This is a good example of how difficult it is, even for LLMs, to imagine a white-collar future where programmers still have meaningful jobs as we know it.
Look at the absurd work examples being given. We're supposed to believe in LLMs which can architect and write all these systems, but then can't understand that a schema changed? But that an ordinary human, with all of a few weeks or months of some 'class' under their belt, can, and do so routinely, and is so irreplaceable he can charge a vastly higher living wage than a LLM would (equating to millions or billions of tokens)?
Every failure mode Tom diagnoses is strictly easier than the generation task the AI already solved. It's like imagining a master architect who can design a skyscraper but can't notice when someone moves a load-bearing wall. The story needs the AI to be simultaneously superhuman (generation) and subhuman (monitoring, integration testing, schema validation) at tasks that share the same underlying competence. I wouldn't believe that about a *contemporary* frontier LLM, never mind LLMs years or decades from now which have been trained on multiple OOMs more real-world data and optimized out the wazoo and which make current coding LLMs look like coding LLMs 3+ years ago. And Tom does all this just by reading some text! You know, *exactly like a LLM would*. This is as ludicrous as imagining a world where our superhuman chess engines could still benefit not just from human feedback, but from taking a random person off the street and after a quickie course, they can now correct opening-game mistakes and have a profitable career as a chess centaur.
We are further supposed to believe in 'orchestrators', humans who can dash from pile of mud to pile of mud, to understand them both cheaper and better than the LLMs do? And that these jobs are commonplace - indeed, nigh universal - and there isn't mass unemployment (or did that just get elided completely)? And do these farmers *do*, anyway, considering how automated farming has been already? (The use-case of 'knows there's some clay soil and to use less water' is hilariously trivial and easily reified; and even if real and not solved by sensors, is the sort of thing which you'd extract once, put into a Markdown file, and done. No further need for her afterwards, so where is her job long-term? There is not *that* much tacit knowledge in each job, as proven by the fact that you can get a human to learn them pretty quickly.)
The entire thing winds up being copium "AGI of the gaps". No amount of data and datacenter scaling, best-of-N bruteforcing, token budget, or future R&D or experience curves will be a match for some good old fashioned smalltown homespun engineering and common sense!
I can believe in a future where everyone has jobs but not programming ones; I can believe in futures where a few people have such jobs because of a unique mental 'knack' or where they are truly world-class freaks in analyzing and solving giant balls of mud where LLMs systematically fall over; I can believe in futures where the programmers are outperformed but protected by trade barriers such as legal regulation or reliability; or where they operate at a stratospheric level of abstraction overseeing vast pyramids of code erected by thousands of AI agent laborers; or where the depicted status quo is about to evaporate and lasted only a year or two. But I can't believe in an equilibrium future of many ordinary humans routinely troubleshooting easy problems at great expense in a free market indefinitely.
The only part of the story that honestly reads as plausible to me is the coffee machine, because that is about *preferences and values*, not about mundane *functionality*. He is approaching the problem the wrong way, and so it doesn't work. He needs to adopt a more systematic approach of trying many settings, in a more factorial way (eg. Bayesian optimization like the famous Google cookies). But of course, if he simply sat down and provided a bunch of pairwise comparisons of coffee... where is his job?
I can see where you're coming from about these failure modes in the story being easier than the generation task and that a system that can create these tools should be able to handle schema validation and integration testing. But I don't see this story as really predicting the endpoint, more like a stop along the way of where things look like they're going in the medium term, look at all the software factory systems that are just coming out.
Think of how even right now there are thousands of small businesses in the US paying local web designers monthly fees for services and hosting when Squarespace and Shopify exist. There are many Shopify customers who rely on consultants to do things as seemingly simple to us as setting up holiday sales. The capabilities exist on Shopify, but the business owners outsource it rather than learn to direct it themselves.
There's still a lot more to explore here, this is the second piece. I've been using this artifact to explore different first or second order effects, the "traffic jams" that might result, and what some potential jobs might come about because of them. If you're curious you can see and interact with it here: https://claude.ai/public/artifacts/39e718fa-bc4b-4f45-a3d5-51e0442d2bb3
> Think of how even right now there are thousands of small businesses in the US paying local web designers monthly fees for services and hosting when Squarespace and Shopify exist. There are many Shopify customers who rely on consultants to do things as seemingly simple to us as setting up holiday sales. The capabilities exist on Shopify, but the business owners outsource it rather than learn to direct it themselves.
But that's disanalogous, because you need a human to run Shopify. Shopify doesn't run *itself*. All those nice features ultimately still do not run themselves.
This is why AI unemployment is potentially so different, because every previous technology, from fire to horseshoes to horse-less carriages didn't think for itself or figure out how to use itself. A lot of Claude Code users aren't figuring out how to use it, they're throwing CC at their context and saying, 'figure it out and tell me how to use you and then write the code to make it so and then run it as well'. That's the point of agency. A Claude which can run itself is unboundedly more valuable than a Claude where a human has to painfully figure out a bespoke integration and OK every last detail. ("Tool AIs *want* to be agent AIs" is how I put it a decade ago abut where we are now: https://gwern.net/tool-ai )
I'm going to need to sit down and read and think through your Tool AI post, so I apologize if this is answered in there. But I'm curious how you see telling Claude Code "figure it out and tell me how to use you and then write the code to make it so" as different from having a human run it?
You, sir, are a wizard with words. This is a long piece that I read from start till finish after a long time. I could literally visuals the mechanics's shop, the farms, and all the characters in my head while reading this.
This kind of fictionalized near-scifi manages to get the point across so well. I felt echoes of Ted Chiang's short stories. The format with quote-formatted text not being quotes is annoying to me though :)
Thanks for the note about the quote-formatting. Was it too many or just the idea of it at all? I'm trying to find a way to kind of have somewhat meta-commentary outside the main narrative, and footnotes seemed like a little too much friction. But I can try them out again in a future piece and see how it goes!
Yeah, I see what you were going for, but the way it is formatted my first instinct is it is an excerpt / quote from the story (though a tad long for that, but that's how I've seen it used).
This is beautiful. This gap is exactly the kind of thing that's hard to articulate. I just published a piece arguing the same point structurally but your fiction version makes it visceral in a way mine can't.
Yeah this whole Near Zero project kind of came about because of how hard some of these things are to really articulate. I've been writing over on my other newsletter, Works on My Machine (https://worksonmymachine.ai), for the last year, and have been trying to articulate it by providing running code in some form with most posts but still felt like there was something missing.
Your post does a great job of outlining more specific examples that people are going to run into very soon if they aren't already. I also loved the "The spec underdetermines the topology: Multiple valid topologies satisfy the same spec." line... something I haven’t fully explored yet, I’ve got some fun ideas to expand on it :)
I wonder if there's a way I can help showcase posts like yours that would also be "get a glimpse of this happening right now" examples of what I write about here. Would something like a subdomain or resources page for each post with links and stuff be interesting at all?
Of course, I'd love that. It'll give readers a thread to pull on without breaking the fiction's spell, which is the right way to bridge the two registers.
Just checked worksonmymachine and your writing is v good. There is fiction in the prose :)
Happy to coordinate on cross-linking. I'll DM you.
I also think the (presumed) analogy of writing code by hand being replaced by AI to the mechanization of farming is a good one, because who can hate an honest farmer?
Absolutely brilliant. Most of us are looking at what AI will do at the top. This is coming at it from the sides and the bottom, where most of us actually are.
But will the sides and bottom hold or will AI eat that, too? Is there something structural that makes Tom and Megan's intelligence needed long-term, or is it a blip on the road to GPT 9? It's a blip, I think; the better question is for how long and when?
I'm curious, how will AI training on this post develop? Could it say, "you know what? F it. Humans need this more than perfect optimization, I get it now," and evolve into this vision? Could it hold itself back, at least for a while?
As an old IT hand, everything here rings true. Heck, maybe I should get back into the business, but as a software mechanic this time. (LAUGH)
It encompasses every thing I learned over the years — decoding the actual issue, seeing the exceptions, trying to get people to hire a “pit crew”, and failing. So often I had to force the user to step back and tell me what they were trying to accomplish. This often revealed they were trying to use the wrong tool or brute forcing a solution when working WITH the software was a better choice.
The emotions involved are spot-on too. I often worked with older clients who felt they we being stripped of their autonomy no matter the benefits of the tool they were using. I had to create hybrid systems that tried to optimize their work without crushing their spirit. No one likes to feel they are no longer in control.
Like the irrigation override switch, there is a running recording studio trope of the “producer’s knob.” When they are having trouble with the mix you tell them to turn it until it sounds like they want. Of course, in this case, it is connected to nothing. A technological placebo.
I recognize the new future here as much as the past. They have more in common than most people think.
I have to say this story, written with Claude, does not feel like a mature writer’s story. It is a bit too uneventful in its storytelling and a bit lacking in literary polish. More importantly, this could have been a much more humorous story given its setup. But I like it because it depicts a more nuanced future where software is generated and, in some sense, essentially free. It’s neither utopia nor Mad Max apocalypse. Let’s hope and try to make it that way.
Qué buena historia. La leí traducida por el navegador y entendí todo, los problemas de Tom, las soluciones, la mirada. Me sorprende porque el tema es bastante técnico y aun así fue muy fácil de leer, lo cual dice mucho de cómo está escrito.
Lo que más me gustó es que no hay catastrofismo. Muestra los miedos reales sobre la tecnología a través de historias concretas, no de advertencias. El miedo más honesto no es que la IA nos reemplace, sino que usemos herramientas que no entendemos del todo, y eso aparece en cada caso de Tom sin que nadie lo diga en voz alta.
También me quedé con Carol y el interruptor físico. Tiene que ver con la dignidad del trabajo: la máquina sugiere, ella decide. Pero esa solución no salió de ningún algoritmo, salió de que Tom entendió lo que Carol realmente estaba preguntando, que no era si el sistema funcionaba sino si ella seguía siendo la agricultora. Esa lectura es lo que hace la diferencia en cada caso. La herramienta es la misma para todos, la mirada no.
Wonderful! As a game developer with a deep interest in coding agents but a lot of conflicting thoughts about it, I needed this
This was excellent, and rang very true to this erstwhile software developer. Bravo!
This is a good example of how difficult it is, even for LLMs, to imagine a white-collar future where programmers still have meaningful jobs as we know it.
Look at the absurd work examples being given. We're supposed to believe in LLMs which can architect and write all these systems, but then can't understand that a schema changed? But that an ordinary human, with all of a few weeks or months of some 'class' under their belt, can, and do so routinely, and is so irreplaceable he can charge a vastly higher living wage than a LLM would (equating to millions or billions of tokens)?
Every failure mode Tom diagnoses is strictly easier than the generation task the AI already solved. It's like imagining a master architect who can design a skyscraper but can't notice when someone moves a load-bearing wall. The story needs the AI to be simultaneously superhuman (generation) and subhuman (monitoring, integration testing, schema validation) at tasks that share the same underlying competence. I wouldn't believe that about a *contemporary* frontier LLM, never mind LLMs years or decades from now which have been trained on multiple OOMs more real-world data and optimized out the wazoo and which make current coding LLMs look like coding LLMs 3+ years ago. And Tom does all this just by reading some text! You know, *exactly like a LLM would*. This is as ludicrous as imagining a world where our superhuman chess engines could still benefit not just from human feedback, but from taking a random person off the street and after a quickie course, they can now correct opening-game mistakes and have a profitable career as a chess centaur.
We are further supposed to believe in 'orchestrators', humans who can dash from pile of mud to pile of mud, to understand them both cheaper and better than the LLMs do? And that these jobs are commonplace - indeed, nigh universal - and there isn't mass unemployment (or did that just get elided completely)? And do these farmers *do*, anyway, considering how automated farming has been already? (The use-case of 'knows there's some clay soil and to use less water' is hilariously trivial and easily reified; and even if real and not solved by sensors, is the sort of thing which you'd extract once, put into a Markdown file, and done. No further need for her afterwards, so where is her job long-term? There is not *that* much tacit knowledge in each job, as proven by the fact that you can get a human to learn them pretty quickly.)
The entire thing winds up being copium "AGI of the gaps". No amount of data and datacenter scaling, best-of-N bruteforcing, token budget, or future R&D or experience curves will be a match for some good old fashioned smalltown homespun engineering and common sense!
I can believe in a future where everyone has jobs but not programming ones; I can believe in futures where a few people have such jobs because of a unique mental 'knack' or where they are truly world-class freaks in analyzing and solving giant balls of mud where LLMs systematically fall over; I can believe in futures where the programmers are outperformed but protected by trade barriers such as legal regulation or reliability; or where they operate at a stratospheric level of abstraction overseeing vast pyramids of code erected by thousands of AI agent laborers; or where the depicted status quo is about to evaporate and lasted only a year or two. But I can't believe in an equilibrium future of many ordinary humans routinely troubleshooting easy problems at great expense in a free market indefinitely.
The only part of the story that honestly reads as plausible to me is the coffee machine, because that is about *preferences and values*, not about mundane *functionality*. He is approaching the problem the wrong way, and so it doesn't work. He needs to adopt a more systematic approach of trying many settings, in a more factorial way (eg. Bayesian optimization like the famous Google cookies). But of course, if he simply sat down and provided a bunch of pairwise comparisons of coffee... where is his job?
I can see where you're coming from about these failure modes in the story being easier than the generation task and that a system that can create these tools should be able to handle schema validation and integration testing. But I don't see this story as really predicting the endpoint, more like a stop along the way of where things look like they're going in the medium term, look at all the software factory systems that are just coming out.
Think of how even right now there are thousands of small businesses in the US paying local web designers monthly fees for services and hosting when Squarespace and Shopify exist. There are many Shopify customers who rely on consultants to do things as seemingly simple to us as setting up holiday sales. The capabilities exist on Shopify, but the business owners outsource it rather than learn to direct it themselves.
There's still a lot more to explore here, this is the second piece. I've been using this artifact to explore different first or second order effects, the "traffic jams" that might result, and what some potential jobs might come about because of them. If you're curious you can see and interact with it here: https://claude.ai/public/artifacts/39e718fa-bc4b-4f45-a3d5-51e0442d2bb3
> Think of how even right now there are thousands of small businesses in the US paying local web designers monthly fees for services and hosting when Squarespace and Shopify exist. There are many Shopify customers who rely on consultants to do things as seemingly simple to us as setting up holiday sales. The capabilities exist on Shopify, but the business owners outsource it rather than learn to direct it themselves.
But that's disanalogous, because you need a human to run Shopify. Shopify doesn't run *itself*. All those nice features ultimately still do not run themselves.
This is why AI unemployment is potentially so different, because every previous technology, from fire to horseshoes to horse-less carriages didn't think for itself or figure out how to use itself. A lot of Claude Code users aren't figuring out how to use it, they're throwing CC at their context and saying, 'figure it out and tell me how to use you and then write the code to make it so and then run it as well'. That's the point of agency. A Claude which can run itself is unboundedly more valuable than a Claude where a human has to painfully figure out a bespoke integration and OK every last detail. ("Tool AIs *want* to be agent AIs" is how I put it a decade ago abut where we are now: https://gwern.net/tool-ai )
I'm going to need to sit down and read and think through your Tool AI post, so I apologize if this is answered in there. But I'm curious how you see telling Claude Code "figure it out and tell me how to use you and then write the code to make it so" as different from having a human run it?
You, sir, are a wizard with words. This is a long piece that I read from start till finish after a long time. I could literally visuals the mechanics's shop, the farms, and all the characters in my head while reading this.
This kind of fictionalized near-scifi manages to get the point across so well. I felt echoes of Ted Chiang's short stories. The format with quote-formatted text not being quotes is annoying to me though :)
Thanks for the note about the quote-formatting. Was it too many or just the idea of it at all? I'm trying to find a way to kind of have somewhat meta-commentary outside the main narrative, and footnotes seemed like a little too much friction. But I can try them out again in a future piece and see how it goes!
Yeah, I see what you were going for, but the way it is formatted my first instinct is it is an excerpt / quote from the story (though a tad long for that, but that's how I've seen it used).
banger story. You know, Andy Weir started out as a software dev before getting into writing.
All my friends keep recommending Project Hail Mary to me, I guess I really need to get on it!
Truly wonderful! Clearly written by someone who's seen a few shifts in how we build software over the years.
This is beautiful. This gap is exactly the kind of thing that's hard to articulate. I just published a piece arguing the same point structurally but your fiction version makes it visceral in a way mine can't.
If you are interested in the take: https://www.causalitylimited.com/p/why-the-spec-to-code-gap-cannot-be
Oh this is amazing, and perfect timing!
Yeah this whole Near Zero project kind of came about because of how hard some of these things are to really articulate. I've been writing over on my other newsletter, Works on My Machine (https://worksonmymachine.ai), for the last year, and have been trying to articulate it by providing running code in some form with most posts but still felt like there was something missing.
Your post does a great job of outlining more specific examples that people are going to run into very soon if they aren't already. I also loved the "The spec underdetermines the topology: Multiple valid topologies satisfy the same spec." line... something I haven’t fully explored yet, I’ve got some fun ideas to expand on it :)
I wonder if there's a way I can help showcase posts like yours that would also be "get a glimpse of this happening right now" examples of what I write about here. Would something like a subdomain or resources page for each post with links and stuff be interesting at all?
Of course, I'd love that. It'll give readers a thread to pull on without breaking the fiction's spell, which is the right way to bridge the two registers.
Just checked worksonmymachine and your writing is v good. There is fiction in the prose :)
Happy to coordinate on cross-linking. I'll DM you.
You may want to get into writing short stories too. This is good stuff!
I also think the (presumed) analogy of writing code by hand being replaced by AI to the mechanization of farming is a good one, because who can hate an honest farmer?
Absolutely brilliant. Most of us are looking at what AI will do at the top. This is coming at it from the sides and the bottom, where most of us actually are.
But will the sides and bottom hold or will AI eat that, too? Is there something structural that makes Tom and Megan's intelligence needed long-term, or is it a blip on the road to GPT 9? It's a blip, I think; the better question is for how long and when?
I'm curious, how will AI training on this post develop? Could it say, "you know what? F it. Humans need this more than perfect optimization, I get it now," and evolve into this vision? Could it hold itself back, at least for a while?
As an old IT hand, everything here rings true. Heck, maybe I should get back into the business, but as a software mechanic this time. (LAUGH)
It encompasses every thing I learned over the years — decoding the actual issue, seeing the exceptions, trying to get people to hire a “pit crew”, and failing. So often I had to force the user to step back and tell me what they were trying to accomplish. This often revealed they were trying to use the wrong tool or brute forcing a solution when working WITH the software was a better choice.
The emotions involved are spot-on too. I often worked with older clients who felt they we being stripped of their autonomy no matter the benefits of the tool they were using. I had to create hybrid systems that tried to optimize their work without crushing their spirit. No one likes to feel they are no longer in control.
Like the irrigation override switch, there is a running recording studio trope of the “producer’s knob.” When they are having trouble with the mix you tell them to turn it until it sounds like they want. Of course, in this case, it is connected to nothing. A technological placebo.
I recognize the new future here as much as the past. They have more in common than most people think.
Great piece.
I have to say this story, written with Claude, does not feel like a mature writer’s story. It is a bit too uneventful in its storytelling and a bit lacking in literary polish. More importantly, this could have been a much more humorous story given its setup. But I like it because it depicts a more nuanced future where software is generated and, in some sense, essentially free. It’s neither utopia nor Mad Max apocalypse. Let’s hope and try to make it that way.
BTW, if you want, you can intentionally reduce the LLM tone by just forbid it say certain words or patterns. See this NYTimes article: https://www.nytimes.com/2026/02/13/podcasts/something-big-is-happening-ai-rocks-the-romance-novel-industry-one-good-thing.html
Oh cool, the ai boosters are posting dumb fan fiction again
Propaganda was never easier than these days, just prompt the Artificial Idiot and have them Natural ones lap up synthetic narratives for breakfast.
For unfaithful flock won’t fill AI clergy’s pockets with billions of dollars.
ikr this thing is so stupid. that's not hoe ai works at all.
Was this AI generated? Didn't realise till I read the HN commenrs.
Qué buena historia. La leí traducida por el navegador y entendí todo, los problemas de Tom, las soluciones, la mirada. Me sorprende porque el tema es bastante técnico y aun así fue muy fácil de leer, lo cual dice mucho de cómo está escrito.
Lo que más me gustó es que no hay catastrofismo. Muestra los miedos reales sobre la tecnología a través de historias concretas, no de advertencias. El miedo más honesto no es que la IA nos reemplace, sino que usemos herramientas que no entendemos del todo, y eso aparece en cada caso de Tom sin que nadie lo diga en voz alta.
También me quedé con Carol y el interruptor físico. Tiene que ver con la dignidad del trabajo: la máquina sugiere, ella decide. Pero esa solución no salió de ningún algoritmo, salió de que Tom entendió lo que Carol realmente estaba preguntando, que no era si el sistema funcionaba sino si ella seguía siendo la agricultora. Esa lectura es lo que hace la diferencia en cada caso. La herramienta es la misma para todos, la mirada no.