explain this part me". AI is like a personal teacher, if you use it that way.
For simple concepts, it can probably do a good job on that front, most of the time, in well-trodden fields. For more complex concepts and less well-trodden fields, I'd be very wary. Now the problem is, that no matter its ability in doing a good job, it will always dutifully and confidently give you answer and teach you. So if you "train" yourself to trust it with simple questions and you start asking harder question and listen to it, sometimes you will be led astray without realizing. I hope Claude bot is n
Re: (Score:2)
Using it to do things that you're bad at means you won't get any better (fine, a lot of the time) but also you can't evaluate the output properly as you don't understand it, and of course you get impressed by observing (sometimes just apparent) competency. Using it to do things that you're good at, erodes your skills.
Are you the guy that programs in machine language?
Re: (Score:2)
No? ...
Re: (Score:2)
No? ...
Deep thoughts, I actually did some machine language programming. Don't recommend it, but since It is an olde school thing, I was fine to lose my skills.
It can be a little scary, but is there a point where AI programming will be pretty good, perhaps being rusty won't mean anything.
As I noted, there is an SDR program out now that is not the equivalent of standard offerings - but better. It is AetherSDR, and It is pretty darn nice - written in Claude. A human reviews it for any egregious mistakes, but ca
Re: (Score:2)
Closest I've done is a bit of assembly, and I'm jealous sometimes that I wasn't born a decade earlier to do *myself* more of everything while it happens.
I'm not one of those people who refuse to use it because it's "not good enough". It's out of principle, skill erosion avoidance, and because I don't want to contribute to the continuation of this travesty of a situation, where a few companies have the world by the balls. Wrt AetherSDR, I can Godwin the discussion and e.g. say that "shouldn't these nice med
Re: (Score:2)
Closest I've done is a bit of assembly, and I'm jealous sometimes that I wasn't born a decade earlier to do *myself* more of everything while it happens. I'm not one of those people who refuse to use it because it's "not good enough". It's out of principle, skill erosion avoidance, and because I don't want to contribute to the continuation of this travesty of a situation, where a few companies have the world by the balls. Wrt AetherSDR, I can Godwin the discussion and e.g. say that "shouldn't these nice medical advances not exist?"
It was interesting days, for myself, I started out using punchcards on an IBM360, fixing a program that if there was any error anywhere, the computer would dump everything to the printer. We knew it was happening because "The printer's singing again!" as the dot matrix printer sounded different.
Can't say I miss that all that much.
The real fun started later, when I started working with personal computers. I really got into it at that point. It was an exciting digital cowboy experience.
Now the second thi
Re: (Score:2)
Using it to do things that you're bad at means you won't get any better (fine, a lot of the time) but also you can't evaluate the output properly as you don't understand it, and of course you get impressed by observing (sometimes just apparent) competency. Using it to do things that you're good at, erodes your skills.
That doesn't hold up. If you hand over your carpentry business to a nephew, you'll still always know what a fucked up coffee table looks like. Lots of things are like that, it's easier to grade an essay than write one. Hell it's always easier to read and judge, and get good at it, than it will ever be to write something original, they're separate skills. There might even be a universal information theory rule somewhere in there. Getting rusty happens, but losing the ability to judge the quality of someone e
Re: (Score:3)
AI is great for thing that do not matter... (Score:5, Interesting)
But when you have to review and take responsibility for thousands of lines AI spews it is quite another story...
I have to deliver more and take responsibility for something that is harder and makes more time to review...
I hate my work now... AI took fun from me... in return I got much more responsibility and boring review work...
Re:AI is great for thing that do not matter... (Score:5, Insightful)
And that's the difference between it being forced on you, and it simply being made available, like it is to Torvalds.
It sounds like he's using it for hobby projects. Not the kind of assignment an employer would give you.
Re: (Score:2)
Yes, this. And your last statement is the one that makes this a catastrophe: Actually competent people will burn out, get depressed and go into different fields. And nobody will be there to do the absolutely critical review anymore. There already is a mountain of unreviewed code that cannot go into production.
Yes, LLMs are nice to create mock-ups, demonstrations and the like. They are a massive risk for doing production code unless carefully reviewed by competent humans. Anything else may wok 95% of the tim
Re: (Score:2)
But when you have to review and take responsibility for thousands of lines AI spews it is quite another story...
I would generalise this and say that AI code is useful when it is easier to validate than it is to create. AI, as it is defined today, is an approximation engine that only asymptotically approaches reliable correctness. Thus, you must always assume that the output is wrong. Either being wrong doesn't matter of you must take the time to validate the result. My area is designing the logic that goes into chips. Coding effort is dwarfed by verification and debug. Correctness is critical. Once a chip is m
But how do you use it ? (Score:2)
When trying to do something I know nothing about, it just gives me standard boilerplate examples, but I could find them by myself just reading the doc.
When learning very obscure or uncommon protocols, it is no better than me.
When dealing with a large codebase (big makefile or cmake), you can't upload the whole thing to chatgpt, so how do you tell it to refactor something ? Maybe I'm showing my age but I don't get it. Maybe it works with
Re: (Score:1)
Not here to do your tech support for you, but you need a coding *agent*. It's a process that has access to your source code, and maybe your jira. It works a bit like you do, so you tell it "here at 5 jira tickets, solve each one, raise a PR for it and assign the PR to me". Then you go home for the night. By morning, your 5 PRs will be waiting for you to review.
Alternatively, skip Jira, and just say "here's a load of code, refactor it so that the Sonarqube complexity score is below 10 for all functions". The
Re: (Score:2)
When learning very obscure or uncommon protocols, it is no better than me.
Same. And even worse: LLMs fail long before I run into trouble understanding things. They are just not worth my time. These days I just occasionally look at the "free" LLM answers with web-search to confirm it is still 30% hallucinations, 50% unhelpful crap and 10% stuff that would actually be helpful if it could give me good sources. The remaining 10% I can find in the first few genuine web-hits from stuff written by people.
My only explanation why so many people are in awe of LLM results is that they are r
Re: (Score:3)
When dealing with a large codebase (big makefile or cmake), you can't upload the whole thing to chatgpt, so how do you tell it to refactor something ? Maybe I'm showing my age but I don't get it. Maybe it works with the kernel because the entire code is publicly available, but for a private codebase ?
I'm about as far from an LLM booster as possible, and hate to be that guy, but "you're holding it wrong".
ChatGPT shouldn't be used for real software development. It's a toy chatbot. If you're doing development you should be running a harness like Claude Code or OpenCode or whatever locally on your development machine. That gives the model and tools full access to your source, including private repositories, libraries, databases, documentation, bug tracker, logs, etc. These tools are also much, much better
C-Suite: you are doing it wrong (Score:3)
The C-Suite is doing it wrong, the idea is not to require code reviews by people, let AI do the code reviews of the AI generated code. That will produce a second layer of thankless jobs and unmotivated employees sucking down salary and health care. Clearly, the answer is to go one layer further and have a bot review a bot's review of a bot's code.
You can see where this is headed, yes? Eventually you have an employee base that is unmotivated and still sucking down salary and health care, at least the ones still left. You now need to make the Quantum Leap. Put those AI review bots in a loop. With enough iterations they will approach if not actually reach a fixed point. The result will be totally inscrutable and unassailable code able to withstand the slings and arrows of customers' lawyers.
You, as a C-Suite Big Swinging Dick can spend the intervening years before the shit hits the fan accumulating wealth so that when the maelstrom reaches its zenith, you can skip out to a life of luxury and trophy wives who can suck the chrome off a trailer hitch.
Re: (Score:2)
LLMs cannot competently review LLM results. That claim is a pure hallucination.
Re: (Score:2)
LLMs cannot competently review LLM results. That claim is a pure hallucination.
You are getting more and more skilled in your trolling! Really nicely you put a false claim there, for which you can respond to all counterexamples with No true Scotsman fallacy.
Re: (Score:2)
You should read some actual research occasionally, instead of mindlessly believing marketing claims like a complete moron.
Re: (Score:2)
Re: (Score:2)
LLMs cannot competently review LLM results. That claim is a pure hallucination.
GP acknowledges this in his closing.
Re: (Score:2)
Yes, I dead that. But it cannot be stated enough.
Better article (Score:4, Informative)
He just figured that out now... (Score:3)
I woke up to AI assisted coding over 2 years ago and never looked back.
Claude is the best lab partner I've ever had.
I've forgotten volumes of cryptic syntax and replaced it with lots of dreaming, learning and investigations that previously would have been impractical.
With Claude's help I can very effectively program in Rust despite not knowing it at all.
Anyone not using a coding agent is getting left behind.
The human brain is not built for coding. Our strength is being intuitive, thinking out of the box, using everything we've learned in the past to create new things.
Re: (Score:2)
The human brain is not built for coding.
My intuitive, out-of-the-box 2 cents: it sounds like you're not autistic. That puts you in the minority here.
Torvalds is at the top (Score:5, Interesting)
Linus Torvalds is at the top. He's a great programmer and he doesn't have to worry about the negative effects of AI. He can take just the good parts. But humanity as a whole cannot do that. He's missing the point.
Re: (Score:2)
I don't think he is. But Linux knows that his responsibility to the human race is the kernel, not some global LLM strategy or the like. And if you actually read and understand the article, "LLMs are great for production work!" is not what you come away with.
"So, I use AI to do the things that I'm bad at" (Score:1)
This quote from Torvalds shows exacly who is using AI and why.
Re: (Score:2)
Indeed. And Linus is not bad at anything important that he does.
GIGO (Score:2)
About a month before retiring, my boss said to me "AI is going to make everyone more productive: it will make the good, smart employees more productive, and we want that. It will also make the bad, dumb and lazy employees more productive, and I'm retiring so good luck, buddy!"