3 ms·

Somewhere around 2005-2007, when people were wondering if the Internet was done, PG was fond of saying "It has decades to run. Social changes take longer than

Somewhere around 2005-2007, when people were wondering if the Internet was done, PG was fond of saying "It has decades to run. Social changes take longer than technical changes." I think we're at a similar point with LLMs. The technical stuff is largely "done" - LLMs have closer to 10% than 10x headroom in how much they will technologically improve, we'll find ways to make them more efficient and burn fewer GPU cycles, the cost will come down as more entrants mature. But the social changes are going to be vast. Expect huge amounts of AI slop and propaganda. Expect white-collar unemployment as execs realize that all their expensive employees can be replaced by an LLM, followed by white-collar business formation as customers realize that product quality went to shit when all the people were laid off. Expect the Internet as we loved it to disappear, if it hasn't already. Expect new products or networks to arise that are less open and so less vulnerable to the propagation of AI slop. Expect changes in the structure of governments. Mass media was a key element in the formation of the modern nation state, mass cheap fake media will likely lead to its fragmentation as any old Joe with a ChatGPT account can put out mass quantities of bullshit. Probably expect war as people compete to own the discourse.

tossandthrow 6mo ago
You are very strong on the "slop" bias. Why? In managing a large to enterprise sized code base, I experience the opposite. I can guarantee a much more homogenous quality of the code base. It is the opposite of slop I am seeing. And that at a lower cost. Today,I literally made a large and complex migration of all of our endpoints. Took ai 30 minutes, including all frontends using these endpoints. Works flawlessly, debt principal down.
- chaps 6mo ago
  Which company do you work at so we can avoid your migrated endpoints?
  - tossandthrow 6mo ago
    Wtf. You don't even know what the migration was about?
    - chaps 6mo ago
      I mean, I'm always down for learning something new. But I hope what I learn includes the name of the company I'd like to avoid.
      - tossandthrow 6mo ago
        Your tone is in conflict with the statement that you are curious.
        chaps 6mo ago
        It's because you're deflecting. :)
        tossandthrow 6mo ago
        Deflecting from what? Telling the company name so you can avoid it due to your incredibly curious nature?
        chaps 6mo ago
        Sigh. Look friend, I really hope you can realize how you sound in your post. You're extraordinarily confidently saying that you refactored some ambiguous endpoints in 30 minutes. Whenever I see someone act that confidently towards refactoring, thousands alarms go off in my head. I hope you see how it sounds to others. Like, at least spend longer than a lunch break on it with just a tad more diligence. Or hell, maybe even consider LIEing about how much time you spent on it. But my point is that your shortcuts will burn you. If you want to go down that path, I'm happy to be a witness to eventual schadenfreude. My issue isn't with the fact that you used AI. My issue is with how confident you are that it worked well and exactly to spec. I'm very well aware of what these systems can do. Hell, I've been able to get postgres to boot inside linux inside postgres inside linux inside postgres recently with these tools. But I'm also acutely aware of the aggressive modes that these systems can break in. So again, which company should we all avoid so that we can avoid your, specifically your, refactoring?
  - bsmith 6mo ago
    All big tech companies are mandating employees to use AI for tasks. Unless there's a similar movement to open source that is AI-free, you're going to need to be tech-free of you want to avoid companies that use AI.
- apsurd 6mo ago
  One point: yes, you're speaking from the power position. God-mode over a fleet of minions has always been an engineer's wet-dream. That's not even bad per-say. It's the collateral damage down stream that's at issue. Maybe you don't see any damage, but that's largely the point. Is it really up to you to say?
  - tossandthrow 6mo ago
    What is the collateral damage? In ensuring that a bunch of endpoints use the same structure using LLMs?
    - apsurd 6mo ago
      Let's not debate that it's possible to make very large very safe changes. It is possible that you did that. This is about "slop bias". I'd wager that empowering everyone, especially power-positions to ship 50x more code will produce more code that is slop than not. You strongly oppose this because it's possible for you to update an API? I'm stuck on the power-position thing because I'm living it. I'm pro-AI but there are AI-transformation waves coming in and mandating top-down. From their green-field position it's undeniable crush-mode killin' it. Maintenance of all kinds is separate and the leaders and implementors don't pay this cost. Maybe AI will address everything at every level. But those imposing this world assume that to be true, while it's the line-engineers and sales and customer service reps that will bear the reality.
      - tossandthrow 6mo ago
        > Maybe AI will address everything at every level. I think this is the idea you need to entertain / ponder more on. I largely agree with you, what I don't agree with is the weighting about the individual elements. My point was that I could do a 30 minutes cleanup in order to streamline hundreds of endpoints. Without AI I would not have been able to justify this migration due to business reasons. We get to move faster, also because we can shorten deprication tails and generally keep code bases more fit more easily. In particular, we have dropped the external backoffice tool, so we have a single mono repo. An Ai does tasks all the way from the infrastructure (setting policies to resources) and all the way to the frontends. Equally, if resources are not addressed in our codebase, we know at a 100% it is not in use, and can be cleaned up. Unused code audits are being done on a weekly schedule. Like our sec audits, robustness audits, etc.
- hliyan 6mo ago
  > Today, I literally made a large and complex migration of all of our endpoints. Took ai 30 minutes, including all frontends using these endpoints. Works flawlessly, debt principal down. This is either a very remarkable or a very frightening statement. You're claiming flawless execution within the same day as the change. If you're unable to tell us which product this is, can you at least commit to report back in a month as to how well this actually went?
  - tossandthrow 6mo ago
    It is a part of the smoke testing process right now. But we run 90% test coverage, e2e test etc. None of which had been altered, and are all passing. Migrations are generally not that high risk if you have a code base in alright shape.
- peterbell_nyc 6mo ago
  Seeing plenty of this. The quality of agentic code is a function of the quantity and quality of adversarial quality gates. I have seen no proof that an agentic system is incapable of delivering code that is as functional, performant and maintainable as code from a great team of developers, and enough anecdotes in the other direction to suggest that AI "slop" is going to be a problem that teams with great harnesses will be solving fairly soon if they haven't already.
  - apsurd 6mo ago
    I take your point but then it makes me think is there no more value in diversity? [Philosophy disclaimer] So in a code-base diversity is probably a bad idea, ok that makes sense. But in an agentic world, if everything is run through the Perfect Harness then humans are intentionally just triggers? Not even that, like what are humans even needed for? Everything can be orchestrated. I'm not against this world, this is an ideal outcome for many and it's not my place to say whether it's inevitable. What I'm conflicted on is does it even "work" in terms of outcomes. Like have we lost the plot? Why have any humans at all. 1 person billion dollar company incoming. Software aside, is the premise even valid? 1 person's inputs multiplied by N thousand agents -> ??? -> profit
    - tossandthrow 6mo ago
      These are the right questions to ask.
    - SturgeonsLaw 6mo ago
      > Why have any humans at all Why have humans do work at all? We could have a radically better existence. It would mean that the few at the top of the pyramid lose their privileged position relative to the rest of us, but we could, actually, have that world of abundance for all. Work in the current sense arguably isn't even desirable Maybe I've just read too many Culture books.
- bluecheese452 6mo ago
  Ironically the post saying it is not slop sounds exactly like ai slop.
  - tossandthrow 6mo ago
    Too. Many spelling errors for that to be slop...
- skeeter2020 6mo ago
  >> Works flawlessly, debt principal down. I don't doubt it completed the initial coding work in a short time, but the fact that you've equated that with flawless execution is on the concerning-scary spectrum. I can only assume you're talking "compiles-runs-ship it" The danger is not generating obvious slop, it's accepting decent and convincing outputs as complete and absolving ourselves of responsibility.
  - tossandthrow 6mo ago
    You are right, and it happens that the output looks decent. Code idioms, or patterns if you will, is largely our solution. We have small pattern/[pattern].md files througout the code base where we explain how certain things should be done. In this case, the migration was a normalization to the specific pattern specified in the pattern file for the endpoints. Semantics was not changed and the transform was straight forward. Just not task I would be able to justify spending time on from a business perspective. Now, the more patterns you have, and the more your code base adheres to these patterns, the easier you can verify the code (as you recognize the patterns) and the easier you cal call out faulty code. It is easier to hear an abnormality in music than in atmospheric noise. It is the same with code.
hn_throwaway_99 6mo ago
> Somewhere around 2005-2007, when people were wondering if the Internet was done Literally who wondered that? Drives me nuts when people start off an argument with an obvious strawman. I remember the time period of 2005-2007 very well, and I don't remember a single person, at least in tech, thinking the Internet was done. I don't know, maybe some ragebait articles were written about it, but being knee-deep in web tech at that time, I remember the general feeling is that it was pretty obvious there was tons to do. E.g. we didn't necessarily know what form mobile would take, but it was obvious to most folks that the tech was extremely immature and that it would have a huge impact on the Internet as it progressed. That's just one example - social media was still in its nascent stages then so it was obvious there would be a ton of work around that as well.
- magicalist 6mo ago
  > I don't know, maybe some ragebait articles were written about it, but being knee-deep in web tech at that time, I remember the general feeling is that it was pretty obvious there was tons to do Almost definitely professional ragebaiters in Wired or Time or whatever, yeah.
- nostrademons 6mo ago
  If you were in tech in 2005-2007 you were part of a small minority of the general population. It often didn't feel like a small minority because, well, you knew all those other people on the Internet, but that's a pretty strong selection bias. There is, of course, the Paul Krugman quote from 1998 that by 2005 the Internet would be no more important than a fax machine. [1] Here's Wired in 2007 saying, in reference to Facebook, "no company in its right mind would give it a $15 billion valuation". [2] I remember, being at Google in ~2011, we used to laugh at the Wall Street analysts because they would focus on CPC numbers to forecast a valuation, which is important only if the number of clicks is remaining constant. We knew, of course that total Internet usage was still growing quite rapidly and that queries had increased by roughly 4x over the 2009-2013 timeframe. And a lot of people will say "If you're so smart, why aren't you rich?", and I'll point out that many people who assumed the Internet had lots of room to grow in 2005-2007 did end up very rich. Google stock has increased roughly 20x since 2007 (and 40x from its 2009 lows). Meta is now worth $1.6T, a 100x increase over the $15B valuation that everyone thought was insane in 2007. Amazon is also up about 100x. It would not be possible to take the other side of the trade and make these kind of profits if the majority of people did not think the Internet was largely over. [1] https://www.snopes.com/fact-check/paul-krugman-internets-effect-economy/ https://www.snopes.com/fact-check/paul-krugman-internets-eff... [2] https://www.wired.com/2007/10/facebook-future/ https://www.wired.com/2007/10/facebook-future/
vharuck 6mo ago
I agree with the gist of your points, but not much with these two: >followed by white-collar business formation as customers realize that product quality went to shit when all the people were laid off. These will be rare boutique affairs. Based on how mass production and cheap shipping played out, most people value price over quality. The economy will rearrange itself around those savings, making boutique products and services expensive. >mass cheap fake media will likely lead to its fragmentation as any old Joe with a ChatGPT account can put out mass quantities of bullshit. We have this today. And that's not a "same as it ever was" dismissal. Today, there are a lot of terminally online people posting the equivalent of propaganda (and actual propaganda). Social media pushes hot takes in audiences' faces, a portion of them reshare it, and it spreads exponentially. The only limitation to propaganda today is how much time the audience spends staring at the "correct" content provider.

tossandthrow 6mo ago

You are very strong on the "slop" bias. Why? In managing a large to enterprise sized code base, I experience the opposite. I can guarantee a much more homogenous quality of the code base. It is the opposite of slop I am seeing. And that at a lower cost. Today,I literally made a large and complex migration of all of our endpoints. Took ai 30 minutes, including all frontends using these endpoints. Works flawlessly, debt principal down.

chaps 6mo ago
Which company do you work at so we can avoid your migrated endpoints?
- tossandthrow 6mo ago
  Wtf. You don't even know what the migration was about?
  - chaps 6mo ago
    I mean, I'm always down for learning something new. But I hope what I learn includes the name of the company I'd like to avoid.
    - tossandthrow 6mo ago
      Your tone is in conflict with the statement that you are curious.
      - chaps 6mo ago
        It's because you're deflecting. :)
        tossandthrow 6mo ago
        Deflecting from what? Telling the company name so you can avoid it due to your incredibly curious nature?
        chaps 6mo ago
        Sigh. Look friend, I really hope you can realize how you sound in your post. You're extraordinarily confidently saying that you refactored some ambiguous endpoints in 30 minutes. Whenever I see someone act that confidently towards refactoring, thousands alarms go off in my head. I hope you see how it sounds to others. Like, at least spend longer than a lunch break on it with just a tad more diligence. Or hell, maybe even consider LIEing about how much time you spent on it. But my point is that your shortcuts will burn you. If you want to go down that path, I'm happy to be a witness to eventual schadenfreude. My issue isn't with the fact that you used AI. My issue is with how confident you are that it worked well and exactly to spec. I'm very well aware of what these systems can do. Hell, I've been able to get postgres to boot inside linux inside postgres inside linux inside postgres recently with these tools. But I'm also acutely aware of the aggressive modes that these systems can break in. So again, which company should we all avoid so that we can avoid your, specifically your, refactoring?
- bsmith 6mo ago
  All big tech companies are mandating employees to use AI for tasks. Unless there's a similar movement to open source that is AI-free, you're going to need to be tech-free of you want to avoid companies that use AI.
apsurd 6mo ago
One point: yes, you're speaking from the power position. God-mode over a fleet of minions has always been an engineer's wet-dream. That's not even bad per-say. It's the collateral damage down stream that's at issue. Maybe you don't see any damage, but that's largely the point. Is it really up to you to say?
- tossandthrow 6mo ago
  What is the collateral damage? In ensuring that a bunch of endpoints use the same structure using LLMs?
  - apsurd 6mo ago
    Let's not debate that it's possible to make very large very safe changes. It is possible that you did that. This is about "slop bias". I'd wager that empowering everyone, especially power-positions to ship 50x more code will produce more code that is slop than not. You strongly oppose this because it's possible for you to update an API? I'm stuck on the power-position thing because I'm living it. I'm pro-AI but there are AI-transformation waves coming in and mandating top-down. From their green-field position it's undeniable crush-mode killin' it. Maintenance of all kinds is separate and the leaders and implementors don't pay this cost. Maybe AI will address everything at every level. But those imposing this world assume that to be true, while it's the line-engineers and sales and customer service reps that will bear the reality.
    - tossandthrow 6mo ago
      > Maybe AI will address everything at every level. I think this is the idea you need to entertain / ponder more on. I largely agree with you, what I don't agree with is the weighting about the individual elements. My point was that I could do a 30 minutes cleanup in order to streamline hundreds of endpoints. Without AI I would not have been able to justify this migration due to business reasons. We get to move faster, also because we can shorten deprication tails and generally keep code bases more fit more easily. In particular, we have dropped the external backoffice tool, so we have a single mono repo. An Ai does tasks all the way from the infrastructure (setting policies to resources) and all the way to the frontends. Equally, if resources are not addressed in our codebase, we know at a 100% it is not in use, and can be cleaned up. Unused code audits are being done on a weekly schedule. Like our sec audits, robustness audits, etc.
hliyan 6mo ago
> Today, I literally made a large and complex migration of all of our endpoints. Took ai 30 minutes, including all frontends using these endpoints. Works flawlessly, debt principal down. This is either a very remarkable or a very frightening statement. You're claiming flawless execution within the same day as the change. If you're unable to tell us which product this is, can you at least commit to report back in a month as to how well this actually went?
- tossandthrow 6mo ago
  It is a part of the smoke testing process right now. But we run 90% test coverage, e2e test etc. None of which had been altered, and are all passing. Migrations are generally not that high risk if you have a code base in alright shape.
peterbell_nyc 6mo ago
Seeing plenty of this. The quality of agentic code is a function of the quantity and quality of adversarial quality gates. I have seen no proof that an agentic system is incapable of delivering code that is as functional, performant and maintainable as code from a great team of developers, and enough anecdotes in the other direction to suggest that AI "slop" is going to be a problem that teams with great harnesses will be solving fairly soon if they haven't already.
- apsurd 6mo ago
  I take your point but then it makes me think is there no more value in diversity? [Philosophy disclaimer] So in a code-base diversity is probably a bad idea, ok that makes sense. But in an agentic world, if everything is run through the Perfect Harness then humans are intentionally just triggers? Not even that, like what are humans even needed for? Everything can be orchestrated. I'm not against this world, this is an ideal outcome for many and it's not my place to say whether it's inevitable. What I'm conflicted on is does it even "work" in terms of outcomes. Like have we lost the plot? Why have any humans at all. 1 person billion dollar company incoming. Software aside, is the premise even valid? 1 person's inputs multiplied by N thousand agents -> ??? -> profit
  - tossandthrow 6mo ago
    These are the right questions to ask.
  - SturgeonsLaw 6mo ago
    > Why have any humans at all Why have humans do work at all? We could have a radically better existence. It would mean that the few at the top of the pyramid lose their privileged position relative to the rest of us, but we could, actually, have that world of abundance for all. Work in the current sense arguably isn't even desirable Maybe I've just read too many Culture books.
bluecheese452 6mo ago
Ironically the post saying it is not slop sounds exactly like ai slop.
- tossandthrow 6mo ago
  Too. Many spelling errors for that to be slop...
skeeter2020 6mo ago
>> Works flawlessly, debt principal down. I don't doubt it completed the initial coding work in a short time, but the fact that you've equated that with flawless execution is on the concerning-scary spectrum. I can only assume you're talking "compiles-runs-ship it" The danger is not generating obvious slop, it's accepting decent and convincing outputs as complete and absolving ourselves of responsibility.
- tossandthrow 6mo ago
  You are right, and it happens that the output looks decent. Code idioms, or patterns if you will, is largely our solution. We have small pattern/[pattern].md files througout the code base where we explain how certain things should be done. In this case, the migration was a normalization to the specific pattern specified in the pattern file for the endpoints. Semantics was not changed and the transform was straight forward. Just not task I would be able to justify spending time on from a business perspective. Now, the more patterns you have, and the more your code base adheres to these patterns, the easier you can verify the code (as you recognize the patterns) and the easier you cal call out faulty code. It is easier to hear an abnormality in music than in atmospheric noise. It is the same with code.

hn_throwaway_99 6mo ago

> Somewhere around 2005-2007, when people were wondering if the Internet was done Literally who wondered that? Drives me nuts when people start off an argument with an obvious strawman. I remember the time period of 2005-2007 very well, and I don't remember a single person, at least in tech, thinking the Internet was done. I don't know, maybe some ragebait articles were written about it, but being knee-deep in web tech at that time, I remember the general feeling is that it was pretty obvious there was tons to do. E.g. we didn't necessarily know what form mobile would take, but it was obvious to most folks that the tech was extremely immature and that it would have a huge impact on the Internet as it progressed. That's just one example - social media was still in its nascent stages then so it was obvious there would be a ton of work around that as well.

magicalist 6mo ago
> I don't know, maybe some ragebait articles were written about it, but being knee-deep in web tech at that time, I remember the general feeling is that it was pretty obvious there was tons to do Almost definitely professional ragebaiters in Wired or Time or whatever, yeah.
nostrademons 6mo ago
If you were in tech in 2005-2007 you were part of a small minority of the general population. It often didn't feel like a small minority because, well, you knew all those other people on the Internet, but that's a pretty strong selection bias. There is, of course, the Paul Krugman quote from 1998 that by 2005 the Internet would be no more important than a fax machine. [1] Here's Wired in 2007 saying, in reference to Facebook, "no company in its right mind would give it a $15 billion valuation". [2] I remember, being at Google in ~2011, we used to laugh at the Wall Street analysts because they would focus on CPC numbers to forecast a valuation, which is important only if the number of clicks is remaining constant. We knew, of course that total Internet usage was still growing quite rapidly and that queries had increased by roughly 4x over the 2009-2013 timeframe. And a lot of people will say "If you're so smart, why aren't you rich?", and I'll point out that many people who assumed the Internet had lots of room to grow in 2005-2007 did end up very rich. Google stock has increased roughly 20x since 2007 (and 40x from its 2009 lows). Meta is now worth $1.6T, a 100x increase over the $15B valuation that everyone thought was insane in 2007. Amazon is also up about 100x. It would not be possible to take the other side of the trade and make these kind of profits if the majority of people did not think the Internet was largely over. [1] https://www.snopes.com/fact-check/paul-krugman-internets-effect-economy/ https://www.snopes.com/fact-check/paul-krugman-internets-eff... [2] https://www.wired.com/2007/10/facebook-future/ https://www.wired.com/2007/10/facebook-future/

vharuck 6mo ago

I agree with the gist of your points, but not much with these two: >followed by white-collar business formation as customers realize that product quality went to shit when all the people were laid off. These will be rare boutique affairs. Based on how mass production and cheap shipping played out, most people value price over quality. The economy will rearrange itself around those savings, making boutique products and services expensive. >mass cheap fake media will likely lead to its fragmentation as any old Joe with a ChatGPT account can put out mass quantities of bullshit. We have this today. And that's not a "same as it ever was" dismissal. Today, there are a lot of terminally online people posting the equivalent of propaganda (and actual propaganda). Social media pushes hot takes in audiences' faces, a portion of them reshare it, and it spreads exponentially. The only limitation to propaganda today is how much time the audience spends staring at the "correct" content provider.