Linus Torvalds used Claude Code to build a Linux page cache fix that hit 23x read throughput, but kernel reviewers found a race before anything merged.
LLMs make slop code faster than humans can meaningfully review. If you make LLMs the only review layer, then slop gets through because they have similar biases. Stop making slop.
Cool assumption. But pretty sure it’s not how things work at all but sounds great…
What actually happens is that you have a long list of multiple maintainers reviewing , and multiple branches for testing, and multiple layers of review too.
Yes this means the review process is working. What is important to track is if LLM code causes a disproportionately large number of reviews or reviews that take longer because whoever generated the code doesn’t actually know what it is doing. Hopefully not the case here but it does happen a lot with LLM coding.
“AI bad. AI do bad thing. Let’s point and laugh at AI doing bad thing to delegitimize the technology in the public’s eyes.”
Meanwhile, at work, I interact with coding agents on a regular basis, and the results are pretty good. Then it goes through a local blind review by a more powerful agent, get pushed out to the company’s code review agent, which is designed to be as critical as possible, check the ticket details, make sure everything matches what the design plan says. Depending on the repo, it might even go through an extra security review.
Then, because we aren’t fucking stupid, it goes through the standard code review process by real developers. Sometimes they still find something, sometimes the agent code reviews found way more than a human can. Overall, I trust the current process over our previous one, where some guy spends barely two minutes looking at the code, rubberstamps it, and then QA discovers a few bugs that need fixing, because the damn code review wasn’t as through and critical as it should have been.
Thank you. This is a tool just like everything else. Garbage in equals garbage out. Generative coding in the right hands is an extremely powerful tool.
It’s a difference between kids playing with fireworks. (Vibe coders that know nothing about software development)
And rocket scientists at NASA building a new shuttle. (Real developers with controlled thoughtful AI workflows)
You can dev as long as you like, doesnt make you good at using a new tool by default. Its a tool, it works for millions of people. Why do you think you’re special.
Real developer who has been programming daily since the 80s. The quality of the models used to vary pretty wildly, but here in the past year or two, everything has gotten way better and most models are pretty capable. Claude, Gemini, GLM, Kimi, DeepSeek. The prices haven’t, but my advice is to ignore the bullshit and just pick the models that work for you. Claude isn’t 25x better than DeepSeek, so it doesn’t deserve to be 25x the cost.
I think Claude and OpenAI have started to realize that nobody’s buying their overpriced token costs. Although, frankly, none of these prices even make sense compared the cost of compute for an average cloud-based GPU cluster. I can’t see how any of these companies are making a profit, just riding the private equality wave until the money runs out. Maybe Amazon, Microsoft, Google, since they own their own hardware, but not OpenAI or Anthropic. China’s going to run them out of business in a few years, anyway.
And of course, if you have the option, use local models. VRAM size is still a bit of a problem, but quantization techniques have gone a long way. You can run a decent model on a personal PC with a bit of tweaking, for home programming tasks, or just a personal model that doesn’t immediately sell your data to Peter Thiel.
As an electrical engineer who has been doing full stack dev, firmware all the way up to board design and layout for about 16 years now, it’s getting better every single day. My boss is finally coming around after being the biggest shit talker for the last year and half. Is the code perfect? No. Does anyone write perfect code… NO! It is a tool that needs to be driven correctly. If you don’t like it, don’t use it! Is it going away? No! And it doesn’t look like anything we say is going to change that. Live with it, it’s the only option.
Everyone is entitled to their own opinions. Everyone has their own experiences too.
Be fucking nice tho, why is everyone so hostile. Go back to reddit if you want to be an asshole.
Not you dread beef, just in general.
My boss has been in it since like the early 80s, programming with basic and assembly. The turning point for him was a month or so ago when Claude found a bug in our mesh firmware that he had been trying to figure out for months. Edge case in hub mode when a message was in route to a device that momentarily dropped from the mesh. Flag gets cleared but no timer set to clear the buffer if the device doesn’t show back up in time. Super fun async RF problems, intermittent and ridiculously hard to find proxy buffer issue that was nearly impossible to reproduce. Old devs made code to lie and say the device was offline when it was still send it’s UID over RDM. So they knew the bug was there for YEARS. With a ton of logs and access to the specific files, Claude found the bug in like 45 mins. It was kind of crazy.
the last refuge for people who cant mount a proper retort. As expected of someone like you. You grab the low hanging fruit, blind to the trap, then think you’re a world leading genius for going directly where you’re directed.
Wow! Says the person who started off throwing insults by calling me smug. I only agreed with your own assement of yourself via your username, if you took that as insult, it’s a you problem. But, at least you know that insulting people is disrespectful. Maybe try heeding your own advice before you start talking shit to strangers on the Internet. How “astoundingly” oblivious of you to make that comment.
AI bad. Bugs never occurred when human built code. We had to come up with a completely new term, unknown in programming before, called “race condition”.
That’s what code review is for lol.
Where is the story here?
LLMs make slop code faster than humans can meaningfully review. If you make LLMs the only review layer, then slop gets through because they have similar biases. Stop making slop.
You had to lean on that assumption that there’s a massive amount being produced to make this point though.
The actual operating point is that the slop is made faster than humans can WRITE it.
Reviewing 100 lines of code thats 90% there and took 3 minutes to produce is faster than writing it from scratch, there’s no getting around that.
For me the issue with it is that all (hyperbole) work is about to become reading instead of creating, dull.
Cool assumption. But pretty sure it’s not how things work at all but sounds great…
What actually happens is that you have a long list of multiple maintainers reviewing , and multiple branches for testing, and multiple layers of review too.
https://www.kernel.org/doc/html/next/process/maintainers.html . See that long list of maintainers for a start.
The actual policy is that everything ends up getting reviewed by a human. If I post a patch, it doesn’t go straight to master
If it’s not being properly tested and maintained, it gets dropped irrespective of what it is (remember devfs?)
Also, I’m not sure about the kernel specifically, but on projects I submitted code to, they often wanted complex patches broken down if possible.
Yes this means the review process is working. What is important to track is if LLM code causes a disproportionately large number of reviews or reviews that take longer because whoever generated the code doesn’t actually know what it is doing. Hopefully not the case here but it does happen a lot with LLM coding.
Linus secretly understands nothing about the Linux kernel
“AI bad. AI do bad thing. Let’s point and laugh at AI doing bad thing to delegitimize the technology in the public’s eyes.”
Meanwhile, at work, I interact with coding agents on a regular basis, and the results are pretty good. Then it goes through a local blind review by a more powerful agent, get pushed out to the company’s code review agent, which is designed to be as critical as possible, check the ticket details, make sure everything matches what the design plan says. Depending on the repo, it might even go through an extra security review.
Then, because we aren’t fucking stupid, it goes through the standard code review process by real developers. Sometimes they still find something, sometimes the agent code reviews found way more than a human can. Overall, I trust the current process over our previous one, where some guy spends barely two minutes looking at the code, rubberstamps it, and then QA discovers a few bugs that need fixing, because the damn code review wasn’t as through and critical as it should have been.
Thank you. This is a tool just like everything else. Garbage in equals garbage out. Generative coding in the right hands is an extremely powerful tool.
It’s a difference between kids playing with fireworks. (Vibe coders that know nothing about software development) And rocket scientists at NASA building a new shuttle. (Real developers with controlled thoughtful AI workflows)
Real developer who has done programming every day for over a decade now. AI is shit. It’s bad.
You can dev as long as you like, doesnt make you good at using a new tool by default. Its a tool, it works for millions of people. Why do you think you’re special.
“Real developer who has done assembly every day for ten years now. Compilers are shit. It’s bad”
Real developer who has been programming daily since the 80s. The quality of the models used to vary pretty wildly, but here in the past year or two, everything has gotten way better and most models are pretty capable. Claude, Gemini, GLM, Kimi, DeepSeek. The prices haven’t, but my advice is to ignore the bullshit and just pick the models that work for you. Claude isn’t 25x better than DeepSeek, so it doesn’t deserve to be 25x the cost.
I think Claude and OpenAI have started to realize that nobody’s buying their overpriced token costs. Although, frankly, none of these prices even make sense compared the cost of compute for an average cloud-based GPU cluster. I can’t see how any of these companies are making a profit, just riding the private equality wave until the money runs out. Maybe Amazon, Microsoft, Google, since they own their own hardware, but not OpenAI or Anthropic. China’s going to run them out of business in a few years, anyway.
And of course, if you have the option, use local models. VRAM size is still a bit of a problem, but quantization techniques have gone a long way. You can run a decent model on a personal PC with a bit of tweaking, for home programming tasks, or just a personal model that doesn’t immediately sell your data to Peter Thiel.
As an electrical engineer who has been doing full stack dev, firmware all the way up to board design and layout for about 16 years now, it’s getting better every single day. My boss is finally coming around after being the biggest shit talker for the last year and half. Is the code perfect? No. Does anyone write perfect code… NO! It is a tool that needs to be driven correctly. If you don’t like it, don’t use it! Is it going away? No! And it doesn’t look like anything we say is going to change that. Live with it, it’s the only option.
Everyone is entitled to their own opinions. Everyone has their own experiences too. Be fucking nice tho, why is everyone so hostile. Go back to reddit if you want to be an asshole. Not you dread beef, just in general.
My boss has been in it since like the early 80s, programming with basic and assembly. The turning point for him was a month or so ago when Claude found a bug in our mesh firmware that he had been trying to figure out for months. Edge case in hub mode when a message was in route to a device that momentarily dropped from the mesh. Flag gets cleared but no timer set to clear the buffer if the device doesn’t show back up in time. Super fun async RF problems, intermittent and ridiculously hard to find proxy buffer issue that was nearly impossible to reproduce. Old devs made code to lie and say the device was offline when it was still send it’s UID over RDM. So they knew the bug was there for YEARS. With a ton of logs and access to the specific files, Claude found the bug in like 45 mins. It was kind of crazy.
the level of smugness in this reply, to compare AI slopists with NASA rocket scientists, its fucking astounding
The smugness in assuming that ‘anyone who uses x tool is stupid.’
Self awareness at critical low.
I was making a point about people who are experts in their field, and people who know nothing about said field.
Username checks out.
the last refuge for people who cant mount a proper retort. As expected of someone like you. You grab the low hanging fruit, blind to the trap, then think you’re a world leading genius for going directly where you’re directed.
Wow! Says the person who started off throwing insults by calling me smug. I only agreed with your own assement of yourself via your username, if you took that as insult, it’s a you problem. But, at least you know that insulting people is disrespectful. Maybe try heeding your own advice before you start talking shit to strangers on the Internet. How “astoundingly” oblivious of you to make that comment.
You okay man? You read so much into my comment it makes me worry a little for ya. Hope you have a better day man.
Slop, not even once
AI bad. Bugs never occurred when human built code. We had to come up with a completely new term, unknown in programming before, called “race condition”.