A viral essay from Citrini about how AI bullishness could be bearish was impactful enough for Bloomberg to give it partial responsibility for a decline in the stock market, and all the cool economics types are talking about it.
If we already assume the singularity, why should we assume compute is a limiting factor? The happy frontier ASI can build specialized contra-bitter-lesson little brothers that run on a potato and handle the simple human level stuff.
Republicans will, of course, demand we leave people out in the cold, making the Holodor look like a picnic and shrinking the population of "useless eaters".
“The ai is smarter than the humans”… so is the calculator!
The idea that ai is going to cause this widespread unemployment is a smoke screen to distort the reality that the only hardware available is self-driving cars. The companies said in 2022 that they’re all 5-10 years away from meaningful progress on hardware that incorporates the learning.
The fact is, ai is just software programs. It’s as likely to destroy jobs as the calculator eliminated accountants. Or the nailgun eliminated carpenters. It’s all hype. It’s all marketing. And to watch the general population fall for it proves the humans aren’t as smart as a calculator.
"Don't Worry About the Vase" is classic! Great article! I wrote something similar a couple of months ago to Citirni's fiction - except mine wasn't, though it was speculative. The more I use AI, and now Openclaw, the more I realize some weaknesses in my article (below), more of a timeframe issue as I think it won't be quite 2028 but closer to 2030-2035 before it hits the fan. But we're already experiencing "end-stage capitalism." We're on an unsustainable path as a society and, like others before us, "it's the economy stupid" LOL AI is just speeding things up to what has been coming for years.
The DoorDash and Uber examples don't make any sense to me. We already have Kayak, and it's not like that has driven Expedia or Priceline out of business, right?
This is a pretty bad rebuttal of Thompson's assertions on DoorDash. One of the primary reasons why vibe coded alternatives to DoorDash wouldn't scale is that they already own the largest network of drivers, restaurants and customers in the real world, which are not "free" to acquire like it seemingly is to "vibe-code" an alternative (which is also not a nonzero cost). If you've ever worked in sales or marketing, you'd realize how hard it is to acquire even one lead.
Thompson's final assertion on how the same argument applies on credit cards is salient since you seem to base a lot of your divergence from the Shitrini report on the credit card model. Well that, can also be applied to DoorDash who acts more than just an aggregator for demand (which in itself is already powerful). Additionally, a standalone agent would not have all the data that DoorDash has with 56 million MAUs... it's speaks more to the ability to commercialize than just complete a transaction.
Isn’t a key piece of the world of this story that acquisition costs go way down? That’s the main difference, not that it’s easier to code up a competitor (though that’s easier too). Everyone has their own agent talking to every other agent, so you don’t need to work hard to find leads. If you’re offering the best deal, their agent connects to your agent which can onboard them immediately.
That's assuming you have multiple platforms that can compete with DoorDash that agents can access (there are only a handful) and that these platforms can actually procure the restaurant and driver network (and the underlying infrastructure like payroll, regulatory, etc. to prop it up). All of this feels very wishy washy since we're talking in theory of some hypothetically perfect marketplace, where in practice its the aggregators with the existing demand able to match supply that will in all likelihood continue to win. That's ignoring the other benefits that DoorDash offers other than just a pure order platform, like the advertising, driver management, safety and trust, subscription, customer support, etc.
If we run with the assumption of zero friction with cheap, perfect agents (load-bearing for the rest of the article), then for the food delivery case:
Restaurant says 'find me customers' to their agent.
Consumer says 'get me some fried chicken' to their agent.
Worker says 'find me (delivery) work' to their agent.
The agents access agent-communication platforms through APIs and make the transactions happen, the consumer gets their chicken and everyone gains from the trade.
Customer support? Just talk to your agent, it can talk to the others for you.
DoorDash creating their own agent doesn't help. They also don't lose to a vibe-coded 'competitor platform' because such platforms no longer need to exist. Why would you need a whole platform dedicated to a specific type of three-party transaction if we have perfect agents which can rapidly negotiate any transaction for you?
If we reject the assumptions then obviously none of this happens. But Thompson never explicitly rejects the assumptions and instead just says "this doesn't happen (even in your imaginary world)". Which is nonsense.
But, if we accept the assumptions, then Citrini is thinking far too small as 'DoorDash goes out of business' would not be a top 100 concern.
It's an interesting thought experiment, but far from a current reality from a technological and behavioral perspective. You assume that people don't want to retain their own agency and submit everything to a faceless AI application.
> You don’t replace $100 in human labor with $70 in AI spending for more than a month or two at most, you replace it with $7 and then $0.70.
> Given the timing there won’t be enough compute for what they describe on this time frame, so compute costs would skyrocket during this scenario much more than they hint at, ruling out many suggested use cases.
Both of these things cannot be true at the same time, Zvi.
Yes they can? If compute costs rise that slows down diffusion, and maybe slows down his timeline for OOM reductions in cost (but in businesses processing there can be OOM reductions in cost from simply improving the efficiency of the initial implementation).
Also, the "sufficiently competent for X" model at T-0 is often itself dramatically cheaper at T-n.
But (much) more likely, compute _costs_ do not rise by so much, compute remains constrained, and diffusion is therefore slower but the cost reductions still apply.
Debt is denominated in nominal dollars. Mortgages, student loans, car payments, etc.
If people have to take lower-paying jobs, debt repayment eats all their discretionary spending. Selling the assets into a deflating market doesn't help. Nor does bankruptcy.
Yes, you can start a business at zero cost, but your total addressable market is ~0 (nobody spends when prices and their wages are falling), and your market share is a millionth of that.
A debt jubilee would appear to be key to getting through this fictional scenario. Otherwise, Japan’s Lost Decades on a much bigger scale and with a much bigger Scoville number.
Yeah, it's tough to watch Zvi trying to argue against the economic logic of this piece without him understanding (anything?) about macroeconomics or past economic crises, even the ones he has personally lived through.
It is good for the explicit taking of the ideas to their logical conclusions, and the reaction to it make very clear that many of these had simply not been presented in the right form before, but the logic would have benefited from someone like yourself, or even a less proficient forecaster.
But therein lies the rub: these conclusions have not been presented like this before because, as presented, they are impossible. But in fact this appears to have been much more efficient communication than the more logically coherent efforts that preceded it!
SaaS is already dying. Look up the rideshare app Empower, which gives 100% of fees to the driver. It's significantly cheaper to the rider than Lyft and Uber, and its use is increasing quickly in NYC.
Between this and AI 2027, it seems we should update on narrative style presentation of AI risk (particularly more hyperbolic scenarios) as having a much larger impact on awareness than more traditional forecasts. Maybe there is something in the average not-hyper-situationally-aware person that is more receptive to "here is a scary scenario (but this is just a hypothesis)" than "my hypothesis is that this scary scenario seems likely." IABIED went this route to some extent, but tried to walk a very tight line of presenting a concrete narrative while also disclaiming the details of that narrative as unimportant at every turn.
I'm left wondering if this is a viable long term strategy, or if it will lead to people getting burned out on worrying exactly when it's becoming more important? There's also some credibility loss when those exact predictions don't come to pass or are modified, like when the AI 2027 authors updated their predictions to be more like AI 2030 and this was somehow interpreted as "see everything is fine."
Got as far as this sentence: "Prices should then drop to match and quality of goods and services should improve, across the board, for this reason and also other reasons described later."
Here a basic level of economics would be helpful. Prices do not just "drop". Prices and wages are both sticky, they generally do not fall, even in the face of deflationary shocks. This is one reason why depressions happen at all. The deflationary spiral leading to depression is a very real risk since it's happened before and can happen again, and price/wage stickiness is one of the main drivers.
Zvi—fix the typo in the subhead “SaaSpocalype”
If we already assume the singularity, why should we assume compute is a limiting factor? The happy frontier ASI can build specialized contra-bitter-lesson little brothers that run on a potato and handle the simple human level stuff.
Republicans will, of course, demand we leave people out in the cold, making the Holodor look like a picnic and shrinking the population of "useless eaters".
“The ai is smarter than the humans”… so is the calculator!
The idea that ai is going to cause this widespread unemployment is a smoke screen to distort the reality that the only hardware available is self-driving cars. The companies said in 2022 that they’re all 5-10 years away from meaningful progress on hardware that incorporates the learning.
The fact is, ai is just software programs. It’s as likely to destroy jobs as the calculator eliminated accountants. Or the nailgun eliminated carpenters. It’s all hype. It’s all marketing. And to watch the general population fall for it proves the humans aren’t as smart as a calculator.
"Don't Worry About the Vase" is classic! Great article! I wrote something similar a couple of months ago to Citirni's fiction - except mine wasn't, though it was speculative. The more I use AI, and now Openclaw, the more I realize some weaknesses in my article (below), more of a timeframe issue as I think it won't be quite 2028 but closer to 2030-2035 before it hits the fan. But we're already experiencing "end-stage capitalism." We're on an unsustainable path as a society and, like others before us, "it's the economy stupid" LOL AI is just speeding things up to what has been coming for years.
https://navyvet1964.substack.com/p/ai-and-the-acceleration-of-end-stage
The DoorDash and Uber examples don't make any sense to me. We already have Kayak, and it's not like that has driven Expedia or Priceline out of business, right?
This is a pretty bad rebuttal of Thompson's assertions on DoorDash. One of the primary reasons why vibe coded alternatives to DoorDash wouldn't scale is that they already own the largest network of drivers, restaurants and customers in the real world, which are not "free" to acquire like it seemingly is to "vibe-code" an alternative (which is also not a nonzero cost). If you've ever worked in sales or marketing, you'd realize how hard it is to acquire even one lead.
Thompson's final assertion on how the same argument applies on credit cards is salient since you seem to base a lot of your divergence from the Shitrini report on the credit card model. Well that, can also be applied to DoorDash who acts more than just an aggregator for demand (which in itself is already powerful). Additionally, a standalone agent would not have all the data that DoorDash has with 56 million MAUs... it's speaks more to the ability to commercialize than just complete a transaction.
Isn’t a key piece of the world of this story that acquisition costs go way down? That’s the main difference, not that it’s easier to code up a competitor (though that’s easier too). Everyone has their own agent talking to every other agent, so you don’t need to work hard to find leads. If you’re offering the best deal, their agent connects to your agent which can onboard them immediately.
That's assuming you have multiple platforms that can compete with DoorDash that agents can access (there are only a handful) and that these platforms can actually procure the restaurant and driver network (and the underlying infrastructure like payroll, regulatory, etc. to prop it up). All of this feels very wishy washy since we're talking in theory of some hypothetically perfect marketplace, where in practice its the aggregators with the existing demand able to match supply that will in all likelihood continue to win. That's ignoring the other benefits that DoorDash offers other than just a pure order platform, like the advertising, driver management, safety and trust, subscription, customer support, etc.
If we run with the assumption of zero friction with cheap, perfect agents (load-bearing for the rest of the article), then for the food delivery case:
Restaurant says 'find me customers' to their agent.
Consumer says 'get me some fried chicken' to their agent.
Worker says 'find me (delivery) work' to their agent.
The agents access agent-communication platforms through APIs and make the transactions happen, the consumer gets their chicken and everyone gains from the trade.
Customer support? Just talk to your agent, it can talk to the others for you.
DoorDash creating their own agent doesn't help. They also don't lose to a vibe-coded 'competitor platform' because such platforms no longer need to exist. Why would you need a whole platform dedicated to a specific type of three-party transaction if we have perfect agents which can rapidly negotiate any transaction for you?
If we reject the assumptions then obviously none of this happens. But Thompson never explicitly rejects the assumptions and instead just says "this doesn't happen (even in your imaginary world)". Which is nonsense.
But, if we accept the assumptions, then Citrini is thinking far too small as 'DoorDash goes out of business' would not be a top 100 concern.
It's an interesting thought experiment, but far from a current reality from a technological and behavioral perspective. You assume that people don't want to retain their own agency and submit everything to a faceless AI application.
> You don’t replace $100 in human labor with $70 in AI spending for more than a month or two at most, you replace it with $7 and then $0.70.
> Given the timing there won’t be enough compute for what they describe on this time frame, so compute costs would skyrocket during this scenario much more than they hint at, ruling out many suggested use cases.
Both of these things cannot be true at the same time, Zvi.
Yes they can? If compute costs rise that slows down diffusion, and maybe slows down his timeline for OOM reductions in cost (but in businesses processing there can be OOM reductions in cost from simply improving the efficiency of the initial implementation).
Also, the "sufficiently competent for X" model at T-0 is often itself dramatically cheaper at T-n.
But (much) more likely, compute _costs_ do not rise by so much, compute remains constrained, and diffusion is therefore slower but the cost reductions still apply.
Debt is denominated in nominal dollars. Mortgages, student loans, car payments, etc.
If people have to take lower-paying jobs, debt repayment eats all their discretionary spending. Selling the assets into a deflating market doesn't help. Nor does bankruptcy.
Yes, you can start a business at zero cost, but your total addressable market is ~0 (nobody spends when prices and their wages are falling), and your market share is a millionth of that.
A debt jubilee would appear to be key to getting through this fictional scenario. Otherwise, Japan’s Lost Decades on a much bigger scale and with a much bigger Scoville number.
Fortunately, we have worse things to worry about.
Yeah, it's tough to watch Zvi trying to argue against the economic logic of this piece without him understanding (anything?) about macroeconomics or past economic crises, even the ones he has personally lived through.
It is good for the explicit taking of the ideas to their logical conclusions, and the reaction to it make very clear that many of these had simply not been presented in the right form before, but the logic would have benefited from someone like yourself, or even a less proficient forecaster.
But therein lies the rub: these conclusions have not been presented like this before because, as presented, they are impossible. But in fact this appears to have been much more efficient communication than the more logically coherent efforts that preceded it!
SaaS is already dying. Look up the rideshare app Empower, which gives 100% of fees to the driver. It's significantly cheaper to the rider than Lyft and Uber, and its use is increasing quickly in NYC.
Between this and AI 2027, it seems we should update on narrative style presentation of AI risk (particularly more hyperbolic scenarios) as having a much larger impact on awareness than more traditional forecasts. Maybe there is something in the average not-hyper-situationally-aware person that is more receptive to "here is a scary scenario (but this is just a hypothesis)" than "my hypothesis is that this scary scenario seems likely." IABIED went this route to some extent, but tried to walk a very tight line of presenting a concrete narrative while also disclaiming the details of that narrative as unimportant at every turn.
I'm left wondering if this is a viable long term strategy, or if it will lead to people getting burned out on worrying exactly when it's becoming more important? There's also some credibility loss when those exact predictions don't come to pass or are modified, like when the AI 2027 authors updated their predictions to be more like AI 2030 and this was somehow interpreted as "see everything is fine."
Got as far as this sentence: "Prices should then drop to match and quality of goods and services should improve, across the board, for this reason and also other reasons described later."
Here a basic level of economics would be helpful. Prices do not just "drop". Prices and wages are both sticky, they generally do not fall, even in the face of deflationary shocks. This is one reason why depressions happen at all. The deflationary spiral leading to depression is a very real risk since it's happened before and can happen again, and price/wage stickiness is one of the main drivers.