Given Torvalds’ attitude towards AI, it’s not surprising.
“They’ve gotten good enough to find bugs because LLMs, AI, are just pattern matchers and code is nothing but patterns in a specific format.”
True, but let’s not forget that 2 + 2 = 4 can be achieved through a million different patterns[1], so just because AI can find a working pattern doesn’t make it good
Linus comparing AI to text editors and fucking compilers shows how high on his own farts he is. Given the same input, a compiler will always produce the same output, down to the last bit. The same cannot be said about AI - throw the same version of claude code with the same prompt on the same computer file and it will produce different results, maybe with a “similar enough” end result, but not identical.
AI is not a tool, it’s a million monkeys with keyboards that cannot be held accountable for their output.
(((((64 / 2) / 2) / 2) / 2) / 2) + (((((64 / 2) / 2) / 2) / 2) / 2) is a valid pattern for 2 + 2, for instance ↩︎
just because AI can find a working pattern doesn’t make it good
I have seen so much from AI supporters saying “it can be a good tool” to save time, but I’m not sure I’ve ever heard anyone try to make the claim that it’s a tool that leads to better code. That should be very telling.
I hate to be that guy but there are compilers with non-deterministic optimization strategies and they do tend to be the most performant. They’re just generally disallowed in major project distributions because the maintainers want the artifact to be verifiable by and end user for security reasons.
“LLMs are good enough now,” said Greg Kroah-Hartman, a veteran Linux kernel maintainer, at the Linux Foundation’s Open Source Summit in South Korea in August. “They’ve gotten good enough to find bugs because LLMs, AI, are just pattern matchers and code is nothing but patterns in a specific format.”
Also to be fair, that quote is about finding bugs and the guy might not have been talking about using AI to patch them. The article makes it sound that way, but the article also seems overly confident and pro-AI about some other things too.
You can deterministically find something that is clearly defined, but bugs don’t fall into that category
I know, but even regarding bug catching, there have been a LOT of bogus bug reports “discovered by AI” that could not be replicated in any way, shape or form. Granted, I’ve seen one or two articles saying that AI helped find real software bugs, but what’s the ratio there, 1:500 for useful/useless AI bug reports?
Granted, I’ve seen one or two articles saying that AI helped find real software bugs, but what’s the ratio there, 1:500 for useful/useless AI bug reports?
It’s worth distinguishing between LLM generated bug-reports submitted by randos, and developers themselves using LLMs to look for bugs in their software.
My own experience, with my own code-bases, is that the LLM I use is extremely good at finding real flaws in my code. It’ll also nitpick a lot, but it has has become better at triaging issues than it used to be, so that kind of thing is easy to ignore.
Because of that, having an LLM take a look at my code before I push it has become another QC step, along with running tests, linters, and formatters. The end result is higher quality software
bogus bug reports “discovered by AI” that could not be replicated in any way, shape or form
And we NEED to be more serious about use of secure verifiable identities for people so we can properly filter inputs from internet contributions. Anonymous sources should be de-prioritized to just above the known consistently bad actors. That’s nothing about AI or LLMs, that’s just the world continuing to be far too trusting of basically anonymous contributions.
If you want to submit a bunch of AI slop under your secure ID, be prepared for that ID to be de-prioritized beneath anonymous contributions.
If you want submissions from your secure ID to be taken seriously, at least task your LLM with building a proof-of-concept replication of the bugs in the reports your other LLM session creates before submitting it for humans to spend their time evaluating.
We have a human test department overseas that gets a little eager to pump up their metrics once in a while, we haven’t resorted to penalizing them (yet) but when they send an “observation” with no way to reproduce it, no log files to back it up, we tend to ignore that observation until the review meeting and then in the review meeting ask “have you seen this again? do you have any way to reproduce it? do you have screen shots or log files?” and when all those things are missing, the observation is closed with a “cannot reproduce” tag, waiting for another tester to see it again and produce more evidence.
Anybody writing articles about AI, or evaluating articles about AI, needs to get very specific: what month are they talking about, because if your reference point for “overly confident” is based on where things were 8-12 months ago, that’s very different from where things are today, and many times experts in the field project progress to what they expect for the future - that is of course unreliable speculation, but so far most of the reasonable optimism from the last year or two for progress in LLM software authorship and maintenance proficiency has been borne out in reality. I’m not talking about the “software is solved” numbskulls, but the progress in terms of being a valuable tool set the way C compilers were valuable tools to former Assembly programmers - yeah… it’s gone there.
I also recall trying to use Borland Turbo C++ in 1991 and abandoning it because it was too buggy, wrote the project du jour in C instead, saved the C++ rewrite for 1996.
Being a million monkeys or not, it is a tool. The relevant definition for tool is:
something (such as an instrument or apparatus) used in performing an operation
That’s not to say it’s a good tool, or that it’s a tool being used appropriately by many people. What even is appropriate use of AI?
Accountability should be definitely be properly assigned to the user. I’ve heard of tradies blaming their tools for shitty work, but I’ve never heard of one actually holding their hammer accountable…
Also, sometimes the responsibility falls onto the tool maker. Remember Crowdstrike? That software update that caused widespread outages in Windows computers used by many organizations, including in airports? It wasn’t any user’s fault, surprisingly it wasn’t even Microslop’s fault, it was Crowdstrike’s fault alone.
I’m not sure that’s appropriate even if it is the most common use …
I think there’s a difference between responsible and accountable. Accountable requires that you be held to account if things go pear shaped. Crowdstrike were responsible for their outage but I’m pretty sure their responsibility ended at at “here’s a new patch to fix our fuck up”. No one was charging or fining them for the outage or putting those responsible in prison. I think a toolmaker might be held accountable of their tool exploded when it wasn’t supposed to and was killing people. For example the Ford Pinto. A car is a tool for getting people from one place to another, and the toolmaker was held accountable for their shit design. If you were operating the car out of spec or driving it into walls, then you as the operator would be held responsible and most likely accountable.
Given the same input, a compiler will always produce the same output, down to the last bit. The same cannot be said about AI
LLM’s have something like a temperature setting that is producing the intentional randomness, so that you prompting it is not going around in circles all the time. (if you are running an army of agents this is actually important for usefulness)
Anyway this is not an inherent LLM thing, it can be turned off.
Fame does something to people, either that or he’s burnt out and this is a midlife crisis. I love unix based operating systems, but sometimes I feel Linux was the wrong thing to get popular unfortunately. It abandons so much of the core philosophy.
just because AI can find a working pattern doesn’t make it good
It’s a wee bit more sophisticated than that, and will tend to find 2 + 2 even when 16/4 - 8/4 gives the same result. And if you suspect it’s settled on an over-complicated solution you can add the prompt: “Search for simpler architectures which meet all requirements with less complexity.” It won’t always find the simplest possible, but generally speaking - I find LLM agents to be a bit better than my human colleagues in terms of finding good simple solutions and also finding problems within implementations.
None of us are perfect, and LLMs are trained on our writings. LLM agents within coding frameworks have progressed easily +10 years of human experience in their perfomance levels in the last 12 months. It’s still an blueberries to tangerines comparison, a lot of what makes the LLM agents “good” is their speed and thereby ability to experiment with alternatives in a reasonable time.
Given Torvalds’ attitude towards AI, it’s not surprising.
True, but let’s not forget that
2 + 2 = 4can be achieved through a million different patterns[1], so just because AI can find a working pattern doesn’t make it goodLinus comparing AI to text editors and fucking compilers shows how high on his own farts he is. Given the same input, a compiler will always produce the same output, down to the last bit. The same cannot be said about AI - throw the same version of claude code with the same prompt on the same computer file and it will produce different results, maybe with a “similar enough” end result, but not identical.
AI is not a tool, it’s a million monkeys with keyboards that cannot be held accountable for their output.
(((((64 / 2) / 2) / 2) / 2) / 2) + (((((64 / 2) / 2) / 2) / 2) / 2)is a valid pattern for 2 + 2, for instance ↩︎I have seen so much from AI supporters saying “it can be a good tool” to save time, but I’m not sure I’ve ever heard anyone try to make the claim that it’s a tool that leads to better code. That should be very telling.
I hate to be that guy but there are compilers with non-deterministic optimization strategies and they do tend to be the most performant. They’re just generally disallowed in major project distributions because the maintainers want the artifact to be verifiable by and end user for security reasons.
Huh, TIL. Off the rabbit hole of compiler (non)determinism I go
That quote was someone else
Also to be fair, that quote is about finding bugs and the guy might not have been talking about using AI to patch them. The article makes it sound that way, but the article also seems overly confident and pro-AI about some other things too.
You can deterministically find something that is clearly defined, but bugs don’t fall into that category
I know, but even regarding bug catching, there have been a LOT of bogus bug reports “discovered by AI” that could not be replicated in any way, shape or form. Granted, I’ve seen one or two articles saying that AI helped find real software bugs, but what’s the ratio there, 1:500 for useful/useless AI bug reports?
It’s worth distinguishing between LLM generated bug-reports submitted by randos, and developers themselves using LLMs to look for bugs in their software.
My own experience, with my own code-bases, is that the LLM I use is extremely good at finding real flaws in my code. It’ll also nitpick a lot, but it has has become better at triaging issues than it used to be, so that kind of thing is easy to ignore.
Because of that, having an LLM take a look at my code before I push it has become another QC step, along with running tests, linters, and formatters. The end result is higher quality software
And we NEED to be more serious about use of secure verifiable identities for people so we can properly filter inputs from internet contributions. Anonymous sources should be de-prioritized to just above the known consistently bad actors. That’s nothing about AI or LLMs, that’s just the world continuing to be far too trusting of basically anonymous contributions.
If you want to submit a bunch of AI slop under your secure ID, be prepared for that ID to be de-prioritized beneath anonymous contributions.
If you want submissions from your secure ID to be taken seriously, at least task your LLM with building a proof-of-concept replication of the bugs in the reports your other LLM session creates before submitting it for humans to spend their time evaluating.
We have a human test department overseas that gets a little eager to pump up their metrics once in a while, we haven’t resorted to penalizing them (yet) but when they send an “observation” with no way to reproduce it, no log files to back it up, we tend to ignore that observation until the review meeting and then in the review meeting ask “have you seen this again? do you have any way to reproduce it? do you have screen shots or log files?” and when all those things are missing, the observation is closed with a “cannot reproduce” tag, waiting for another tester to see it again and produce more evidence.
Anybody writing articles about AI, or evaluating articles about AI, needs to get very specific: what month are they talking about, because if your reference point for “overly confident” is based on where things were 8-12 months ago, that’s very different from where things are today, and many times experts in the field project progress to what they expect for the future - that is of course unreliable speculation, but so far most of the reasonable optimism from the last year or two for progress in LLM software authorship and maintenance proficiency has been borne out in reality. I’m not talking about the “software is solved” numbskulls, but the progress in terms of being a valuable tool set the way C compilers were valuable tools to former Assembly programmers - yeah… it’s gone there.
I also recall trying to use Borland Turbo C++ in 1991 and abandoning it because it was too buggy, wrote the project du jour in C instead, saved the C++ rewrite for 1996.
Being a million monkeys or not, it is a tool. The relevant definition for tool is:
That’s not to say it’s a good tool, or that it’s a tool being used appropriately by many people. What even is appropriate use of AI?
Accountability should be definitely be properly assigned to the user. I’ve heard of tradies blaming their tools for shitty work, but I’ve never heard of one actually holding their hammer accountable…
Industrial scale misinformation and scamming.
Also, sometimes the responsibility falls onto the tool maker. Remember Crowdstrike? That software update that caused widespread outages in Windows computers used by many organizations, including in airports? It wasn’t any user’s fault, surprisingly it wasn’t even Microslop’s fault, it was Crowdstrike’s fault alone.
I’m not sure that’s appropriate even if it is the most common use …
I think there’s a difference between responsible and accountable. Accountable requires that you be held to account if things go pear shaped. Crowdstrike were responsible for their outage but I’m pretty sure their responsibility ended at at “here’s a new patch to fix our fuck up”. No one was charging or fining them for the outage or putting those responsible in prison. I think a toolmaker might be held accountable of their tool exploded when it wasn’t supposed to and was killing people. For example the Ford Pinto. A car is a tool for getting people from one place to another, and the toolmaker was held accountable for their shit design. If you were operating the car out of spec or driving it into walls, then you as the operator would be held responsible and most likely accountable.
LLM’s have something like a temperature setting that is producing the intentional randomness, so that you prompting it is not going around in circles all the time. (if you are running an army of agents this is actually important for usefulness)
Anyway this is not an inherent LLM thing, it can be turned off.
Fame does something to people, either that or he’s burnt out and this is a midlife crisis. I love unix based operating systems, but sometimes I feel Linux was the wrong thing to get popular unfortunately. It abandons so much of the core philosophy.
It’s a wee bit more sophisticated than that, and will tend to find 2 + 2 even when 16/4 - 8/4 gives the same result. And if you suspect it’s settled on an over-complicated solution you can add the prompt: “Search for simpler architectures which meet all requirements with less complexity.” It won’t always find the simplest possible, but generally speaking - I find LLM agents to be a bit better than my human colleagues in terms of finding good simple solutions and also finding problems within implementations.
None of us are perfect, and LLMs are trained on our writings. LLM agents within coding frameworks have progressed easily +10 years of human experience in their perfomance levels in the last 12 months. It’s still an blueberries to tangerines comparison, a lot of what makes the LLM agents “good” is their speed and thereby ability to experiment with alternatives in a reasonable time.