Eh, I didn't really like this post. Chiefly, because it feels like the author picked a conclusion that hasn't come to pass yet ("AI will make all coding / intellectual work / blogging / paper writing stop completely") and taking it as fact argues backwards how terrible that'll be. And, like, yes, I don't want us to reach that world either.
But I also don't really feel like that's the direction we're heading? Even at my super AI-bullish company nobody is saying "do black-box development and let the AI do it all". We are expected to know how they work and strongly encouraged to use it, but there is still the understanding that we "own" the code and are expected to understand and maintain it. I understand that's an n = 1 anecdote, but it's not like OP's arguments are on any stronger foundations.
With Math, I'm seeing a strong pushback as well, as mathematicians point out that LLMs may confirm the truthfulness of statements, but their raw proofs aren't nearly as useful as one constructed by a person.
"You write a blog post: who’s going to read it?" / "You design a revolutionary new programming language: who cares?" / "You write a library, and publish it on GitHub: who cares?" This question trio is what ultimately prompted me to write this comment, because I find it frankly a little ridiculous. Even if the current anti-AI sentiment wanes (which is very much not a foregone conclusion yet), people will always want to read cool writeups and hand-tinker on stuff.
Now they talk about “prompts”, “harnesses”, “loops”. The discourse is narrower, shallower, and more repetitive. There’s only so many times I can hear about “agentic harnesses” before I lose my mind.
This one I'll give to the author. I also experience it and it gives me such fatigue. The constant churn of "you need to use BowelMovement-Sloptastic-420-69-QED with this particular fork of vllm.cpp or else you're missing out" or "BigModel 10.5 does 10 points higher on the benchmarks than LargeModel 7.7, but LargeModel 7.7 is much cheaper, so you should still use it"...
It's like people are discussing Pokémon or something. And not even just the base game, but minmaxing some artificially kneecapped challenge.
I agree that it's a bit reductive and defeatist (and the bleak world it paints is more of a hypothetical at this point), but I think this is somewhat different from the author's argument:
people will always want to read cool writeups
There will always be people who like to read writeups, but it's not the same as a vibrant intellectual community of peers the author is talking about. Commodification of good-at-first-glance work removes quality signals and diminishes the prestige of doing good work.
In addition, any new useful signal comes with an incentive to fake it*, and that's only a prompt away.
Maybe what we'll see is tighter and more closed professional groups and an increase of the value of in-network reputation, which is much more difficult to fake. This would be sad, because our industry became a social lift for so many people precisely because it had less emphasis on networks and more on contributions than other highly-paid industries.
*: as an example, "rigorous benchmarking vs the alternatives" used to be a useful signal. Look at how slop benchmarking got weaponised by a certain racist Danish-American entrepreneur lately.
There will always be people who like to read writeups, but it's not the same as a vibrant intellectual community of peers the author is talking about.
Fair, however, I still don't think is necessarily a given just because LLMs are a thing now. For one, LLM generated texts are ghoulish and even people otherwise open to the tech recoil in disgust. Unless they somehow figure out how to autonomously give the models a voice that doesn't have such a visceral effect (I'm doubtful!), that means production of content will always have a human element.
As for consumption, I just don't see why people would completely stop intellectually engaging with things. LLMs are very impressive in many aspects, but they aren't curious nor innovative. You can let them lose on problems you might not be able to solve yourself, but you still have to have the knowledge about these problems' existence and the gist of them. Then there are all the people who use LLMs as more "research partners" / "assistants", while still taking charge. I don't expect them to stop wanting to improve themselves in the field either.
I suppose the argument boils down to whether you expect LLMs to improve to an extent where they'll "human better" than actual humans. I personally don't. I expect there to be a theoretical limit and I expect said limit to be well below AGI.
Maybe what we'll see is tighter and more closed professional groups and an increase of the value of in-network reputation, which is much more difficult to fake.
Funny, but I expect this less in professional circles and more in hobbyist communities. The more AI can ragebait and act like a nuisance, the more communities are incentivized to vet their new members, which could result in more vouching systems ala Lobsters.
In professional contexts, I think there is already a sort of inherent filtering. You can make it as a shoddy programmer, but if you want to get into specific branches and certain companies, you must be able to rise above the rest. And an LLM won't necessarily let you do that (because you either don't have access to it or the domain knowledge is simply so complex that no generalist model can hope to understand it like a person who spent quite a bit of time on it).
On your last point: Or because the people at the branches and companies you want to get into are also using LLM tools effectively, so using an LLM yourself becomes just table stakes, not a competitive advantage.
Chiefly, because it feels like the author picked a conclusion that hasn't come to pass yet
The 'yet' is doing a good deal of heavy lifting, since the position that you're complaining about is the openly and explicitly stated goal of the large AI labs.
Your "super AI bullish" position is effectively a bet that the AI companies will fail.
It's interesting to see people work as hard as possible to build a world they insist they don't want, while claiming that the thing they're working on enabling can't possibly succeed.
Uh, I'm sorry, but I really struggle to understand this comment. It really sounds like you're pointing the finger at me for being pro-AI, when I'm really not.
Your "super AI bullish" position is effectively a bet that the AI companies will fail.
Where did I say I'm AI bullish? I called the company where I work at. Betting on failure is bearish and as I mentioned, I expect LLMs to plateau way below AGI, i.e. fail the labs' goal, so I'm a lot more bearish than bullish myself.
It's interesting to see people work as hard as possible to build a world they insist they don't want
It really feels like one of us is misunderstanding the other here. If you are talking in general, then sorry, I'm the one who misunderstood you and yeah, I agree, I don't like the altruism-washing I see from AI companies. If you did mean me, I think you're knocking on the wrong door. I don't use LLMs, except in a very limited capacity at work, let alone pay or advocate for them or do work for labs.
Now they talk about “prompts”, “harnesses”, “loops”. The discourse is narrower, shallower, and more repetitive. There’s only so many times I can hear about “agentic harnesses” before I lose my mind.
But this is the same “-30 IQ” discourse the author is criticizing. It’s a reductionist reality making the wrong extrapolation.
Look at radiology: the first of the physician class jobs that would be wiped out by NNs. What’s the job status there? Do we think radiologists are no longer looking at diag scans?
Coding as a craft is not anywhere near dead and once we wake up from this ai fever dream we are going to realize that humans continue to excel at many coding tasks.
I also think there is massive conflation going on between enterprise coding, which has always been slop with rare exception, and coding the craft. Enterprise coding was a means to an end: it motivated us to learn about software engineering.
Just because corpos are shoving LLMs down our throat and we can be brain dead doesn’t mean coding as craft is dead.
We are expected to know how they work and strongly encouraged to use it, but there is still the understanding that we "own" the code and are expected to understand and maintain it.
Is this any different from the usual corporate spectacle of nominating a scapegoat to sacrifice when things go wrong? Someone has to take the fall and not everyone is in a position to claim that the LLM they were using went "rogue".
But I also don't really feel like that's the direction we're heading?
I feel like that's the direction we're heading in (assuming we're not there already) depending on how one views LLM generated code. If one views it as a means to end, everything's fine because the benchmark is reduced to "does it work?" and ideas, which are hollow in isolation, are only judged by whether they work with what has been generated rather than whether it should be made to work at all in the first place.
Both LLVM and GCC now have LLM generated commits and since bootstrapping begins with C, one should wonder if our increasing dependency on these tools that we cannot replicate, recreate, and control (not without doing the same reprehensible things that OpenAI and Anthropic do) as independent communities is healthy. But as I said, depending on how one views generated code, this may not be a problem at all.
There’s your own private garden of code, which you can grow infinitely in all directions with the help of AI, but you never have to leave the garden and go to the bazaar to trade with people.
This is the main reason I am miserable. I know many people will view this as a personal failing, and yet it's not just me, a heck of a lot of people are suffering loneliness right now. We no longer need each other, it's just the same as how we can get all our material needs delivered to our door without saying a single word to another human.
This is another axis we may (and thus, many will due to how we are optimized for less immediately effortful actions) choose to not engage with each-other on. Hard not to see it as corporations mining (stealing) yet more of our crumbling networks of interdependence in order to transmute them into a network of dependence that terminates at capital holders.
I really enjoyed this article. I do tend to think of myself as someone intrinsically motivated but I do love to share my work and if there's no audience for it or chances to cooperate on software with like-minded individuals, that's a sad state indeed. If the future of FOSS is just random drive-by PRs generated by LLMs, how lonely a prison we've developed for ourselves...
This whole notion that you don't have to understand how anything works because AI can do it is beyond absurd. Anybody who's worked with this tools in any serious capacity can see that you can't let LLMs make any serious decisions in a project. You still have to understand algorithms, data structures, and architecture to make something work properly. The very nature of LLMs is such that they're just stochastic token generators, they fundamentally lack the ability to understand the requirements in a human sense. For author's conclusion to hold, there would need to be a qualitative leap in AI technology where it becomes genuinely intelligent. It might happen one day, but it's pretty clear we're not there yet, and LLMs alone are not the solution.
What I think is far more likely is that we'll see programming move towards formal methods and contracts. Humans will focus on high level concepts such as crafting the state machine that describes a particular workflow, then handing a formal spec to the LLM to fill in the implementation details.
Good article and it resonated with me. I still enjoy programming (even with AI, believe it or not!), but I feel I've lost a lot of the desire of sharing ideas, progress, and results. Looking back, all my social media posting was primarily about that and it has reduced to zero throughout this year. Because... why bother?
Luckily I'm attending BazelCon next week. Let's see if the social connection that comes with it renews my energy.
I don't think this is the full picture. I am a person who likes to talk about type systems and compilers because I can use them at work. These are two things that are increasingly hard to justify in a work context. Why care about nice type systems when the all-seeing LLM just gets everything right? I'm already seeing swaths of our lovely Haskell code just reimplemented in Python because, well why not? The end result is the same (to some outside observers, anyway), but it's miserable to those of who do care. So no, it's not just a job, at least not to me. Maybe I have been horribly misaligned and done a bad job of separating work from pleasure
nemin | 15 hours ago
Eh, I didn't really like this post. Chiefly, because it feels like the author picked a conclusion that hasn't come to pass yet ("AI will make all coding / intellectual work / blogging / paper writing stop completely") and taking it as fact argues backwards how terrible that'll be. And, like, yes, I don't want us to reach that world either.
But I also don't really feel like that's the direction we're heading? Even at my super AI-bullish company nobody is saying "do black-box development and let the AI do it all". We are expected to know how they work and strongly encouraged to use it, but there is still the understanding that we "own" the code and are expected to understand and maintain it. I understand that's an
n = 1anecdote, but it's not like OP's arguments are on any stronger foundations.With Math, I'm seeing a strong pushback as well, as mathematicians point out that LLMs may confirm the truthfulness of statements, but their raw proofs aren't nearly as useful as one constructed by a person.
"You write a blog post: who’s going to read it?" / "You design a revolutionary new programming language: who cares?" / "You write a library, and publish it on GitHub: who cares?" This question trio is what ultimately prompted me to write this comment, because I find it frankly a little ridiculous. Even if the current anti-AI sentiment wanes (which is very much not a foregone conclusion yet), people will always want to read cool writeups and hand-tinker on stuff.
This one I'll give to the author. I also experience it and it gives me such fatigue. The constant churn of "you need to use
BowelMovement-Sloptastic-420-69-QEDwith this particular fork of vllm.cpp or else you're missing out" or "BigModel 10.5 does 10 points higher on the benchmarks than LargeModel 7.7, but LargeModel 7.7 is much cheaper, so you should still use it"...It's like people are discussing Pokémon or something. And not even just the base game, but minmaxing some artificially kneecapped challenge.
dangroshev | 11 hours ago
I agree that it's a bit reductive and defeatist (and the bleak world it paints is more of a hypothetical at this point), but I think this is somewhat different from the author's argument:
There will always be people who like to read writeups, but it's not the same as a vibrant intellectual community of peers the author is talking about. Commodification of good-at-first-glance work removes quality signals and diminishes the prestige of doing good work.
In addition, any new useful signal comes with an incentive to fake it*, and that's only a prompt away.
Maybe what we'll see is tighter and more closed professional groups and an increase of the value of in-network reputation, which is much more difficult to fake. This would be sad, because our industry became a social lift for so many people precisely because it had less emphasis on networks and more on contributions than other highly-paid industries.
*: as an example, "rigorous benchmarking vs the alternatives" used to be a useful signal. Look at how slop benchmarking got weaponised by a certain racist Danish-American entrepreneur lately.
nemin | 11 hours ago
Fair, however, I still don't think is necessarily a given just because LLMs are a thing now. For one, LLM generated texts are ghoulish and even people otherwise open to the tech recoil in disgust. Unless they somehow figure out how to autonomously give the models a voice that doesn't have such a visceral effect (I'm doubtful!), that means production of content will always have a human element.
As for consumption, I just don't see why people would completely stop intellectually engaging with things. LLMs are very impressive in many aspects, but they aren't curious nor innovative. You can let them lose on problems you might not be able to solve yourself, but you still have to have the knowledge about these problems' existence and the gist of them. Then there are all the people who use LLMs as more "research partners" / "assistants", while still taking charge. I don't expect them to stop wanting to improve themselves in the field either.
I suppose the argument boils down to whether you expect LLMs to improve to an extent where they'll "human better" than actual humans. I personally don't. I expect there to be a theoretical limit and I expect said limit to be well below AGI.
Funny, but I expect this less in professional circles and more in hobbyist communities. The more AI can ragebait and act like a nuisance, the more communities are incentivized to vet their new members, which could result in more vouching systems ala Lobsters.
In professional contexts, I think there is already a sort of inherent filtering. You can make it as a shoddy programmer, but if you want to get into specific branches and certain companies, you must be able to rise above the rest. And an LLM won't necessarily let you do that (because you either don't have access to it or the domain knowledge is simply so complex that no generalist model can hope to understand it like a person who spent quite a bit of time on it).
koreth | 8 hours ago
On your last point: Or because the people at the branches and companies you want to get into are also using LLM tools effectively, so using an LLM yourself becomes just table stakes, not a competitive advantage.
orib | 5 hours ago
The 'yet' is doing a good deal of heavy lifting, since the position that you're complaining about is the openly and explicitly stated goal of the large AI labs.
Your "super AI bullish" position is effectively a bet that the AI companies will fail.
It's interesting to see people work as hard as possible to build a world they insist they don't want, while claiming that the thing they're working on enabling can't possibly succeed.
nemin | 4 hours ago
Uh, I'm sorry, but I really struggle to understand this comment. It really sounds like you're pointing the finger at me for being pro-AI, when I'm really not.
Where did I say I'm AI bullish? I called the company where I work at. Betting on failure is bearish and as I mentioned, I expect LLMs to plateau way below AGI, i.e. fail the labs' goal, so I'm a lot more bearish than bullish myself.
It really feels like one of us is misunderstanding the other here. If you are talking in general, then sorry, I'm the one who misunderstood you and yeah, I agree, I don't like the altruism-washing I see from AI companies. If you did mean me, I think you're knocking on the wrong door. I don't use LLMs, except in a very limited capacity at work, let alone pay or advocate for them or do work for labs.
erock | 18 minutes ago
Agreed and to pile on:
But this is the same “-30 IQ” discourse the author is criticizing. It’s a reductionist reality making the wrong extrapolation.
Look at radiology: the first of the physician class jobs that would be wiped out by NNs. What’s the job status there? Do we think radiologists are no longer looking at diag scans?
Coding as a craft is not anywhere near dead and once we wake up from this ai fever dream we are going to realize that humans continue to excel at many coding tasks.
I also think there is massive conflation going on between enterprise coding, which has always been slop with rare exception, and coding the craft. Enterprise coding was a means to an end: it motivated us to learn about software engineering.
Just because corpos are shoving LLMs down our throat and we can be brain dead doesn’t mean coding as craft is dead.
ayushnix | 8 hours ago
Is this any different from the usual corporate spectacle of nominating a scapegoat to sacrifice when things go wrong? Someone has to take the fall and not everyone is in a position to claim that the LLM they were using went "rogue".
I feel like that's the direction we're heading in (assuming we're not there already) depending on how one views LLM generated code. If one views it as a means to end, everything's fine because the benchmark is reduced to "does it work?" and ideas, which are hollow in isolation, are only judged by whether they work with what has been generated rather than whether it should be made to work at all in the first place.
Both LLVM and GCC now have LLM generated commits and since bootstrapping begins with C, one should wonder if our increasing dependency on these tools that we cannot replicate, recreate, and control (not without doing the same reprehensible things that OpenAI and Anthropic do) as independent communities is healthy. But as I said, depending on how one views generated code, this may not be a problem at all.
elliotmorris | 11 hours ago
This is the main reason I am miserable. I know many people will view this as a personal failing, and yet it's not just me, a heck of a lot of people are suffering loneliness right now. We no longer need each other, it's just the same as how we can get all our material needs delivered to our door without saying a single word to another human.
This is another axis we may (and thus, many will due to how we are optimized for less immediately effortful actions) choose to not engage with each-other on. Hard not to see it as corporations mining (stealing) yet more of our crumbling networks of interdependence in order to transmute them into a network of dependence that terminates at capital holders.
rattboi | 15 hours ago
I really enjoyed this article. I do tend to think of myself as someone intrinsically motivated but I do love to share my work and if there's no audience for it or chances to cooperate on software with like-minded individuals, that's a sad state indeed. If the future of FOSS is just random drive-by PRs generated by LLMs, how lonely a prison we've developed for ourselves...
Yogthos | 9 hours ago
This whole notion that you don't have to understand how anything works because AI can do it is beyond absurd. Anybody who's worked with this tools in any serious capacity can see that you can't let LLMs make any serious decisions in a project. You still have to understand algorithms, data structures, and architecture to make something work properly. The very nature of LLMs is such that they're just stochastic token generators, they fundamentally lack the ability to understand the requirements in a human sense. For author's conclusion to hold, there would need to be a qualitative leap in AI technology where it becomes genuinely intelligent. It might happen one day, but it's pretty clear we're not there yet, and LLMs alone are not the solution.
What I think is far more likely is that we'll see programming move towards formal methods and contracts. Humans will focus on high level concepts such as crafting the state machine that describes a particular workflow, then handing a formal spec to the LLM to fill in the implementation details.
jmmv | 14 hours ago
Good article and it resonated with me. I still enjoy programming (even with AI, believe it or not!), but I feel I've lost a lot of the desire of sharing ideas, progress, and results. Looking back, all my social media posting was primarily about that and it has reduced to zero throughout this year. Because... why bother?
Luckily I'm attending BazelCon next week. Let's see if the social connection that comes with it renews my energy.
singpolyma | 10 hours ago
I think it has been a long time since the majority of software developers were interested in CS topics or were even particularly nerdy.
This is ok. It's just a job after all. A job anyone can do if they want to, and (probably because of pay) many have wanted to.
The people who liked to talk about type systems and compilers still like those things.
ocharles | 7 hours ago
I don't think this is the full picture. I am a person who likes to talk about type systems and compilers because I can use them at work. These are two things that are increasingly hard to justify in a work context. Why care about nice type systems when the all-seeing LLM just gets everything right? I'm already seeing swaths of our lovely Haskell code just reimplemented in Python because, well why not? The end result is the same (to some outside observers, anyway), but it's miserable to those of who do care. So no, it's not just a job, at least not to me. Maybe I have been horribly misaligned and done a bad job of separating work from pleasure