"AI" is a failed technology, in the sense that it's neither fit for purpose, nor a likely path to its stated end goal of "AGI", but a lot of people's profits currently depend on pretending it's not.
-
Re: provided its limitations are understood and appropriately managed
Those are some extremely large qualifications.
The number of people capable of understanding the limitations of LLMs (which I'm pretty sure is what everyone means by AI in these discussions) across a large number of knowledge domains is all too small and probably still greater than the number of people in management capable of managing them appropriately.
Dear fellow Mastodonians,
As much as I appreciate the responses to my earlier discussion with Mr. Rochko, who seems to have left the building
, I'd like to clarify one thing:My intention was to invite a debate about the overall argument, not to have every illustrative example picked apart in isolation or deal with strange cryptic name calling.
So, by all means, challenge my reasoning, disagree with my conclusions, and bring your counterarguments. That's precisely what I'm here for!
So in essence: let's focus on the substance of the discussion rather than nitpicking individual examples. Let's have a lively, respectful debate. Thank you!
-
@skavish Well, in music, the opinions of listeners generally are considered to be of primary importance, and it's perfectly common for music critics to be taken seriously despite playing no instruments at all.
So, while the fact that you use the tools gives you some basis for an informed opinion, I'd still tend to favor the opinion of someone who has to work with the code that you produce in ways that you didn't anticipate, such as by incorporating it into their open source project. Good code should be reusable, right?
I don't know whether you have that experience -- you haven't said -- but @Gargron certainly does.
food critics do not need to be chefs, but if you have never cooked, maybe do not lecture chefs about what their kitchen can produce
i have been a software engineer for 30 years and review 10–30 pull requests a day. the majority involve ai-generated code. i see firsthand what these tools produce, good and bad. you are arguing about what you imagine they can do, while i work with the results every day. code review is code review, regardless of who wrote the code.
-
food critics do not need to be chefs, but if you have never cooked, maybe do not lecture chefs about what their kitchen can produce
i have been a software engineer for 30 years and review 10–30 pull requests a day. the majority involve ai-generated code. i see firsthand what these tools produce, good and bad. you are arguing about what you imagine they can do, while i work with the results every day. code review is code review, regardless of who wrote the code.
-
@crazyeddie you are missing the point. his engineering expertise isn't in question. he explicitly said he's never used AI. so how can he confidently judge what AI-generated code looks like when the tools are used effectively?
@skavish
I have looked over several GNN generated codebases – most of them were the criticised spagetthi code, usually with bonus points for being instant abandonware.I have also seen a few that were actually well-maintained and based on their history it was obvious that a lot of effort was put into avoiding spagetthi, making sure their tests deserved the name, adding and refining long design documents, ...
Based on what I saw I’m very sceptical that they ended up being significantly faster using this technique than just limiting their GNN usage to interactive research (or even none at all if they are already well-versed in the given domain and programming environment) and then self-writing the code by hand. They definitely would have used only a fraction of the electricity and water that way for sure.
Why bother even trying to get into something you know is way worse in externalities than the status quo if there is no way it will be able to compensate in any way if done correctly?
@crazyeddie -
uggh, please read what i wrote again. we are going in circles. AI can produce shitty PRs when used badly. maybe that is what he has seen, i have no idea. used properly, it can produce great PRs. looking at PRs alone does not tell you what the tools are capable of when used well. to judge that, you need to actually use them. which he has explicitly said he does not want to do. and that is fine.
-
uggh, please read what i wrote again. we are going in circles. AI can produce shitty PRs when used badly. maybe that is what he has seen, i have no idea. used properly, it can produce great PRs. looking at PRs alone does not tell you what the tools are capable of when used well. to judge that, you need to actually use them. which he has explicitly said he does not want to do. and that is fine.
looking at PRs alone does not tell you what the tools are capable of when used well.
Yes it does! If the tools produce good PRs, then looking at those PRs tells you that the tools are capable of doing something good!
to judge that, you need to actually use them.
No you don't! You need to know about code, and you need to observe what these tools actually produce!
-
@skavish
I have looked over several GNN generated codebases – most of them were the criticised spagetthi code, usually with bonus points for being instant abandonware.I have also seen a few that were actually well-maintained and based on their history it was obvious that a lot of effort was put into avoiding spagetthi, making sure their tests deserved the name, adding and refining long design documents, ...
Based on what I saw I’m very sceptical that they ended up being significantly faster using this technique than just limiting their GNN usage to interactive research (or even none at all if they are already well-versed in the given domain and programming environment) and then self-writing the code by hand. They definitely would have used only a fraction of the electricity and water that way for sure.
Why bother even trying to get into something you know is way worse in externalities than the status quo if there is no way it will be able to compensate in any way if done correctly?
@crazyeddielooks like half the people in this thread have never used the tools themselves but have very strong opinions about them. that is fine, but it is impossible to have a meaningful discussion when people argue from assumptions rather than experience.
-
looking at PRs alone does not tell you what the tools are capable of when used well.
Yes it does! If the tools produce good PRs, then looking at those PRs tells you that the tools are capable of doing something good!
to judge that, you need to actually use them.
No you don't! You need to know about code, and you need to observe what these tools actually produce!
jesus fucking christ, you are arguing in circles. if you have seen good prs, then great, the tools can produce good code. if you have only seen bad ones, you know they can produce bad code. neither tells you what happens when someone who actually knows how to use the tools uses them.
this is pretty basic logic. I think I am done here
-
calling code that’s currently being generated by AI agents “spaghetti code” is a pretty sure sign that you either haven’t actually used it seriously or don’t know how to use it.
AI-generated code can absolutely be terrible. so can human-generated code. the interesting question is whether you know how to get the agent to produce good code and how to review, test, and evolve it.
@skavish @Gargron
I program for a living. I've reviewed AI generated code. It sucked. I've had AI review my code. It always got obvious shit wrong.It will continue to suck for the foreseeable future, because it's an imitation engine whose training data prominently features badly-written code and StackOverflow examples labeled ""Here's what I wrote; why doesn't it work?", and it has no actual intelligence.
The more reliable the coding AI gets, the more useless it will actually become, because folks will rely on it for bigger things, and finding the bugs that will inevitably turn up in major projects will be beyond them.
-
jesus fucking christ, you are arguing in circles. if you have seen good prs, then great, the tools can produce good code. if you have only seen bad ones, you know they can produce bad code. neither tells you what happens when someone who actually knows how to use the tools uses them.
this is pretty basic logic. I think I am done here
@skavish So the PR that someone who actually knows how to use the tool made doesn't tell you anything about their ability to use the tool?
The sound that I hear when someone plays a guitar doesn't tell me anything about their ability to play the guitar?
Do you listen to yourself?
-
@skavish @Gargron
I program for a living. I've reviewed AI generated code. It sucked. I've had AI review my code. It always got obvious shit wrong.It will continue to suck for the foreseeable future, because it's an imitation engine whose training data prominently features badly-written code and StackOverflow examples labeled ""Here's what I wrote; why doesn't it work?", and it has no actual intelligence.
The more reliable the coding AI gets, the more useless it will actually become, because folks will rely on it for bigger things, and finding the bugs that will inevitably turn up in major projects will be beyond them.
i have bad news for you: you do not know how to use it.
i do not know whose opinion you consider authoritative. how about Linus? he is famous for being strict about code quality, yet even he sees AI as a useful tool.
-
calling code that’s currently being generated by AI agents “spaghetti code” is a pretty sure sign that you either haven’t actually used it seriously or don’t know how to use it.
AI-generated code can absolutely be terrible. so can human-generated code. the interesting question is whether you know how to get the agent to produce good code and how to review, test, and evolve it.
@skavish I find takes like this increasingly hilarious. There’s the magical mythical engineers “doing it right” while the vast majority of users - by my own observation - are pumping out absolute trash. The incentive for quality just does not exist. Some people can build quality systems in spite of that but it is not the norm.
-
@skavish So the PR that someone who actually knows how to use the tool made doesn't tell you anything about their ability to use the tool?
The sound that I hear when someone plays a guitar doesn't tell me anything about their ability to play the guitar?
Do you listen to yourself?
ok, look, you probably know who Linus Torvalds is, right? if not google it. then check it out: https://thenewstack.io/torvalds-linux-ai-stance/
-
@skavish I find takes like this increasingly hilarious. There’s the magical mythical engineers “doing it right” while the vast majority of users - by my own observation - are pumping out absolute trash. The incentive for quality just does not exist. Some people can build quality systems in spite of that but it is not the norm.
on that I completely agree, but it does not mean the tool is bad, you just need to know how to use it
-
ok, look, you probably know who Linus Torvalds is, right? if not google it. then check it out: https://thenewstack.io/torvalds-linux-ai-stance/
@skavish I read the email from Linus that led to that whole incident, actually, so I know that he was talking about stuff like Codex, the bug-finding tool. Yeah, LLMs are pretty dang useful for when you have a decent idea of the kind of code that will trigger a bug, and need loads and loads of variants with different timing and stuff. Enhanced fuzzing.
Code quality is quite irrelevant for that purpose.
Is it your opinion that code quality is irrelevant, generally?
-
You sound like a cocaine user
RE: https://beige.party/@fuzzy/117417199780071912
Hint: most of the 1,266 accounts that I block are anti-AI.
Cc @skavish
-
@skavish I read the email from Linus that led to that whole incident, actually, so I know that he was talking about stuff like Codex, the bug-finding tool. Yeah, LLMs are pretty dang useful for when you have a decent idea of the kind of code that will trigger a bug, and need loads and loads of variants with different timing and stuff. Enhanced fuzzing.
Code quality is quite irrelevant for that purpose.
Is it your opinion that code quality is irrelevant, generally?
-
you're extrapolating from the worst possible use of AI to all uses of AI.
of course, millions of people can generate mountains of garbage with it. millions of people also write terrible code without AI.
dismissing AI without examining it is just ideology, not engineering.
correct.
worst possible use = most common use - at a massive scale = all use.
also - open source maintainers *definitely* need to inspect a lot of code - so not sure how you extrapolated that one.
anyway, blocked.
-
@skavish Uh, yeah, I guess I remembered the name of the agentic review tool wrong.
Ya got me.
-
@Gargron counterpoint: we are well past the point of no return for agentic engineering. There will always be purists who insist on writing code by hand but we aren’t returning to that being the default even after the bubble pops. And I think quiet use of AI as glue will propagate more or less unseen through a lot of tech until it is the expected/desired thing vs a bolted-on buzzword.
Are there myriad problems with both of these? Of course. But that much is not “failure” in any meaningful sense.
@mellowfish @Gargron “Agentic engineering” causes numerous problems, has serious quality control issues, and is extremely resource intensive. Currently, it is supported by hype, and by the widespread availability of subsidised compute. Following a market correction, those factors will evaporate; It is extremely premature and unfounded to assume the practice will continue in any form.
