Does anyone else think it's weird how many times someone says “if you showed me a current LLM in 2020, I wouldn't even believe it was real” and then they follow that up with “but there's no way robots or AI will be replacing humans in the next 10 years”
Maybe they will, maybe they won't, but doubting the pace of development seems like a particularly dubious thing to hang your hat on
I think there's a difference between "will an AI do some jobs that some humans are doing now" and "will AI permanently disemploy tens of millions of people such that there is no productive work for them to do."
I think that's setting the bar too high though. If the AI revolution leads to 30 million people taking a 25% wage haircut as they're forced into taking lower wage jobs, that would be politically and economically catastrophic. And that frankly seems like a best case scenario - home health aids do not get paid 75% of what software developers or middle managers get
I mean it sort of depends what you mean by setting the bar. I'm mostly trying to distinguish between whether AI's impact on the labor market is a difference of degree or kind. If it's degree it can be quite bad but all our normal economic policy responses and thinking remain intact, if it's a difference in kind it sort of destroys everything we know about how to make people's lives better.
I get that, but I think much of the dramatic improvement of the last couple of years, which is certainly observable in the software development domain, comes largely from throwing a massive amount of hardware and power at the problem. The fundamental limitations of LLMs are still there.
I mean that's what I'm saying, doubting the pace of change seems weird to me. LLMs can do x and y sure, but there's NO WAY they'll do z!
Objections rooted in things like people like getting served by other humans, that makes way more sense than "I don't think the tech will be good enough"
But gaming out this *specific* solution, throwing lots and lots of compute at it, means you’re introducing very real physical restraints that are already showing up: Chip shortages; capital expenditures greater than anything ever done before; power needs.
As Ben Thompson likes to point out, once you enter this physical realm, companies start to more like utilities. It’s just a very different realm where “tech gets better” assumptions need to be scrutinized.
Right, the "imagine what it's like in 5 years" thing assumes continuing the trend rather than mean regression. Could end up being like Disco Stu's graph of "disco record sales" that ends in 1978.
My point of view is biased by working in tech, which is where AI seems to especially shine and is the industry that has probably adopted AI the fastest and deepest. That said, AI does best when you have well-documented and sensible processes and standards. In my experience, most companies do not have that and if they do, they don't have them thoroughly documented (another pre-req for easy adoption of AI). I think eventually AI adoption will be able to collapse a whole team of people into a small number of people who spec the work really well, and then validate it after AI has had at it. But before that can happen, companies who have oriented all of their tools and processes around humans being able to understand and actuate them will have to completely reconfigure those tools and processes to be AI-understandable.
I don’t know yet what the context of the “Rich people love quiet” article is in the podcast, but I have a *huge*, very personal bone to pick with it, as someone who was born and raised on the Upper East Side.
Gonzalez is talking about Fifth through Park Avenues, and ignoring where I grew up, between 2nd and 3rd Avenue. Those, with Lexington and 1st, were normal mixed-use buildings, with plenty of noise and life at all times of day and into the night, and plenty of people sitting on stoops on the side streets. If you ever watch the opening credits of the Jack Lemmon vehicle *The Prisoner Of Second Avenue*, where a car simply rides down the titular thoroughfare, you’ll see exactly what I mean. On top of this, my street, 75th, has a firehouse and a junior high school with a huge courtyard to play in. I can literally sleep through anything!
I grant that she’s writing about the UES a decade after I went away for college, and at one point the schoolyard was taken over with tennis courts, the nets left up even when no one was using them, so “rich people” mores may have even encroached my old neighborhood. But the article is pretty clear she’s painting the entire Upper East Side with the broad brush of “rich snobs”, which just wasn’t true! Let me have my memories, for god’s sake!
I really must know: Was it too quiet in there, where she was shown the private art collection away from the party? Surely she should have invited the other party-goers into this back room to restore the authenticity she craves?
Really enjoyed this conversation. Found myself coming down on Matt's side for the most part.
Some thoughts:
1. This is an incredible time to be an entrepreneur. Find your competitive advantage, hone that, and then outcompete every incumbent by doing everything else more cheaply than they are, with AI (The Argument is a good example of this!)
2. Self-checkouts are the better counter-example to the Starbucks case than chess. There are plenty of services where people prefer the antisocial, cheaper option.
3. One thing stopping wider adoption of AI in customer service is jailbreaking (e.g. customers getting plane tickets for free by "emotionally manipulating" a chat bot) - it'll be interesting to see if industries accept this as a cost of doing business (like retail shrinkage) and/or if there'll be a push to criminalize such activity
4. "The Argument" as a publication name is really annoying to have to refer to when you're having an argument.
Metaculus puts a general purpose household robot at late 2030, so I think you should probably read more about what that process is going to look like, Jerusalem. LLMs still having gotcha questions like the car wash thing don't seem too relevant to me. It's not that we're going to put a weak LLM inside a robot and it will make logical mistakes, it's that Opus 7 is going to program incredibly efficient training sims for robots that will accelerate progress on physical interaction until they get to an acceptable gaffe rate, at which point mass production and learning curves will work their magic.
The discussion about the white paper kind of demonstrates why this kind of thing is so difficult in America: cars are actually oppression! Everyone thinks cute, walkable downtowns are more charming and nice, but if we all feel like we have to drive there and there's no parking, the strip mall wins out. But this is because people just assume they have to drive everywhere. American cities are set up like this, but it's extraordinarily uncommon in the rest of the world because a mixed-use neighborhood where you walk to most of your everyday stores and can take efficient public transit to anywhere less common is how cities work pretty much everywhere else.
I also think we should just grant noise ordinance exemptions to bars and restaurants on Fridays and Saturdays. It irritates me so much that the Oakland police enforce noise complaints against bars that have been around for decades when they get new rich neighbors who grew up in some quiet suburb and can't deal with the noise of a bar they knew existed when they bought the place. It's ruining many people's good time for people who shouldn't have bought a house on a block with a bar. The bartenders have to go home and sleep through all the lawnmowers and construction noise that people for some reason start at 8am, the least people can do is let them make as much money as they can by keeping the patio open until closing time.
One of the things hanging in the background of this topic is what does "massive" job displacement mean? I got the sense that fundamentally, the reason why Jerusalem sounds sanguine and Matt sounds pessimistic on this front is that they have a fundamental disagreement over how large the scale of disruption has to be to be a big deal.
Matt referenced it a bit obliquely in the episode, but the automation/off-shoring of manufacturing in the 80s and 90s is a rorschach test for this. If you're an economist, you might look at it and think, "Goods are cheaper, unemployment is low, it was a win-win for the economy, so there's no problem!" But it was and remains to this day a massive political problem! When a factory leaves a small town where it's the main employer, it depresses the town and can even ruin it forever! Disruption caused by AI doesn't have to be unprecedented. If it does the same to lawyers and admins as automation did to textile workers in the 80s, that *is* a massive problem.
Does anyone else think it's weird how many times someone says “if you showed me a current LLM in 2020, I wouldn't even believe it was real” and then they follow that up with “but there's no way robots or AI will be replacing humans in the next 10 years”
Maybe they will, maybe they won't, but doubting the pace of development seems like a particularly dubious thing to hang your hat on
I think there's a difference between "will an AI do some jobs that some humans are doing now" and "will AI permanently disemploy tens of millions of people such that there is no productive work for them to do."
I think that's setting the bar too high though. If the AI revolution leads to 30 million people taking a 25% wage haircut as they're forced into taking lower wage jobs, that would be politically and economically catastrophic. And that frankly seems like a best case scenario - home health aids do not get paid 75% of what software developers or middle managers get
I mean it sort of depends what you mean by setting the bar. I'm mostly trying to distinguish between whether AI's impact on the labor market is a difference of degree or kind. If it's degree it can be quite bad but all our normal economic policy responses and thinking remain intact, if it's a difference in kind it sort of destroys everything we know about how to make people's lives better.
I get that, but I think much of the dramatic improvement of the last couple of years, which is certainly observable in the software development domain, comes largely from throwing a massive amount of hardware and power at the problem. The fundamental limitations of LLMs are still there.
I mean that's what I'm saying, doubting the pace of change seems weird to me. LLMs can do x and y sure, but there's NO WAY they'll do z!
Objections rooted in things like people like getting served by other humans, that makes way more sense than "I don't think the tech will be good enough"
But gaming out this *specific* solution, throwing lots and lots of compute at it, means you’re introducing very real physical restraints that are already showing up: Chip shortages; capital expenditures greater than anything ever done before; power needs.
As Ben Thompson likes to point out, once you enter this physical realm, companies start to more like utilities. It’s just a very different realm where “tech gets better” assumptions need to be scrutinized.
Right, the "imagine what it's like in 5 years" thing assumes continuing the trend rather than mean regression. Could end up being like Disco Stu's graph of "disco record sales" that ends in 1978.
My point of view is biased by working in tech, which is where AI seems to especially shine and is the industry that has probably adopted AI the fastest and deepest. That said, AI does best when you have well-documented and sensible processes and standards. In my experience, most companies do not have that and if they do, they don't have them thoroughly documented (another pre-req for easy adoption of AI). I think eventually AI adoption will be able to collapse a whole team of people into a small number of people who spec the work really well, and then validate it after AI has had at it. But before that can happen, companies who have oriented all of their tools and processes around humans being able to understand and actuate them will have to completely reconfigure those tools and processes to be AI-understandable.
I don’t know yet what the context of the “Rich people love quiet” article is in the podcast, but I have a *huge*, very personal bone to pick with it, as someone who was born and raised on the Upper East Side.
Gonzalez is talking about Fifth through Park Avenues, and ignoring where I grew up, between 2nd and 3rd Avenue. Those, with Lexington and 1st, were normal mixed-use buildings, with plenty of noise and life at all times of day and into the night, and plenty of people sitting on stoops on the side streets. If you ever watch the opening credits of the Jack Lemmon vehicle *The Prisoner Of Second Avenue*, where a car simply rides down the titular thoroughfare, you’ll see exactly what I mean. On top of this, my street, 75th, has a firehouse and a junior high school with a huge courtyard to play in. I can literally sleep through anything!
I grant that she’s writing about the UES a decade after I went away for college, and at one point the schoolyard was taken over with tennis courts, the nets left up even when no one was using them, so “rich people” mores may have even encroached my old neighborhood. But the article is pretty clear she’s painting the entire Upper East Side with the broad brush of “rich snobs”, which just wasn’t true! Let me have my memories, for god’s sake!
And now, today, I find out that Gonzalez is also a giant hypocrite: https://www.nytimes.com/2026/07/26/nyregion/metropolitan-diary.html
I really must know: Was it too quiet in there, where she was shown the private art collection away from the party? Surely she should have invited the other party-goers into this back room to restore the authenticity she craves?
Really enjoyed this conversation. Found myself coming down on Matt's side for the most part.
Some thoughts:
1. This is an incredible time to be an entrepreneur. Find your competitive advantage, hone that, and then outcompete every incumbent by doing everything else more cheaply than they are, with AI (The Argument is a good example of this!)
2. Self-checkouts are the better counter-example to the Starbucks case than chess. There are plenty of services where people prefer the antisocial, cheaper option.
3. One thing stopping wider adoption of AI in customer service is jailbreaking (e.g. customers getting plane tickets for free by "emotionally manipulating" a chat bot) - it'll be interesting to see if industries accept this as a cost of doing business (like retail shrinkage) and/or if there'll be a push to criminalize such activity
4. "The Argument" as a publication name is really annoying to have to refer to when you're having an argument.
Metaculus puts a general purpose household robot at late 2030, so I think you should probably read more about what that process is going to look like, Jerusalem. LLMs still having gotcha questions like the car wash thing don't seem too relevant to me. It's not that we're going to put a weak LLM inside a robot and it will make logical mistakes, it's that Opus 7 is going to program incredibly efficient training sims for robots that will accelerate progress on physical interaction until they get to an acceptable gaffe rate, at which point mass production and learning curves will work their magic.
The discussion about the white paper kind of demonstrates why this kind of thing is so difficult in America: cars are actually oppression! Everyone thinks cute, walkable downtowns are more charming and nice, but if we all feel like we have to drive there and there's no parking, the strip mall wins out. But this is because people just assume they have to drive everywhere. American cities are set up like this, but it's extraordinarily uncommon in the rest of the world because a mixed-use neighborhood where you walk to most of your everyday stores and can take efficient public transit to anywhere less common is how cities work pretty much everywhere else.
I also think we should just grant noise ordinance exemptions to bars and restaurants on Fridays and Saturdays. It irritates me so much that the Oakland police enforce noise complaints against bars that have been around for decades when they get new rich neighbors who grew up in some quiet suburb and can't deal with the noise of a bar they knew existed when they bought the place. It's ruining many people's good time for people who shouldn't have bought a house on a block with a bar. The bartenders have to go home and sleep through all the lawnmowers and construction noise that people for some reason start at 8am, the least people can do is let them make as much money as they can by keeping the patio open until closing time.
Love All Souls
One of the things hanging in the background of this topic is what does "massive" job displacement mean? I got the sense that fundamentally, the reason why Jerusalem sounds sanguine and Matt sounds pessimistic on this front is that they have a fundamental disagreement over how large the scale of disruption has to be to be a big deal.
Matt referenced it a bit obliquely in the episode, but the automation/off-shoring of manufacturing in the 80s and 90s is a rorschach test for this. If you're an economist, you might look at it and think, "Goods are cheaper, unemployment is low, it was a win-win for the economy, so there's no problem!" But it was and remains to this day a massive political problem! When a factory leaves a small town where it's the main employer, it depresses the town and can even ruin it forever! Disruption caused by AI doesn't have to be unprecedented. If it does the same to lawyers and admins as automation did to textile workers in the 80s, that *is* a massive problem.