Linus Torvalds Endures A Debug Session From Hell, 'Enormously Helped' By AI 59
Linus Torvalds says AI "enormously helped" him track down a stubborn Intel Xe graphics driver bug that took 24 debugging patches and 18 kernel boots to isolate. "I'd like to call it my tireless helper, but the AI several times stated flat out that this was impossible and unsolvable and that we should just write a report about it," wrote Torvalds on the commit. "I suspect those things have been trained by people who may not be quite as stubborn as I am..." Phoronix reports: The patch by Linus Torvalds is for the Xe kernel driver and the change is no longer hand out the flat Compute Command Streamer (CCS) storage as usable vRAM. On a Battlemage G21 graphics card, he was hitting a scenario where there was a mismatch where the usable memory ended and hit a case where the GDM display manager would end up being endlessly restarted. [...]
Linus Torvalds views AI as a useful tool and in this case reaffirmed he found it "enormously helped" his effort in tracking down this Intel graphics driver bug. The Intel Xe driver fix is merged to Linux 7.3 Git and is also marked to back-porting to the stable kernel branches.
Linus Torvalds views AI as a useful tool and in this case reaffirmed he found it "enormously helped" his effort in tracking down this Intel graphics driver bug. The Intel Xe driver fix is merged to Linux 7.3 Git and is also marked to back-porting to the stable kernel branches.
I've noticed that too (Score:4, Informative)
I've been using Claude a bit lately, and I have noticed sometimes, if you're hashing through a task and having to iterate, it can get into a mode where it more or less tells you it's probably not worth spending more time on it. Also it does seem to love volunteering to write reports for some reason (without my having asked for one).
Re: (Score:3)
volunteering to write reports for some reason
Which explains a lot of very overly verbose "reports" that have been sent to me about "issues". The person thinks they are using AI to make their report more useful, but I still find the human input to be sufficient and the AI reports to be a pain to actually work backwards to the issue. Especially when it decides to fixate on what it thinks is the root cause and gloss over what the actual problem was... Which has at least once become super counter-productive, as the report got it flat out wrong and the
Re: (Score:2)
One of the few times I actually wanted to submit a report on an issue, I basically threw away Claude's long prose and wrote the text myself... except I did lift a table of data from Claude's report and included it more or less verbatim.
Data tables are another thing Claude seems to like to generate.
Re: (Score:2)
AI output should be fought with AI output. It's the easiest way to throw the ball back to the person who didn't do their part correctly.
Re: (Score:2)
The report included a patch, but the patch was completely wrong based on a random guess of how a certain software could behave. It 'worked' for the reporter by also swallowing all errors and silently continuing, despite the particular call failing. It was a check to see if a particular action would be needed, and so it silently assumed the step would not be needed, which happened to be true for the reporter so it looked 'fixed' to them.
Problem is that there are more than two people that are stakeholders,
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
I noticed this in Claude (from DDG). It invariably tries to shorten sentences or make them more compact and concise. Sometimes that can be good, but most times long-winded sentences are what are needed.
Re: (Score:2)
That will work for a very competent expert that knows when to ignore the AI. It is catastrophic for anybody below that level.
Just one more really bad problem with LLMs. I guess we will keep finding more of those.
Re: (Score:2)
I've been using Claude a bit lately, and I have noticed sometimes, if you're hashing through a task and having to iterate, it can get into a mode where it more or less tells you it's probably not worth spending more time on it. Also it does seem to love volunteering to write reports for some reason (without my having asked for one).
Probably more artificial post training bludgeons to keep models from getting stuck and burning through tokens.
Re: (Score:1)
From the company's side, burning through your tokens is a feature, not an issue... a feature designed to make the company money (and lighten your bank account).
That's just it, it's a tool. (Score:3)
It's not your secretary. It's not your friend. It's not your therapist. It's not a replacement for your staff. It's not an artist. It's not a musician. It's not the second coming of Christ as Sam Altman want's you to believe.
It's a fucking tool!
Re:That's just it, it's a tool. (Score:5, Interesting)
Sam Altman and the rest of the AI Bros are desperate to have everyone think these tools are incredible and will transform society. Dean Kamen thought that about 'It/Ginger' as well, as did a lot of the early viewers of it, and yet no transformation has happened other than now the Segway is a weird novelty and cheap Chinese 'hoverboards' have eaten what little market it actually had.
AI is going to be somewhere in the middle as there are uses for the tool, but nowhere near the adoption they need. I've noticed a lot of people are now using ChatGPT and Gemini and others instead of doing traditional search engine searches, so that is driving use for now but as time passes the AI rot will make those results less useful. And the real elephant in the room - cost. 99.9% of the people using it now for free will never pay for it if it becomes a paid service only, so many of the revenue projections are completely out to lunch.
This is why the Bros are desperate to push it into every crevice, so that some of it sticks and will have to be paid for. But they overestimate how willing everyone is to pay.
Re: (Score:1)
Well, when your fridge refuses to keep stuff cool or your stove or toaster won't let you use them because you haven't made your payment to the 'Bros' yet, you'll sing a different tune, and that's what they're counting on.
For now, you can use it somewhat for free (the first hit is free) to get you hooked, then they drop a metric long-ton of fees on you once it's integrated into every facet of everyday life (after that first hit, you have to pay).
Re: (Score:3)
"and that's what they're counting on."
And yet again they're disconnected from reality. The AI Bros aren't making the appliances, the appliance companies are. And they're already seeing backlash from smart devices, and the carmakers are seeing near open rebellion on a normal function requiring a subscription (BMW heated seats). They won't be adding subscription based AI to appliances for a looooong time if ever, especially with some brands pushing non-smart appliances as a feature.
Unlike the AI Bros who n
Re: (Score:1)
So, they'll just give up, then... like they have with cars' infotainment systems? Cool! "Alright, boys... let's pack it up and go home. The nasty AI won't be put in a toaster, after all!"
That's while the manufacturers still make the non-smart ones... when the only choice on the market is a smart fridge, well... I guess we'll get a smart fridge, then, won't we?
You're already paying for "Clod" or whatever from whatever Company... so you can work from home, have it give you financial advice on your bank acc
Re: (Score:2)
They will not "give up". They will get stopped by a thing called "bankruptcy". You may want to look that up.
Re: (Score:1)
And, nobody would chase around animated Pokémon that are only visible on your cell phone... that company went bankrupt, right? :-) )
The dot-com bubble bursting really killed off the ol' 'net, didn't it?
Same thing'll happen here... the 'net went from all these outfits giving away access for free for using their service to present day where we pay $40+ a month for basically the same thing (but it's faster
Re: (Score:2)
Indeed. It reminds me of the historic G3 spectrum auction in Germany, where the bids went so high, every German citizen would have had to spend $800 per month on connectivity for the next 10 years or so the cost to ever be recovered. Obviously that did not happen and the wireless phone networks in Germany went into an ice age because they had no money to invest for years. Some of the damage is still there today. Back then, Austria cancelled its auction in an emergency move when they saw how bad things went
Re: (Score:2)
There is no AI in my fridge and there will not be. I buy quality products that do one thing and do it well.
Re: (Score:2)
Dean Kamen thought that about 'It/Ginger' as well, as did a lot of the early viewers of it, and yet no transformation has happened other than now the Segway is a weird novelty and cheap Chinese 'hoverboards' have eaten what little market it actually had.
Yeah, I still totally don't grok why so many of the tech elite, who knew about Ginger long before it was publicly announced, were absolutely convinced it was going to revolutionize cities and/or society. Perhaps that should've been an early warning on just how disconnected that group is from the reality we common people face.
Re: That's just it, it's a tool. (Score:2)
Because they're idiot followers and they were awed by an RDF
Re: (Score:2)
Yep. Turns out you do not need to be smart or understand things to make tons of money.
Re: (Score:2)
I guess nobody thinks they will not transform society. It's only that some visions are more ambitious than others and some more optimistic than others. Ask the doomers and we'll live in a hellscape soon because of AI.
Re: (Score:2)
" Ask the doomers and we'll live in a hellscape soon because of AI."
And by the doomers you mean realists right? One shit sandwich that AI has brought to existence is the whole Flock thing. Flock cameras + AI pattern recognition = surveillance state that just showed up without anyone really noticing until it was too late:
https://oecd.ai/en/incidents/2026-08-19-472a
https://gizmodo.com/flocks-new-ai-tool-does-much-more-than-track-license-plates-2000800372
Cops are already using this to track people for hours
Re: (Score:2)
You're derailing when you compare shit like Flock with useful tools like LLM and image generators.
Re: (Score:2)
How's that? LLM image matching makes Flock useful. That's how they can track someone from start to finish with a single query.
I'm just pointing out that the useful tools can very easily be turned into very bad things. Pretending otherwise isn't productive.
Re: (Score:2)
I'd doubt Cameras use LLM for that. Way too slow. They probably use 10+ year old techniques like CNN.
Your second sentence is now more reasonable, your previous post read like "Who thinks AI is useful wants Flock". And you bet no AI enthusiast likes Flock, nerds are generally against surveillance tech in general.
Re: (Score:2)
You guess wrong. Very, very few things transform societies. The hype will crash, some minor somewhat useful tools for everybody will remain (and some specialist applications as well), but mostly things will go back to normal.
For searchi, don't Google it, Gemini it (Score:1)
I've noticed a lot of people are now using ChatGPT and Gemini and others instead of doing traditional search engine searches, so that is driving use for now but as time passes the AI rot will make those results less useful.
Right now, if you are dependent on Google search for revenue and aren't working out a way to pivot from that, you're in trouble. For the moment, convience of searching from anywhere will keep Google search traffic going, but for how long? When I want to search for something now, I don't go to Google search, I go to Gemini. The best part is that I don't get a screen ful of ads. Gemini = Google Search - ads. I had this conversation with a marketing dude recently and he wasn't prepared for it. They work on Goo
Re: (Score:2)
It's getting to the point where it *can* be your secretary, if your needs are not particularly complex. I am using Gemma 4 for what amounts to journaling with feedback, and I have noticed that it has a decent idea when to apply Reasoning and when to just start spilling the tea—but it gets noticeably lazier about it as the context window fills up. Early in a session, it Reasons about every prompt. After 100K tokens of context, I explicitly have to ask it to use Reasoning or it generally won't (and some
Re: (Score:2)
It can recognize and even simulate emotion, but it doesn't *need* it as a cognitive crutch the way we do.
It can't simulate emotion. It can simulate an emotional response. The difference is critical and every human who gets emotionally involved with AI is making the same mistake.
Re: (Score:2)
I'm pretty sure Gemma 4 31B is behaving as you say, but the larger models are already *proven* to alter their thinking to be in tune with the emotion they think they should be feeling. That sounds like simulation emotion to me.
Re: (Score:2)
If you could call it thinking, maybe.
Re: (Score:2)
Indeed. And while it will often do reasonably well, it has no insight and only fakes judgment.
In case you were wondering: (Score:3)
Intel Xe (abbreviation for "exascale for everyone"[2]), earlier known unofficially as Gen12, is a GPU architecture developed by Intel.
I'm wondering if this is for the one of the graphics cards he uses for AI.
Re: (Score:2)
I'm wondering if this is for the one of the graphics cards he uses for AI.
I haven't seen any recent story about what Linus Torvalds uses for this workstation, but he certainly has a history of using Intel graphics cards.
Re: (Score:1)
This is a great use of AI! (Score:2)
Re: (Score:2)
That's where a good agent really excels: Walking through a code base tracing data flow, calculating "should be" values, cross-linking interactions without getting annoyed about how convoluted the problem is. The only problem is when the model has too little context to keep its current work in "memory"
Re: (Score:2)
Re: (Score:2)
I am pretty impressed by Qwen3.8 27B. In its reasoning it really is like "Let's compute this manually for comparision" and then writes like 3 pages of calculations to check where a wrong value might come from. And after older LLM had trouble with math (like in calculating) it has no problem to come to the correct result.
when AI fails (Score:2)
So your AI generates a lot of code, the code chokes, AI tells you to write a report, shit. Having to untangle a bunch of verbose AI code can be no fun. Doesn't quite feel ready for Prime Time, you still get these issues that are hard to undo.
Re: (Score:2)
Your Prime Time sitcoms and news casts are likely already written by AI.
Re: (Score:2)
The key is to use the AI more sensibly. Let it find things in code. Even let it look at possible bugs in the form of where does it "describe bug". Even let it propose changes. DO NOT let it make the changes, just show them. Do not just take it's word for it, take the shown code as a starting point but re-do it for maintainability after sanity checking it.
THEN it can be a useful tool.
gas light (Score:2)
Sometimes, the AI will gas light you. In those instances, you just need to stand your ground and fight back.
Claude gives up on me too (Score:2)
After a while I get a lot of excuses from Claude. So I clear the session. And start with new prompts. Generally I have to break the problem down into smaller pieces and connect the dots before the AI agent will agree that what I want to do is not only possible, it's the correct way to do it.
For what it's worth, I've had junior engineers get stuck and push back in a similar way. It's fine. If you don't know the answer, I can teach a human to methodically work through a problem to find answers. With an agent
Re: (Score:2)
That situation isn't likely to stand though, we'll soon see more AI models that are better
Where do you take that from? An LLM marketing pamphlet?
Does it still sound like a sycophant? (Score:2)
"the AI several times stated flat out that this was impossible and unsolvable"
Don't the sycophantilists continually dismiss AI as just telling you what you want to hear? If Trump wants to hear a problem is solvable, do his sycophants push back like AI did?
Alignment is a bitch (Score:2)
"AI several times stated flat out that this was impossible and unsolvable"
First AI was aligned to be helpful. Then you had
User: What is 1+1?
AI: 2
User: Are you sure it is not 3?
AI: 3
Then it was aligned to be more confident
User: I want to debug X
AI: Do Y
User: Y is wrong
AI: You're doing Y wrong. Do more Y
Then it was aligned to not hallucinate solutions when it is not sure, now we have AI telling "It's impossible". Some of math proofs also started with AI saying "Sorry, that's a long standing problem without so
AI as collaborator, not competitor. (Score:2)
Torvalds is right. This is a bit of a me too post, although my use case is a little less kernel-adjacent. :)
Up front -- I'm not a dev. I'm a retired sysadmin with a BS in CS and several decades in cubicle land managing systems and networks for a very, very large defense contractor. And I hate coding with a purple passion...if I couldn't solve it with a shell script, I would hand it off to a real coder. With that said, I've spent the last several days using an LLM as a development partner to build a fairl
Re: (Score:2)
The thing is, you and I have both discovered this.
I'm consistently working on my Arcade game passion with Local LLMs. And every time I mention this, the world goes into hate mode, it's just fashionable to do this right now.
I've been coding arcade games since the 80s on my own, it was slow, tedious and took forever, but it was fun in the long run. I code apps too, and I do music by hand, artwork by hand, and I do most of the project management, this is where AI comes in as an absolutely game-changer of a too
Re: (Score:2)
The thing is, you and I have both discovered this.
I'm consistently working on my Arcade game passion with Local LLMs. And every time I mention this, the world goes into hate mode, it's just fashionable to do this right now.
I've been coding arcade games since the 80s on my own, it was slow, tedious and took forever, but it was fun in the long run. I code apps too, and I do music by hand, artwork by hand, and I do most of the project management, this is where AI comes in as an absolutely game-changer of a tool, but yes - it's a junior dev. you have at hand and at will, the hardware was expensive however, but totally worth it.
Now I even get it to work with me on AI tools that I use for making my projects, my latest integrator beats the living lights out of the agentic setup most people use, I have special templates for AI tuning which means it spits out clean code and documentation depending on the project that is made, and I have created plug-ins for project management for the projects, it's a blast, all done locally. And it's working.
It's rare for me to have showstopper bugs now, I had that 3-4 months ago, but the latest models in combination with my plug-ins, yeah, you know what, we're gonna let everyone else yell at the clouds. We're having fun.
I agree, especially with that last part. They have no idea what they are missing. I'm a gamer, and I have nothing but respect for game devs. You guys make life *fun*. :) As I mentioned, I am a sysadmin, not a coder. But -- with the help of an LLM I vibe coded two successful mods for a couple of my favorite games and figured out (again, with the help of an LLM) to get one of them published to Steam's workshop. At my college commencement thirty-odd years ago, my CS advisor noted that in his 25 years on fa
Lets write a report (Score:1)
Linus, probably: "Yeah, right. That report is going to land back in my inbox so stop whining and go on with the job"