5 ms·

Generative AI Is an Engineering Disaster

116 points by latexr 3mo ago

latexr 3mo ago
https://archive.ph/20260715223901/https://www.theatlantic.com/technology/2026/07/generative-ai-engineering-disaster/687901/ https://archive.ph/20260715223901/https://www.theatlantic.co...
maxcb 3mo ago
Comparing how traditional tech companies scale to how frontier AI companies scale doesn't seem very fair to me. AI is still relatively new territory, in comparison traditional software companies benefit from decades of investment in cloud infrastructure, networking, and hardware. I'm certain those technologies also went through periods of heavy investment before they became scalable. The author also talks about advances like smartphones as though they arrived without significant challenges or trade-offs. Every major technological shift has come with some cost, e.g environmental, shortages, investments. I'd argue most things are difficult to scale in the beginning. Just because today's LLMs are expensive and resource intensive doesn't mean the technology is fundamentally flawed, it may simply mean the current approach isn't the one we'll end up with. But naturally, someone working on another approach would have something to criticise about others. Finally, I'm not convinced by the claim that we're "stuck" with LLMs just because they're heavily marketed. They create value for many people, which is why they've been adopted so quickly. As the author points out, investors care about economics. I'm sure people would listen if someone developed a cheaper/more sustainable technology that delivered the same value.
- pixl97 3mo ago
  Right, if you came out with a technology that worked just as well as LLMs at half the power people/companies would jump at it from the absolutely massive savings in hardware and power. AI companies are looking at every other method they can, and inventing new ones, none of them have worked as well as the transformer so far.
- watwut 3mo ago
  > They create value for many people, which is why they've been adopted so quickly. A lot of that adoption is completely useless crap tho. Like ai in vacuum cleaner that ads absolutely nothing useful to anything. AI buttons intentionally at places I randomly click at, so that I am forced to open it. Google search that defaults to ai, so that we have to use it after then nerfed real search. There are useful usages of LLMs. But huge bulk of the adoption is companies realizing they wont get investors money if they dont add ai button, it does not have to be useful.
  - SoftTalker 3mo ago
    Yeah there's a huge part of it that is "the investors want tulips, so we will plant tulips"
    - tharmas 3mo ago
      Thanks. You're comment made me laugh.
  - bryanlarsen 3mo ago
    Which happens for every new technology. Blue LED's in everything, capacitive buttons on everything, car buttons replaced with ipads, et cetera. Some of the random placements will prove out useful but most won't and will disappear.
    - watwut 3mo ago
      Frankly, no, totally not. And not nearly to this extend. I have seen enough of "new technologies" to know that it was not as crazy at it is now with AI. The hysterical doom trolling, the hype, the tokenmaxxing, the personification of it, all of that is off the charts compared to anything I have seen in the past.
    - doctorwho42 3mo ago
      Yes, but corporations didn't bet the Capex of the entire GDP of the world on it
- JackSlateur 3mo ago
  "They create value for many people" They create value for many people because these people are not paying the cost of the tool;
- tedggh 3mo ago
  I think LLMs, with only a few exceptions, have mostly created the illusion of value. It’s undeniable I can write code, troubleshoot and document much faster than in the pre LLM era, but putting a dollar value to that it’s not particularly easy. One group of people being 100x faster at one task doesn’t mean the organization as a whole runs at 100x the speed, in fact in most cases where the bottle neck is somewhere else, the productivity gains of the 100x group have little to no impact, or even a negative impact in some cases due to token overspending.
  - naveen99 3mo ago
    How about sole proprietors?
    - andrekandre 3mo ago
      that would be interesting to see with a large longitudinal study any new technology also introduces new bottlenecks, so i wonder if its a 2 steps forwards 1 step backwards (or vice versa) situation (e.g llm summarized my email conversation to be more clear but left out a key detail; now i have to rework my proposal and i lost 1 week of income etc)
      - naveen99 3mo ago
        Resnet’s solve the regression problem.
  - unknownfuture 3mo ago
    Yup. We're just all re-learning Amdahl's Law.
andy_ppp 3mo ago
The ultra wealthy will out compete you for the resources you need if we don’t start taxing them has been a discussion from the left for quite some time. It seems to keep coming true in new and disastrous ways, we are going to see more and more extreme concentrations of capital and power distort things and pretty soon it won’t just be middle and poor people who cannot afford anything, the whole value chain is being redesigned around things billionaires want and what could be better for them than agents who replace all the white collar workers. You might think you’re safe from this, I don’t. It’s worth remembering that during the industrial revolution in Britain, the fastest growing country the world had ever seen, most people were in abject poverty. This tech revolution might end up being worse.
- efields 3mo ago
  Yup. This is a policy problem.
- throwaway323929 3mo ago
  Instead of just rejecting the future like a Roman fatalist because you've decided that it's just going to be someone else controlling you, maybe a better plan is to adopt emerging technology when it comes into play (like for example when the Internet came into play) and use it to start your own company, or make yourself more productive, or disrupt an industry that desperately needs it. I shouldn't have to explain the benefits of productivity improvements and computer technology to people on Hacker News, but the place for some reason has been hijacked by neo luddites from Reddit that apparently have nothing better to do than troll about AI on a web site that is literally dedicated to an industry they don't like. Sam Altman literally used to run Y Combinator. Why not use that energy you're wasting to adopt emerging technology instead of bashing it? There are places that recognized manufacturing was going to go away, transitioned a service economy, and did well economically, and places that ignored reality and are now in disrepair. Now's a great time to decide if you want to be Seattle or Cleveland. Don't fight the future. And don't think it's a billionaires fault when certain places are more successful then others because they didn't ban AI and datacenters while other places squandered the opportunity because they didn't confront the pitchfork mob driven into a frenzy by yellow journalism.
vanuatu 3mo ago
i dont think the comparison to traditional vc saas is very good yes, in traditional venture you want cost per marginal user to decrease and leverage your platform at scale but improving llms shifts the frontier of their capability and unlocks entirely new use cases. so far, every mega training run has resulted in a model that has paid itself off profitably fully loaded. perhaps the TAM of intelligence has no ceiling? not to say that we shouldnt be investing in efficient models, but the efficiency comes after we create another mega shoggoth that we can make more efficient
- simianwords 3mo ago
  you think companies are not focusing on efficiency but in reality they are both almost fully correlated. if you make a model more efficient, you get to make it do more reasoning and we know that the upper limit for reasoning is very very high. we have real proof - GPT 5.6 Luna is 50% cheaper than 5.5. And this keeps repeating.
scotty79 3mo ago
I can't imagine a successful civilization where non-computing energy usage isn't a rounding error.
- nancyminusone 3mo ago
  I can. AI says it will solve all problems, like curing cancer. Suppose it does. Now there are no more problems and no reason to run the AIs, at which time computing energy will fall rapidly. The alternative is that there will always be problems, and we could spend as much as possible trying to compute solutions, but they will not be found.
  - inigyou 3mo ago
    Maybe we'll all upload ourselves to the Dyson sphere, so there will be nothing left but computation.
    - bigbuppo 3mo ago
      42.
  - scotty79 3mo ago
    > Now there are no more problems and no reason to run the AIs, Are you sure we are talking about the same species?
    - JackSlateur 3mo ago
      Creating fake problems is not a caracteristic I would give to "a successful civilization" :)
      - scotty79 3mo ago
        Each solution generates new problems. That's the general physics of intelligence in our universe. Also label of "fake" is completely arbitrary. Because someone might call cancer a fake problem, because people shouldn't live too long and they need a reason to die.
srhtftw 3mo ago
Approaching an AI version of https://hitchhikers.fandom.com/wiki/Shoe_Event_Horizon https://hitchhikers.fandom.com/wiki/Shoe_Event_Horizon
rglover 3mo ago
It is, in some ways, but no matter your opinion, AI in engineering will continue unabated. Wrote about this recently (from a vibe coding perspective) [1]. This applies to corporations, too, as they're betting their futures on fewer engineers, more AI. [1] https://graybearding.bearblog.dev/they-got-something-on-the-screen/ https://graybearding.bearblog.dev/they-got-something-on-the-...
erelong 3mo ago
https://archive.ph/xo7RG https://archive.ph/xo7RG
bilater 3mo ago
https://x.com/PalmerLuckey/status/2077216959930728889 https://x.com/PalmerLuckey/status/2077216959930728889 Ya'll are not ready for AI to just keep getting better and better. The slop-midwit-doomerism you are being fed as part of some last ditch copium that your SE identity matters and is going to be needed prevents you from seeing what is inevitable. You can't beat em. Try joining em.
- selicos 3mo ago
  The month to month changes and improvements across free tools is noticeable. A number of skills SE and other roles are built on (or people started with) like more simple scripting and automation work are reachable by a BA with proper requirements, some background with development/coding, and access to AI tools. It's a continuation of both low code development tools and the ongoing compaction of development groups through DevOps. AI is enabling (vibe) code development by non devs. It's enabling a BA or even direct business rep to take requirements into a prototype or example. LLMs are out of the box and will continue developing. How/If you use them is up to you.
- simianwords 3mo ago
  [flagged]
devin 3mo ago
The rentier class needs a very strong correction. A lot of the conversation focused on what AI is and isn't is often just a proxy for frustration at the structural problems in our society which are in large part due to the inability for hard-working people to live good quality lives because of the incredible greed of the few.
- Varelion 3mo ago
  WOAH! Communism in my homefront? That's an NSPM-7 violation there, buddy.
simianwords 3mo ago
[flagged]
- unknownfuture 3mo ago
  > This is obviously untrue so either the author is knowingly lying or is plain incompetent. We had LLMs not being able to do simple high school mathematics a year back, now it is solving open problems in mathematics. Fields medalists in physics and mathematics are using it on a daily basis. I'm confused, how do you think this disproves the claim in the article? What do you think that quoted portion was talking about?
  - simianwords 3mo ago
    The two claims are not related, my bad for making it seem so. I about that a serious person would still needs proof that AI is becoming more efficient. If you need proof, here's one: > A year ago, we verified a preview of an unreleased version of @OpenAI o3 (High) that scored 88% on ARC-AGI-1 at est. $4.5k/task > Today, we’ve verified a new GPT-5.2 Pro (X-High) SOTA score of 90.5% at $11.64/task > This represents a ~390X efficiency improvement in one year https://x.com/arcprize/status/1999182732845547795 https://x.com/arcprize/status/1999182732845547795 This is not even including newest improvements. GPT 5.6 beats GPT 5.5 by 1 OOM. https://x.com/GregKamradt/status/2075274981794300113 https://x.com/GregKamradt/status/2075274981794300113
    - unknownfuture 3mo ago
      That's still not the kind of efficiency the article is talking about. They're referring to raw training and inference scaling and it's relationship (or lack thereof) to traditional economies of scale that we've seen with past technologies where they get cheaper as adoption increases, not more expensive. Its true that the models requiring fewer turns and tokens due to increasing sophistication improves cost efficiency for users but that doesn't address the fundamental computational scaling problems of LLM architectures.
      - 3mo ago

116 points by latexr 3mo ago

https://archive.ph/20260715223901/https://www.theatlantic.com/technology/2026/07/generative-ai-engineering-disaster/687901/ https://archive.ph/20260715223901/https://www.theatlantic.co...

maxcb 3mo ago

Comparing how traditional tech companies scale to how frontier AI companies scale doesn't seem very fair to me. AI is still relatively new territory, in comparison traditional software companies benefit from decades of investment in cloud infrastructure, networking, and hardware. I'm certain those technologies also went through periods of heavy investment before they became scalable. The author also talks about advances like smartphones as though they arrived without significant challenges or trade-offs. Every major technological shift has come with some cost, e.g environmental, shortages, investments. I'd argue most things are difficult to scale in the beginning. Just because today's LLMs are expensive and resource intensive doesn't mean the technology is fundamentally flawed, it may simply mean the current approach isn't the one we'll end up with. But naturally, someone working on another approach would have something to criticise about others. Finally, I'm not convinced by the claim that we're "stuck" with LLMs just because they're heavily marketed. They create value for many people, which is why they've been adopted so quickly. As the author points out, investors care about economics. I'm sure people would listen if someone developed a cheaper/more sustainable technology that delivered the same value.

pixl97 3mo ago
Right, if you came out with a technology that worked just as well as LLMs at half the power people/companies would jump at it from the absolutely massive savings in hardware and power. AI companies are looking at every other method they can, and inventing new ones, none of them have worked as well as the transformer so far.
watwut 3mo ago
> They create value for many people, which is why they've been adopted so quickly. A lot of that adoption is completely useless crap tho. Like ai in vacuum cleaner that ads absolutely nothing useful to anything. AI buttons intentionally at places I randomly click at, so that I am forced to open it. Google search that defaults to ai, so that we have to use it after then nerfed real search. There are useful usages of LLMs. But huge bulk of the adoption is companies realizing they wont get investors money if they dont add ai button, it does not have to be useful.
- SoftTalker 3mo ago
  Yeah there's a huge part of it that is "the investors want tulips, so we will plant tulips"
  - tharmas 3mo ago
    Thanks. You're comment made me laugh.
- bryanlarsen 3mo ago
  Which happens for every new technology. Blue LED's in everything, capacitive buttons on everything, car buttons replaced with ipads, et cetera. Some of the random placements will prove out useful but most won't and will disappear.
  - watwut 3mo ago
    Frankly, no, totally not. And not nearly to this extend. I have seen enough of "new technologies" to know that it was not as crazy at it is now with AI. The hysterical doom trolling, the hype, the tokenmaxxing, the personification of it, all of that is off the charts compared to anything I have seen in the past.
  - doctorwho42 3mo ago
    Yes, but corporations didn't bet the Capex of the entire GDP of the world on it
JackSlateur 3mo ago
"They create value for many people" They create value for many people because these people are not paying the cost of the tool;
tedggh 3mo ago
I think LLMs, with only a few exceptions, have mostly created the illusion of value. It’s undeniable I can write code, troubleshoot and document much faster than in the pre LLM era, but putting a dollar value to that it’s not particularly easy. One group of people being 100x faster at one task doesn’t mean the organization as a whole runs at 100x the speed, in fact in most cases where the bottle neck is somewhere else, the productivity gains of the 100x group have little to no impact, or even a negative impact in some cases due to token overspending.
- naveen99 3mo ago
  How about sole proprietors?
  - andrekandre 3mo ago
    that would be interesting to see with a large longitudinal study any new technology also introduces new bottlenecks, so i wonder if its a 2 steps forwards 1 step backwards (or vice versa) situation (e.g llm summarized my email conversation to be more clear but left out a key detail; now i have to rework my proposal and i lost 1 week of income etc)
    - naveen99 3mo ago
      Resnet’s solve the regression problem.
- unknownfuture 3mo ago
  Yup. We're just all re-learning Amdahl's Law.

andy_ppp 3mo ago

The ultra wealthy will out compete you for the resources you need if we don’t start taxing them has been a discussion from the left for quite some time. It seems to keep coming true in new and disastrous ways, we are going to see more and more extreme concentrations of capital and power distort things and pretty soon it won’t just be middle and poor people who cannot afford anything, the whole value chain is being redesigned around things billionaires want and what could be better for them than agents who replace all the white collar workers. You might think you’re safe from this, I don’t. It’s worth remembering that during the industrial revolution in Britain, the fastest growing country the world had ever seen, most people were in abject poverty. This tech revolution might end up being worse.

efields 3mo ago
Yup. This is a policy problem.
throwaway323929 3mo ago
Instead of just rejecting the future like a Roman fatalist because you've decided that it's just going to be someone else controlling you, maybe a better plan is to adopt emerging technology when it comes into play (like for example when the Internet came into play) and use it to start your own company, or make yourself more productive, or disrupt an industry that desperately needs it. I shouldn't have to explain the benefits of productivity improvements and computer technology to people on Hacker News, but the place for some reason has been hijacked by neo luddites from Reddit that apparently have nothing better to do than troll about AI on a web site that is literally dedicated to an industry they don't like. Sam Altman literally used to run Y Combinator. Why not use that energy you're wasting to adopt emerging technology instead of bashing it? There are places that recognized manufacturing was going to go away, transitioned a service economy, and did well economically, and places that ignored reality and are now in disrepair. Now's a great time to decide if you want to be Seattle or Cleveland. Don't fight the future. And don't think it's a billionaires fault when certain places are more successful then others because they didn't ban AI and datacenters while other places squandered the opportunity because they didn't confront the pitchfork mob driven into a frenzy by yellow journalism.

vanuatu 3mo ago

i dont think the comparison to traditional vc saas is very good yes, in traditional venture you want cost per marginal user to decrease and leverage your platform at scale but improving llms shifts the frontier of their capability and unlocks entirely new use cases. so far, every mega training run has resulted in a model that has paid itself off profitably fully loaded. perhaps the TAM of intelligence has no ceiling? not to say that we shouldnt be investing in efficient models, but the efficiency comes after we create another mega shoggoth that we can make more efficient

simianwords 3mo ago
you think companies are not focusing on efficiency but in reality they are both almost fully correlated. if you make a model more efficient, you get to make it do more reasoning and we know that the upper limit for reasoning is very very high. we have real proof - GPT 5.6 Luna is 50% cheaper than 5.5. And this keeps repeating.

scotty79 3mo ago

I can't imagine a successful civilization where non-computing energy usage isn't a rounding error.

nancyminusone 3mo ago
I can. AI says it will solve all problems, like curing cancer. Suppose it does. Now there are no more problems and no reason to run the AIs, at which time computing energy will fall rapidly. The alternative is that there will always be problems, and we could spend as much as possible trying to compute solutions, but they will not be found.
- inigyou 3mo ago
  Maybe we'll all upload ourselves to the Dyson sphere, so there will be nothing left but computation.
  - bigbuppo 3mo ago
    42.
- scotty79 3mo ago
  > Now there are no more problems and no reason to run the AIs, Are you sure we are talking about the same species?
  - JackSlateur 3mo ago
    Creating fake problems is not a caracteristic I would give to "a successful civilization" :)
    - scotty79 3mo ago
      Each solution generates new problems. That's the general physics of intelligence in our universe. Also label of "fake" is completely arbitrary. Because someone might call cancer a fake problem, because people shouldn't live too long and they need a reason to die.

srhtftw 3mo ago

Approaching an AI version of https://hitchhikers.fandom.com/wiki/Shoe_Event_Horizon https://hitchhikers.fandom.com/wiki/Shoe_Event_Horizon

rglover 3mo ago

It is, in some ways, but no matter your opinion, AI in engineering will continue unabated. Wrote about this recently (from a vibe coding perspective) [1]. This applies to corporations, too, as they're betting their futures on fewer engineers, more AI. [1] https://graybearding.bearblog.dev/they-got-something-on-the-screen/ https://graybearding.bearblog.dev/they-got-something-on-the-...

erelong 3mo ago

https://archive.ph/xo7RG https://archive.ph/xo7RG

bilater 3mo ago

https://x.com/PalmerLuckey/status/2077216959930728889 https://x.com/PalmerLuckey/status/2077216959930728889 Ya'll are not ready for AI to just keep getting better and better. The slop-midwit-doomerism you are being fed as part of some last ditch copium that your SE identity matters and is going to be needed prevents you from seeing what is inevitable. You can't beat em. Try joining em.

selicos 3mo ago
The month to month changes and improvements across free tools is noticeable. A number of skills SE and other roles are built on (or people started with) like more simple scripting and automation work are reachable by a BA with proper requirements, some background with development/coding, and access to AI tools. It's a continuation of both low code development tools and the ongoing compaction of development groups through DevOps. AI is enabling (vibe) code development by non devs. It's enabling a BA or even direct business rep to take requirements into a prototype or example. LLMs are out of the box and will continue developing. How/If you use them is up to you.
simianwords 3mo ago
[flagged]

devin 3mo ago

The rentier class needs a very strong correction. A lot of the conversation focused on what AI is and isn't is often just a proxy for frustration at the structural problems in our society which are in large part due to the inability for hard-working people to live good quality lives because of the incredible greed of the few.

Varelion 3mo ago
WOAH! Communism in my homefront? That's an NSPM-7 violation there, buddy.

simianwords 3mo ago

[flagged]

unknownfuture 3mo ago
> This is obviously untrue so either the author is knowingly lying or is plain incompetent. We had LLMs not being able to do simple high school mathematics a year back, now it is solving open problems in mathematics. Fields medalists in physics and mathematics are using it on a daily basis. I'm confused, how do you think this disproves the claim in the article? What do you think that quoted portion was talking about?
- simianwords 3mo ago
  The two claims are not related, my bad for making it seem so. I about that a serious person would still needs proof that AI is becoming more efficient. If you need proof, here's one: > A year ago, we verified a preview of an unreleased version of @OpenAI o3 (High) that scored 88% on ARC-AGI-1 at est. $4.5k/task > Today, we’ve verified a new GPT-5.2 Pro (X-High) SOTA score of 90.5% at $11.64/task > This represents a ~390X efficiency improvement in one year https://x.com/arcprize/status/1999182732845547795 https://x.com/arcprize/status/1999182732845547795 This is not even including newest improvements. GPT 5.6 beats GPT 5.5 by 1 OOM. https://x.com/GregKamradt/status/2075274981794300113 https://x.com/GregKamradt/status/2075274981794300113
  - unknownfuture 3mo ago
    That's still not the kind of efficiency the article is talking about. They're referring to raw training and inference scaling and it's relationship (or lack thereof) to traditional economies of scale that we've seen with past technologies where they get cheaper as adoption increases, not more expensive. Its true that the models requiring fewer turns and tokens due to increasing sophistication improves cost efficiency for users but that doesn't address the fundamental computational scaling problems of LLM architectures.
    - 3mo ago