A serious consideration of what a post AGI economy and labor market might look like
My usual programming focuses more on what is possible now, and what futures seem most likely, given the current trajectory. But every now and then I try to think about extreme scenarios.
The main categories of debate about AI can be summarized in various ways, but here is my high-level framework.
It’s easy to run into people who believe AI will [at some point] be able to do everything a human does, but better. It is difficult debating someone like that, because the answer to everything is just “AI will do that too.” In their mind, things like Jevons Paradox and Comparative advantage, and even GDP itself become meaningless and irrelevant because AI and robots will provide everything to everyone for free. That means the answer to the blue branch in my map above is basically, “yes”. And the answer to everything in the red branch is, that will happen to the most extreme degree you can imagine. This sums up the worldview:
If automation drives the price of many or most goods to zero, then GDP growth will be a very bad proxy for welfare growth.
If robots make a free pill that makes you live forever, GDP does not increase. If robots send whatever food you want to your doorstep for free, GDP does not increase.
This poster does talk in terms of ‘if’, not ‘when’, which is good. I’m not at all throwing shade at this post. From an economics standpoint, it is an interesting premise. The obvious questions it raises are: why would the robots do that? How could they do that for free? Solar powered robots design and build other solar-powered robots in solar-powered factories, and these robots provide everything for free to humans, and both the humans who control the robots (?) and the humans getting everything for free are totally cool with this arrangement1. If you put it that way, it does start to come across as a very unlikely scenario. Various other branches in the map I shared above seem to strongly conspire against this outcome.
As readers of my newsletter will know, I don’t believe that AI is going to be as disruptive as many viral pieces argue. I’m a ‘slow timelines’ guy, and a believer in AI as Normal Technology. But I am actively keeping an open mind, ready to change my mind if facts change, and positioning myself to catch this as early as possible2. An OpenClaw agent, Maggie, is doing almost all the work to grow and build my sideproject Magicdoor.ai. Without my daily input, Maggie would currently make zero progress, and destroy the app within a week. But if the time comes that AI Agents can make good decisions autonomously, I plan to be among the very first in the world to find out.
This also means that every now and then, I will suspend my disbelief and take the most extreme scenarios seriously. So, let’s engage seriously with the idea that automation could drive the price of many or most goods to zero. This means revisiting the microeconomics branch of the map.
A year ago or so, there were few economists weighing in on what would happen if AI really does become capable of doing every job better than every human. Some of the key economics concepts involved like the Jevons Paradox and Comparative Advantage were explained, and most people following AI will have heard of them by now. But it took time for deeper thinking, leaving us with engineers and influencers dominating the conversation. But economists have increasingly started to weigh in, and the debate has increased in quality a lot since then. The Citrini blog post was an important catalyst, as it so blatantly violated basic economics and accounting truths that it ‘woke up’ a lot of economists. As part of this debate, an economist named Alex Dimas posted an essay that got a ton of positive response (and some counters as well):
His point is that with industrialisation, labour shifted from farms to factories. With growing automation, it shifted to the services sector. If AI brings automation to services, it could shift to where human touch is part of the ‘product’. The example he opens with (to save you a click at this point in the story3) is that Starbucks has been reversing a years long drive toward increasing automation. People like being served by humans, and the extra cost is worth it. This lines up with one of the things I’ve been predicting for almost two years now: that AI will put a premium on human generated stuff. Alex provides some experimental evidence, shown in this chart. People are less willing to pay for mass produced goods in general, and even less willing to pay for those created by AI.
I don’t know about you, but I find this extremely obvious. On first glance this creates a fairly narrow list of professions that are safe from automation: Everything in human to human care, a lot of education, and artists and performers. The essay received pushback that spend on human performers historically has not grown with increasing automation. By and large, as people get richer, they have not spent much of that extra money to watch performers. But the category of services where the human touch is the product is much broader. Tailors, chefs, sommeliers, bakers, butchers, baristas, craft brewers, DJs. Alex also lists others like experience designers and community managers. I would add most trainers and coaches, but also bodyguards, drivers, personal assistants, VIP support agents, consultants and even lawyers to this list, and I could keep going for a while.
What will be scarce
This relational component surfaces a key gap in the ‘task based’ model of job replacement. A job is not just a bucket of tasks, where the job disappears if machines can do all the tasks. So far, I have mostly thought about this in terms of task automation being a bad predictor of job automation. But I had not thought as much about the relational/status part of it. Robot coffee has existed since at least the 70s but it's still nicer and higher status to get a barista to make it. Mass produced suits have existed for decades, but people brag about knowing their tailor by name and recommend them to their friends. People spend more money to buy bread from an artisanal baker instead of mass produced bread. People even make their own bread4. There are even still people who make a living as nearly completely unmechanized farmers. On a trip to France a few years ago, I visited a wine maker who did not even have a lawnmower, but used sheep instead to cut his grass. The more high-end you go, the more human touch things get, and there is also a meaningful part of the economy that rejects technology ideologically, or simply for fun (I recently bought another vinyl record player). That is not going to change, and in fact I think human made things will only become higher status as machines produce more stuff. But there is a supply side at play here as well.
So now that have a super long list of products and services where humans are likely willing to pay more if those are made by other humans than if they are made by machines, what does that post-AGI economy start to look like? If we buy into the idea that there will continue to be demand for human made products and services, even if strictly speaking robots can deliver all of those at better quality, how scarce will those things be. If robots took my job, I could become a DJ or open a bakery. But the demand for artisanal bread is finite. If many more bakers start, competition should bring prices down and at some point make it unprofitable to be a baker.
And it’s worse than that. To stick with bread, I am buying expensive sourdough because I don’t like mass-produced supermarket bread. But I started making my own sourdough partly because a loaf is 15 SGD or something. With plenty of time on my hands, I could bake bread that was just as good for $3 per loaf in ingredients. If an AI robot powered bakery truly would sell the same quality bread for say $7 per loaf, how many people would still opt for the human made bread at double the price? would I still do that? To get to the bottom of this, we’d have to open a lot more doors and increase complexity. The supermarket has sourdough. It is ok, but it’s not fresh enough. It had to be transported. Will the robots be distributed everywhere, so that they can come to my house to bake the bread in my own oven? Or will every town have an artisanal robot bakery? At what point does it just not make sense for robots to do something?
This is comparative advantage again. Bread is a perfect example. Armed with one great book, after baking 10 breads or so, I was able to make all kinds of breads in my own kitchen, and they tasted great! If I can bake breads so easily at home that are 80-90% as good as the best artisanal bakery breads, how much upside is there to bread quality? Even if robots are better at it than humans, the gains are small. There isn’t much room for robots to make bread better or faster. You can’t bake one bread 10x faster with ten ovens. You can accelerate the proofing process by adding more yeast, but if anything we bakers are trying to slow this down by proofing overnight in the fridge because a slower rise makes the bread taste better. These factors give humans a strong comparative advantage over robots as bakers.
There is also a floor on costs. No matter how many robots there are, producing flour will always require farmland. Land is scarce. For anything retail, rent is a major expense. Land again. Would robots not need to pay rent on a bakery? Why? The big question I keep coming back to in my mind is about those statements that robots would just provide anything you want for free. Maybe I’m shortsighted here, but I can’t imagine how that would work. Artisanal bakeries operating costs break down roughly like this (according to Gemini):
Raw Materials (28–30% of revenue): Higher costs due to quality ingredients (organic, local, specialized grains).
Labor (35–45% of revenue): This is the highest expense, as artisanal methods are labor-intensive.
Energy (4–6% of revenue): Specialized, slow-fermentation techniques and high-heat ovens are energy-intensive.
Rent (8–15% of revenue): High-traffic locations, necessary for viability, often require high rent.
So, even if the robot was free, the bakery could operate at most at 45% lower cost by eliminating all labour, which if passed through fully to customers would still not make the bread anywhere close to free. Even if everything throughout the entire supply chain would be handled by robots with zero operating cost, the bread would still not be completely free. What am I not seeing that the believers see?
Circling back to the Starbucks example, what would it look like for robots and AI to run a Starbucks. It is already possible to automate nearly all aspects of a Starbucks now. But Starbucks has been reversing some of their automation projects, as explained in this year old article. The company’s just released financials were strong, with ‘experience focused’ store revamps credited as a key contributor to the turnaround. What is the AGI case for Starbucks? Since automating the stores seems to have resulted in lower sales, would a humanoid robot making drinks, talking to customers and writing handwritten notes on the cups be different? Is that what we should be picturing here? Or is the argument that nothing about service level matters if the coffee is nearly free? And does Starbucks provide coffee? Or more than that? A place to work or study, for example.
The labour share of the economy
What percentage of GDP goes to wages? In the USA it has reduced from mid 60s % in the 1950s to 53% today. It is currently 43% in Singapore, and 55% in the Netherlands5. Where would that number be in 20 years from now? Some extremists would probably say it goes to zero. I think that’s clearly absurd. Could it go to 20% or even lower than that? And would that be because employment declined? Or because the ‘robot economy’ comes on top, causing the capital part of the economy to outgrow the labour part massively? It’s hard to answer questions like this even retrospectively because the world is so complex. This whole debate was triggered by this survey of various experts; economists, AI experts, and laypeople. In their responses, every group forecast GDP or labour participation (the % of people who are employed) to change quite modestly even in the case of rapid development of very powerful AI.
On the face of it, this seems like a strange position to take. If AI and robots can run any company at the level of a top CEO and do any manual labor at the level of the best humans, and do all of that at a fraction of the cost6, wouldn’t that result in drastic changes? Alex Dimas and others try to explain how this position could make sense based on economic arguments. On the other hand you have a group who argue that these economic arguments would break down in this scenario. Personally, I believe that there is a lot of merit to Alex’s arguments.
Closing thoughts
There are a few things I learned from this debate, and that I want to share as potential key takeaways:
The cost of AI / Automation is a critical constraint, likely to be more important than the capabilities. It is a critical input to the comparative advantage equation. A totally different economy results if robots+AI can substitute labour at 1/10th the cost compared to if it’s half the cost. It is going to be more difficult to get a robot to bake artisanal breads or pizza compared to getting them to do warehouse logistics.
For any given product, the quality difference is a critical input, again likely more important than the capabilities. In the case of bread or coffee, a robot might make them much faster than a human, but not better (absent malicious tricks like lacing them with some addictive compound).
There are many products and services where quality is a lot more complicated than the product simply being technically good. Things like trust, status, provenance, atmosphere, human attention, belonging, story, and authenticity factor into willingness to pay in ways that are non-obvious.
In a world where AI and robots can do everything better than humans, scarce things include: land, human input, raw materials, machine-time. Even if robots can drive massive efficiency gains through the entire value chain (mining, farming, manufacturing, transport, and construction), they won’t be able to drive even the cost of those things to zero. Land is almost certainly not going to be made less scarce by robots. Human input…well that is the whole question.
Together, this means that a thing can be technically automatable and still not become cheap, not become preferred, not be widely available, or not be priced near cost. There is a hidden chain of assumptions in the idea that AI + robots will drastically reconfigure the economy, that goes something like:
Every one of those arrows could easily be its own long post to unpack the assumptions and nuances underneath. All of that, together with the other forces slowing down AGI like political backlash, technological challenges, energy, makes me continue to believe we really don’t have to spend too much time worrying about this super advanced AI scenario.
It’s kind of cool as a sort of neo-communist world view, with robots and AI as the benevolent dictators, in full control of the means of production, and with such astounding efficiency that they can create everything for free.
‘Positioning’ makes this sound more strategic than it is. I know many people have FOMO and feel they “should be doing more with AI”. Let me assure you, I just got lucky that I am doing all of this entirely for fun. Yes, I’m working my ass off and sacrificing sleep to build more stuff, but I’m having more fun than ever.
Although the post from Alex Dimas is definitely worth a click and full read!
I am personally one of those who started baking sourdough during the Covid Pandemic.
There are various ways to measure this and there is disagreement about historical numbers. I’ve used the data from the international Penn World Tables dataset.
One more time: We are not talking about the likelihood of this scenario, only about what would happen if it occured.