I'm not an economist, so I may be wildly off here, but you don't address the core claim of the Citrini essay, which is that the entire US economy is structured around the assumption that skilled cognitive labor is scarce and valuable, and if that changes very quickly, the financial system and economy are at risk of crisis because it cannot adjust quickly enough to respond. Think of all the "safe" mortgages held by upper middle class professionals that could default if there is massive white collar disruption. Think of all the businesses whose business models rely on the assumption that there will be prosperous white collar workers near by who will spend their discretionary income on these businesses (as Citrini notes, most discretionary consumption is driven by the UMC).
Yes, in the long run, due to comparative advantage and everything else you note, everything will shake out. New jobs and businesses will emerge. Money saved by businesses becoming more efficient gets invested. But this all takes time. If we imagine a world where we get 20% white collar unemployment within a few years due to AI, there simply isn't enough time for society to adjust. It will be like the China shock, but at a national scale. Millions of previous UMC professionals will be forced to work low wage service or blue collar jobs or simply exit the workforce. How could that not depress demand, at least temporarily?
Imagine if the digitization of workplaces happened in 5 years rather than the decades it actually took. You take an office in 1980, and suddenly in 1985 it is as digitized as an office in 2026. Surely that would be a big problem for the labor market and the financial system in the short run, right? Think of all the people doing analog office work--file clerks, typists, mail clerks--who would be suddenly obsoleted.
This seems like a fundamentally different argument -- not that AI inherently destroys economic value but that the uncertainty and change caused by the sudden shock could destabilize markets and push us into a depression.
That story isn't fundamentally impossible but I also don't think it is likely. At a fundamental level the reason you get economic outcomes like depressions is because various parts of the economy (eg wages) are sticky and it takes time for companies, people etc to react and adjust. But those same mechanisms mean that it is unlikely that AI displaces people in a sudden zap. As with every new technology it takes time to work it's way through the system.
Moreover, there are big roles for the government and fed to play in this. At the end of the day the government ultimately does have the power to just not allow banks to fail and to inflate the currency to escape 0 bound traps. Governments are reluctant to do this because it represents huge costs but if AI is really just creating absurd amounts of value out of nothing you can afford it.
In the ultimate extreme the government could just hand out money to everyone. You lose your job from AI we keep paying you at 75% of your previous salary. If it really is true that AI is doing everyone's job for less than a quarter of the original cost (and the AI companies are taxable in the US) even absurdly generous moves like that are affordable. Not saying it's the best move but it illustrates that there are options.
For how long will there be comparative advantage? The new kid on the block looks pretty good.
Utilisation: humans 18% of the year (35 hour work week, weekends, holidays, etc.). AI 90% of the year. A fivefold advantage.
Intensity: From anecdata AIs work at about 6x human speed. (A job that a human expert estimated would take him two weeks took an AI about 6 hours. Hard data difficult to find though.) 6x advantage here.
Cost: Fully burdened cost of a human (incl space leasing, furniture and equipment, other overhead, recruiting costs): maybe $150k/yr. AI: Claude Pro Max costs 200/month, $5k/year. Let’s 10x that.
For $50k/year, maybe you can replace 30 $150k/year humans. A $4.45M saving out of $5M of “human resource” cost.
And today’s AIs are the least competent and slowest AI will ever be.
Comparative advantage will exist for a while in occupations and tasks that AI cannot do, but the firm-level competitive pressure to replace humans is immense. Probably there will be fewer and fewer such occupations every year.
Yes, the debt-deflation trap is a huge risk. Japan’s Lost Decades, turned up to 11.
And unlike the wave of digitization you described, this one doesn’t require procuring any new hardware on the firm level (that’s handled centrally by a limited number of firms) so the diffusion of new ai products can be much faster if they’re not self-hosted. You don’t need buy a new laptop every time Anthropic releases a new model. Then again, it’s possible that the API costs will get to the point where you do need new on-premises hardware to run local models
The real big problem that Citrini pointed at is even more simple. As a rule, stock market indices do not track GDP growth, nor does any other financial asset do so in the short term. So even short term GDP growth explosion is not guaranteed to be good for investors, if they cannot get into the investments that are the capital suppliers to the activity driving that growth. They are still in a position to lose money. If there is extreme shift in which assets are good, the markets can seize up, (since nobody knows who is money good), even if the actual number of good assets stays the same.
I think the problem here is people tend to think about the economy in terms of money and jobs not about making stuff. I mean let's imagine that AI is so great that we literally need virtually no one (say 10% of the previous) to work at any company and we can still produce all the services and goods we currently consume.
At the ultimate extreme that means the government could simply say: alright everyone who lost their job now gets paid 75% of what they made previously directly from the government. Since we are still making at least (likely much more) the same amount of stuff that wouldn't be inflationary and most of us get to do whatever the hell we want now. The few people who keep working get huge salaries.
That's not likely to be how things go exactly because we won't see AI replace all jobs in a flash but I think it shows that -- even though we absolutely need to worry about the impact of big shocks -- the mechanism by which AI just makes us inevitably poor doesn't fly.
I shared your timely post with a great friend of mine (we were roommates in college and now he’s a super successful radiologist). My friend is worried (more like terrified) about his children’s future because of AI—how does something like this ever get started? I mean most people I can understand but a guy who graduated second in his class from NYU—this is crazy.
PS—During my last text conversation with my friend he highlighted the fact that unemployment numbers had spiked due to AI (maybe; maybe not). I thought better than to mention that millions of jobs are lost every month but that only net job numbers are reported). I think that much of the alarm is due to availability bias. Human beings seldom think past the seen to what is unseen. Keep up your great work!
Discount any intuited ideas about AI. It itself is science fiction out of The Cybernetic Imagination in Science Fiction (Warrick). The tech is fatally flawed from its arbitrary origins. There's no path around it.
The comparative advantage argument depends on a scarcity of both types of labour.
In your example AI can do both health-care and coding better than humans, and costs less for both.
So you just use 2 AIs, as there are arbitrarily many instantly available,[1] for the cost of electricity and data centre amortisation.[2] Any job that is worth that amount gets done by AI.
Comparative advantage fails in conditions of an arbitrarily large labour surplus on one side.
1. There may be brief delays when all AI capacity is fully utilized, before equilibrium is achieved. Then more datacentres come online.
2. Because of competition among AI firms, the price falls to the costs of the inputs.
The "costs less" part seems to be an assumption on your part? As I read the example, the price of human labour for in-person care would fall below the cost of electricity and data centre amortization for that same type of artificial labour. I tend to agree with your conclusion, though, that in practice, humans would just be pushed out of such professions, much like human computers lost their jobs to electric computers.
Yeah, I don't know what "10x better" is supposed to mean in a comparison between humans and software, but if there's one thing I've learned from reading this blog, it's that prices should be a model outcome, not a model assumption.
Even in the Citrini report, they did not fantasize collapse. Every individual company in the S and P 500 could lose 30%, every loan could default due to suprise deflation, and still we could rebuild the market structure and move on. There could really be a lot of short term pain.
The extent of pain depends on whether or not market forces are allowed to adjust (the relative price mechanism). So far all significant busts in US history were caused exactly because government intervention prevented such adjustments. And no, not all loans will default. Didn’t happen in Japan deflation, will not happen here.
I'm not an economist, so I may be wildly off here, but you don't address the core claim of the Citrini essay, which is that the entire US economy is structured around the assumption that skilled cognitive labor is scarce and valuable, and if that changes very quickly, the financial system and economy are at risk of crisis because it cannot adjust quickly enough to respond. Think of all the "safe" mortgages held by upper middle class professionals that could default if there is massive white collar disruption. Think of all the businesses whose business models rely on the assumption that there will be prosperous white collar workers near by who will spend their discretionary income on these businesses (as Citrini notes, most discretionary consumption is driven by the UMC).
Yes, in the long run, due to comparative advantage and everything else you note, everything will shake out. New jobs and businesses will emerge. Money saved by businesses becoming more efficient gets invested. But this all takes time. If we imagine a world where we get 20% white collar unemployment within a few years due to AI, there simply isn't enough time for society to adjust. It will be like the China shock, but at a national scale. Millions of previous UMC professionals will be forced to work low wage service or blue collar jobs or simply exit the workforce. How could that not depress demand, at least temporarily?
Imagine if the digitization of workplaces happened in 5 years rather than the decades it actually took. You take an office in 1980, and suddenly in 1985 it is as digitized as an office in 2026. Surely that would be a big problem for the labor market and the financial system in the short run, right? Think of all the people doing analog office work--file clerks, typists, mail clerks--who would be suddenly obsoleted.
This seems like a fundamentally different argument -- not that AI inherently destroys economic value but that the uncertainty and change caused by the sudden shock could destabilize markets and push us into a depression.
That story isn't fundamentally impossible but I also don't think it is likely. At a fundamental level the reason you get economic outcomes like depressions is because various parts of the economy (eg wages) are sticky and it takes time for companies, people etc to react and adjust. But those same mechanisms mean that it is unlikely that AI displaces people in a sudden zap. As with every new technology it takes time to work it's way through the system.
Moreover, there are big roles for the government and fed to play in this. At the end of the day the government ultimately does have the power to just not allow banks to fail and to inflate the currency to escape 0 bound traps. Governments are reluctant to do this because it represents huge costs but if AI is really just creating absurd amounts of value out of nothing you can afford it.
In the ultimate extreme the government could just hand out money to everyone. You lose your job from AI we keep paying you at 75% of your previous salary. If it really is true that AI is doing everyone's job for less than a quarter of the original cost (and the AI companies are taxable in the US) even absurdly generous moves like that are affordable. Not saying it's the best move but it illustrates that there are options.
For how long will there be comparative advantage? The new kid on the block looks pretty good.
Utilisation: humans 18% of the year (35 hour work week, weekends, holidays, etc.). AI 90% of the year. A fivefold advantage.
Intensity: From anecdata AIs work at about 6x human speed. (A job that a human expert estimated would take him two weeks took an AI about 6 hours. Hard data difficult to find though.) 6x advantage here.
Cost: Fully burdened cost of a human (incl space leasing, furniture and equipment, other overhead, recruiting costs): maybe $150k/yr. AI: Claude Pro Max costs 200/month, $5k/year. Let’s 10x that.
For $50k/year, maybe you can replace 30 $150k/year humans. A $4.45M saving out of $5M of “human resource” cost.
And today’s AIs are the least competent and slowest AI will ever be.
Comparative advantage will exist for a while in occupations and tasks that AI cannot do, but the firm-level competitive pressure to replace humans is immense. Probably there will be fewer and fewer such occupations every year.
Yes, the debt-deflation trap is a huge risk. Japan’s Lost Decades, turned up to 11.
And unlike the wave of digitization you described, this one doesn’t require procuring any new hardware on the firm level (that’s handled centrally by a limited number of firms) so the diffusion of new ai products can be much faster if they’re not self-hosted. You don’t need buy a new laptop every time Anthropic releases a new model. Then again, it’s possible that the API costs will get to the point where you do need new on-premises hardware to run local models
The real big problem that Citrini pointed at is even more simple. As a rule, stock market indices do not track GDP growth, nor does any other financial asset do so in the short term. So even short term GDP growth explosion is not guaranteed to be good for investors, if they cannot get into the investments that are the capital suppliers to the activity driving that growth. They are still in a position to lose money. If there is extreme shift in which assets are good, the markets can seize up, (since nobody knows who is money good), even if the actual number of good assets stays the same.
I think the problem here is people tend to think about the economy in terms of money and jobs not about making stuff. I mean let's imagine that AI is so great that we literally need virtually no one (say 10% of the previous) to work at any company and we can still produce all the services and goods we currently consume.
At the ultimate extreme that means the government could simply say: alright everyone who lost their job now gets paid 75% of what they made previously directly from the government. Since we are still making at least (likely much more) the same amount of stuff that wouldn't be inflationary and most of us get to do whatever the hell we want now. The few people who keep working get huge salaries.
That's not likely to be how things go exactly because we won't see AI replace all jobs in a flash but I think it shows that -- even though we absolutely need to worry about the impact of big shocks -- the mechanism by which AI just makes us inevitably poor doesn't fly.
Brian,
I shared your timely post with a great friend of mine (we were roommates in college and now he’s a super successful radiologist). My friend is worried (more like terrified) about his children’s future because of AI—how does something like this ever get started? I mean most people I can understand but a guy who graduated second in his class from NYU—this is crazy.
PS—During my last text conversation with my friend he highlighted the fact that unemployment numbers had spiked due to AI (maybe; maybe not). I thought better than to mention that millions of jobs are lost every month but that only net job numbers are reported). I think that much of the alarm is due to availability bias. Human beings seldom think past the seen to what is unseen. Keep up your great work!
Discount any intuited ideas about AI. It itself is science fiction out of The Cybernetic Imagination in Science Fiction (Warrick). The tech is fatally flawed from its arbitrary origins. There's no path around it.
https://substack.com/@eventperception/p-182707220
The comparative advantage argument depends on a scarcity of both types of labour.
In your example AI can do both health-care and coding better than humans, and costs less for both.
So you just use 2 AIs, as there are arbitrarily many instantly available,[1] for the cost of electricity and data centre amortisation.[2] Any job that is worth that amount gets done by AI.
Comparative advantage fails in conditions of an arbitrarily large labour surplus on one side.
1. There may be brief delays when all AI capacity is fully utilized, before equilibrium is achieved. Then more datacentres come online.
2. Because of competition among AI firms, the price falls to the costs of the inputs.
The "costs less" part seems to be an assumption on your part? As I read the example, the price of human labour for in-person care would fall below the cost of electricity and data centre amortization for that same type of artificial labour. I tend to agree with your conclusion, though, that in practice, humans would just be pushed out of such professions, much like human computers lost their jobs to electric computers.
I took 'better' to mean "costs less and is at least the same quality".
Yeah, I don't know what "10x better" is supposed to mean in a comparison between humans and software, but if there's one thing I've learned from reading this blog, it's that prices should be a model outcome, not a model assumption.
Doomsday prediction meets equilibrating tendencies (Kirzner’s market process). Short term pain is possible. Long term collapse is not.
Even in the Citrini report, they did not fantasize collapse. Every individual company in the S and P 500 could lose 30%, every loan could default due to suprise deflation, and still we could rebuild the market structure and move on. There could really be a lot of short term pain.
The extent of pain depends on whether or not market forces are allowed to adjust (the relative price mechanism). So far all significant busts in US history were caused exactly because government intervention prevented such adjustments. And no, not all loans will default. Didn’t happen in Japan deflation, will not happen here.