unpopular opinion, especially on the fediverse, but: AI is not inherently bad.
it is a tool, that can do everything to a pretty low standard but rather fast. so it is up to the learnt human to use the thing at the right time. i use the company ai at regularly at work, but treat it as an apprentice: every line of code has to be checked, because it will fail hilariously. but it does type faster than me, so simple and boring jobs can be done by it.
i don’t see the appeal of avoiding generated code at all cost. it is not different from human code in that it can fail horribly. the important thing is oversight. is there someone with the knowledge and keen eye to ensure no BS is merged. Linus, in my opinion, spent a lifetime building that reputation.
nothing is good nor evil, but thinking makes it so.
But i see the issues on both sides of the aisle. Kind of like how social media is a great tool that brings about instant worldwide communication. But those merits don’t cancel out how it’s been used to exploit the human psyche, influence elections, and even coax genocides.
In this case, the downwind effects from encouraging use by mature professionals is impacting the impressionable future. And no one seems to care about the future anymore.
I code at a kindergarten level so its been useful for me when I’m trying to get an idea to just work. I tell it what I was trying to do and give it my code and then I ask it to fix then explain what went wrong. Works pretty well in qwen 3.8 27b on my GPU in zed. But I’m not making huge apps or public code just dumb scripts for me. So ive found ive actually gotten better this past year. Its all in how you use it.
I have access to essentially unlimited local compute at my work. Some of my coworkers use it to run LLMs, some use the pre-enshittified corporate models, and I gave all that a shot—like really went for it, but ultimately would rather just work at my own pace.
In my workplace the people who use them come off as kind of frantic to my eye. Like they’re always making these overly engineered greenfield things no one asked for, which gets them a shout out at quarterly and then no one ever thinks of it again…
I’m not saying there’s 0 legitimate use for the technology if you could divorce it from its problematic origins, but I’m not that smart, and people are increasingly coming to me for help when previously they seemed happy to just talk to their little corporate minder.
I feel like chatbot and text generating models generally encourage dark patterns we don’t even have names for yet and I’m just highly wary of the whole thing.
I’ve spent the last 9 months coming in behind these “look what I did! AI!” projects and fixing them. One was turned over to a coworker as a “here I did all this work for you” … Wouldn’t even start because it tried to bind to the same TCP port three times in the init, for some reason.
It’s getting a little better, but I still mostly unfuck shit that people LLMd and then (they) walked away to much fanfare at their “expertise” to do it again.
“oversight” is a myth. I’m pretty sure you can’t oversight your way of plagiarism out of millions of training data sources that you don’t even have on your local disk for comparison and reference.
you are confusing two issues here: oversight for the output of an ANN can very much achieve good results. as i explained in my post.
oversight over the output won’t help you with copyright, water and energy consumption, slave labour and all the ethical issues further up the pipeline. but we don’t need ANNs for big corpos to do all these evil things. we need oversight and real consequences for those corpos - no matter what they produce.
ANNs as a technology are old and have not fundamentally changed since Alan Turing. Sure, we have iterated and improved. But the fundamentals are the same. LLMs just made that old tech quite popular recently and introduced a “line go up” race. i do not believe that we will gain significant improvements from simply feeding more stolen works to the machine. A fraction of the MNIST dataset is enough to train an ANN on a 20 years old laptop to recognise the digits 0-9 reliably. The technology is sound, limited in its usability and detached from the big corpos.
so why are you advocating for not avoiding all generative AI code then? you specifically cited Linus, who does as far as i know not ensure cleared up training data (what would that even be, CC0 only?) like you seem to be advocating for.
once again: oversight of the production line != oversight of the output.
i specifically said that i do trust Linus Torvalds to ensure no bullshit is pushed to the Linux kernel. that is oversight of the output.
we need oversight of the production line. but of all the production lines, not only the ones delivering LLMs. big corporations have a tendency to blatantly break the law and get away with a fee that is smaller than their profit. if they don’t break the law by the letter, they have good lawyers to skirt around it and noone has a morality police. oversight would mean not only catching these things but also punishing such behaviour in a meaningful way.
training an LLM eith CC0 would be fair game. properly bought (declared it is for training some AI) resources as well. i fully disagree with the current training methods: just feeding more to the black box - to that extent, that they buy (improperly) and destroy books just to get some more words into their respective model. i am sure, that other approaches would get better results. but that is slower and more expensive. so the problem is how most AI companies operate - not the actual product. but if it wasn’t “AI”, it would be something else with the exact same bullshit practices.
I don’t think it’s inherently bad. We don’t need to use double the world’s energy to power it (as we see with open weight models) . We don’t need to centralize it in trillion aire companies. We don’t need to leverage an entire global economy on it. We don’t need to use it to replace human labor without a proper economic model.
Those are all aspects of greed and stupidity. None of that is inherent.
I’d agree with you if you mean that what has been marketed as AI to us is inherently bad. That said the word AI has been misused so much that is basically has no meaning. I’ve known researchers who have used things that can be called AI for really important research and it could run on a high end computer from 10 years ago. The externalities of that are quite minimal to what we gain. We really need to find language to discuss AI where when we talk about Big AI we’re not also dragging the proven, and efficient models which are nearly AI but not quite into the discussion.
It really depends on which externalities you’re talking about
The only one I can think of as ‘inherent’ is the ethical concern over copyright. The others are more broadly related to the commercialization of the tool than to the tool itself.
It’s worse than that, even. The calcs and even the Wikis were below or at human error levels, the AIs will peak below 95% accuracy with infinite power and training according to studies by OpenAI in 2020 and Deepmind in 2022 correcting a math error they made.
I think there’s a big difference between
A. Expecting people to pay you and give you credit for something AI made for you
vs
B. Playing with AI and if it spits out something you like a lot you keep it to yourself unless someone else specifically asks you to show it to them.
It’s basically masturbation.
People are gonna do it to themselves and enjoy it but if you try to show me that shit without my consent we are not going to be friends.
We’re not talking about the average clueless vibe coder here. Programming involves a lot more than just writing code. What Torvalds did was essentially “AI assisted coding”. As i understand it, he used it to narrow down the position of the bug in the code and then fixed it himself with a one-liner because apparently AI couldn’t find a solution on its own:
And this was a debug session from hell, enormously helped by an AI
doing much of the grunt-work.
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.
I suspect those things have been trained by people who may not be
quite as stubborn as I am.
But while the AI was ready to give up several times, it did keep
adding debug code and analyzing it faithfully when I pushed. So credit
where credit is due and I let the AI write the commit message above.
This is basically a one-liner fixing a bogus “round_up()” to a
“round_down()”, but there were 24 patches adding more and more debug
information to this, and 18 kernel boot to finally narrow it down to
this. - Linus
unpopular opinion, especially on the fediverse, but: AI is not inherently bad.
it is a tool, that can do everything to a pretty low standard but rather fast. so it is up to the learnt human to use the thing at the right time. i use the company ai at regularly at work, but treat it as an apprentice: every line of code has to be checked, because it will fail hilariously. but it does type faster than me, so simple and boring jobs can be done by it.
i don’t see the appeal of avoiding generated code at all cost. it is not different from human code in that it can fail horribly. the important thing is oversight. is there someone with the knowledge and keen eye to ensure no BS is merged. Linus, in my opinion, spent a lifetime building that reputation.
But i see the issues on both sides of the aisle. Kind of like how social media is a great tool that brings about instant worldwide communication. But those merits don’t cancel out how it’s been used to exploit the human psyche, influence elections, and even coax genocides.
In this case, the downwind effects from encouraging use by mature professionals is impacting the impressionable future. And no one seems to care about the future anymore.
I code at a kindergarten level so its been useful for me when I’m trying to get an idea to just work. I tell it what I was trying to do and give it my code and then I ask it to fix then explain what went wrong. Works pretty well in qwen 3.8 27b on my GPU in zed. But I’m not making huge apps or public code just dumb scripts for me. So ive found ive actually gotten better this past year. Its all in how you use it.
I have access to essentially unlimited local compute at my work. Some of my coworkers use it to run LLMs, some use the pre-enshittified corporate models, and I gave all that a shot—like really went for it, but ultimately would rather just work at my own pace.
In my workplace the people who use them come off as kind of frantic to my eye. Like they’re always making these overly engineered greenfield things no one asked for, which gets them a shout out at quarterly and then no one ever thinks of it again…
I’m not saying there’s 0 legitimate use for the technology if you could divorce it from its problematic origins, but I’m not that smart, and people are increasingly coming to me for help when previously they seemed happy to just talk to their little corporate minder.
I feel like chatbot and text generating models generally encourage dark patterns we don’t even have names for yet and I’m just highly wary of the whole thing.
I’ve spent the last 9 months coming in behind these “look what I did! AI!” projects and fixing them. One was turned over to a coworker as a “here I did all this work for you” … Wouldn’t even start because it tried to bind to the same TCP port three times in the init, for some reason.
It’s getting a little better, but I still mostly unfuck shit that people LLMd and then (they) walked away to much fanfare at their “expertise” to do it again.
It’s a tool, not a replacement for thought.
“oversight” is a myth. I’m pretty sure you can’t oversight your way of plagiarism out of millions of training data sources that you don’t even have on your local disk for comparison and reference.
(This isn’t legal advice. I’m not a lawyer.)
you are confusing two issues here: oversight for the output of an ANN can very much achieve good results. as i explained in my post.
oversight over the output won’t help you with copyright, water and energy consumption, slave labour and all the ethical issues further up the pipeline. but we don’t need ANNs for big corpos to do all these evil things. we need oversight and real consequences for those corpos - no matter what they produce.
ANNs as a technology are old and have not fundamentally changed since Alan Turing. Sure, we have iterated and improved. But the fundamentals are the same. LLMs just made that old tech quite popular recently and introduced a “line go up” race. i do not believe that we will gain significant improvements from simply feeding more stolen works to the machine. A fraction of the MNIST dataset is enough to train an ANN on a 20 years old laptop to recognise the digits 0-9 reliably. The technology is sound, limited in its usability and detached from the big corpos.
so why are you advocating for not avoiding all generative AI code then? you specifically cited Linus, who does as far as i know not ensure cleared up training data (what would that even be, CC0 only?) like you seem to be advocating for.
once again: oversight of the production line != oversight of the output.
i specifically said that i do trust Linus Torvalds to ensure no bullshit is pushed to the Linux kernel. that is oversight of the output.
we need oversight of the production line. but of all the production lines, not only the ones delivering LLMs. big corporations have a tendency to blatantly break the law and get away with a fee that is smaller than their profit. if they don’t break the law by the letter, they have good lawyers to skirt around it and noone has a morality police. oversight would mean not only catching these things but also punishing such behaviour in a meaningful way.
training an LLM eith CC0 would be fair game. properly bought (declared it is for training some AI) resources as well. i fully disagree with the current training methods: just feeding more to the black box - to that extent, that they buy (improperly) and destroy books just to get some more words into their respective model. i am sure, that other approaches would get better results. but that is slower and more expensive. so the problem is how most AI companies operate - not the actual product. but if it wasn’t “AI”, it would be something else with the exact same bullshit practices.
AI is inherently bad. It might be useful, but their externalities are quite inherent to it.
I don’t think it’s inherently bad. We don’t need to use double the world’s energy to power it (as we see with open weight models) . We don’t need to centralize it in trillion aire companies. We don’t need to leverage an entire global economy on it. We don’t need to use it to replace human labor without a proper economic model.
Those are all aspects of greed and stupidity. None of that is inherent.
I could say the same about all marketing, but that won’t make it disappear
I’d agree with you if you mean that what has been marketed as AI to us is inherently bad. That said the word AI has been misused so much that is basically has no meaning. I’ve known researchers who have used things that can be called AI for really important research and it could run on a high end computer from 10 years ago. The externalities of that are quite minimal to what we gain. We really need to find language to discuss AI where when we talk about Big AI we’re not also dragging the proven, and efficient models which are nearly AI but not quite into the discussion.
Perfectly put. I’ve expended paragraphs circling this succinct point.
It really depends on which externalities you’re talking about
The only one I can think of as ‘inherent’ is the ethical concern over copyright. The others are more broadly related to the commercialization of the tool than to the tool itself.
It’s the same as people saying not to use calcs or Wikipedia, new times same issues.
But we very much do enforce times on when not to use calcs. Thars the whole point here.
It’s worse than that, even. The calcs and even the Wikis were below or at human error levels, the AIs will peak below 95% accuracy with infinite power and training according to studies by OpenAI in 2020 and Deepmind in 2022 correcting a math error they made.
I think there’s a big difference between
A. Expecting people to pay you and give you credit for something AI made for you
vs
B. Playing with AI and if it spits out something you like a lot you keep it to yourself unless someone else specifically asks you to show it to them.
It’s basically masturbation.
People are gonna do it to themselves and enjoy it but if you try to show me that shit without my consent we are not going to be friends.
We’re not talking about the average clueless vibe coder here. Programming involves a lot more than just writing code. What Torvalds did was essentially “AI assisted coding”. As i understand it, he used it to narrow down the position of the bug in the code and then fixed it himself with a one-liner because apparently AI couldn’t find a solution on its own:
https://github.com/torvalds/linux/commit/818bebeb63dd6bf5f4e07e145f6cdbace520a34c>
That’s basically how AI should be used. To speed up things that are technically easy to do, but still very time consuming.