996 Won't Save You
Software Problems
In my circle of people, which is admittedly far outside of the tech hubs, I am the person most likely to see things as solvable by software. I make score keepers for card games, I make scheduling apps for community service, I made a web app for my wife's surprise birthday. I am software-pilled. I have internalized the "AI eating software eating the world" mantra at work and at home.
But here's the real issue, even given that, my own preferences betray that mantra. I love craft. I will listen to Brandon Sanderson talk about his Cosmere for hours at a time and buy his books in leather, I love my Thursday sneakers, I pay for a trainer to tell me how to lift, I go to the Indiana State Museum IMAX to see Nolan films in 70mm, I listen to anything Lin Manuel Miranda writes, and if Nilay Patel makes a podcast, I will subscribe.
I hope each of these individuals can use AI to enrich their life and allow them to focus on what makes them happy (though I know for a fact Sanderson and Nolan are as anti-AI as they come) but if they abdicated their thinking, their design, their craft to AI- I would hate that and probably stop my patronage of their work.
This is because not all of my problems (or demands) are software shaped. In fact, I would say MOST of my demands are people shaped; problems where my relation to the fulfiller of the demand is just as meaningful as the fulfillment itself.
My wife and I joke about people who post on social media that marriage is so hard, but we also understand all relationships do require work, sacrifice, and compromise. I could just subscribe to Grok and have a girlfriend who'd never complain when I forget an event on the calendar.
My awe, and the awe of the people I know, is not merely due to the quality of the output, but the craft of it. The human stories, the soul of the individual imprinted on the output, and the Journey which comes Before the Destination (First Ideal- IYKYK).
This belief that people's work is inherently immune to automation is particularly damning to a particular set of ideas coming from the Valley, that soon AI will take all of our jobs and that in its wake will be the emergence of some permanent underclass which will live off of either UBI or some other wealth transfer mechanism existing to drive consumption for the overclass.
The natural extension of this is a performative grind for the VC's in order to escape the underclass in a practice borrowed from the Chinese tech sector- 996: working from 9 AM to 9 PM, 6 Days a week.
My belief is that recognizing that people-shaped problems exist is sufficient to understand that this is nonsense. But I want to take you through some economic principles that I believe are being ignored due to the one-track mind of these tech-bros. I am not trying to hurt anyone's feelings or dissuade anyone from working hard, but there is a severe harm in propagating the core ideas and the buck needs to stop.
A Steel Man for the Permanent Underclass
In full transparency, I have really struggled with the idea that this is even remotely possible through free market- democratic means. Certainly, you could imagine a world wherein there is some electoral capture in which the governmental elite become semi-permanent ruling dynasties- this has happened cyclically throughout history and could realistically perpetuate beyond. But setting aside the hostile takeover of fair and free elections, let us take seriously the idea that a small set of capitalists could seize the means of production, leaving it impossible to be anything but an underling to the progenitors of the likes of Amodei, Altman, and Thiel. (I do not name these individuals to be glib, in my view, this argument works best when just a few model makers, and maybe 1 data infrastructure company with network effects are in a winner take all scenario.)
The best version of the argument goes like this: AI is going to be better at every job in the world, a country of geniuses in a datacenter, and once we reach some level of diffusion, we will likely see that most of the economically valuable work in the economy will be consumed by AI and for every new type of labor which comes up, it too will be consumed by the AI because AI will be better than humans at it- that unlike every industrial revolution before it, that the automation created by AI will displace without creating sufficient complementary demand.
People Shaped Problems are a Disease for Accumulation
Accumulation of wealth is the mechanism for inequality in the permanent underclass model. The assumption would be that the productivity of these small number of firms would be so high that they would naturally accumulate work and then also wealth. The type of work they can accumulate, however, are definitionally the software shaped problems. Meaning, that insofar as there are people shaped problems, the accumulation becomes infected with what is called Cost Disease (Baumol & Bowen, 1966).
Have you wondered why college and Broadway productions have inflated more than the rest of the economy? It is because, in both cases there is little room for automation which means that productivity of the laborers cannot make up for inflationary pressures thus the cost is linearly connected to the cost of living.
Admittedly, college inflation is multi-faceted. There is some debate in the field on the degree to which cost disease accounts for the increase in tuition costs vs administrative bloat and facility arms races.
As AI firms accumulate wealth and their employees' cost of living rises with the broader economy, the cost of every people shaped service they consume rises proportionally. The billionaire still needs a doctor, a teacher for their children, a therapist, a personal trainer. Those costs scale with wages, not with software productivity. And the firms themselves will need to hire for people shaped problems like lawyers, sales, HR, etc., in greater degree as the work that demands those functions scales.
Jevons Paradox
There is a competing and complementary force when you look at the second order effects of automation and productivity gains. As the cost of producing a good falls using automation, demand for that good tends to rise so dramatically that total resource usage increases rather than decreases. This has been true for every transformation in recorded history and it should be noted that the "this time is different" narrative is nearly as universal.
Why does this happen? The technical term is "latent demand". What we observed in the earlier Industrial Revolutions was that as textiles, steel, plastic, etc became more affordably produced, the demand for those finished goods increased and today we have more individuals employed in those supply chains than ever before.
The most common counter example to the labor input for Jevons Paradox is agriculture.
While it is true that the number of farmers worldwide is dramatically decreased, the total number of workers to manage the world's agrifood system is MUCH higher than it was even 30 years ago, with an estimated 1.23 BILLION people employed worldwide.
Looking at software in particular, the demand for code is not just the code itself but also the design, the product roadmapping leading to the correct features, the governance, and the support to ensure product works for you. You might say, "well couldn't you have an AI to do all of that for personalized software"? Theoretically yes, but this mistakes where the bottleneck is. Framing a problem as software requires first observing it, and observation is situated: it lives with the person inside the problem, who feels the friction but usually cannot frame it. The AI can frame, but it cannot observe — it has no access to the friction until someone puts it in front of it. So the two halves sit with two different parties. The person who can see the problem mostly can't specify it; the model that can specify can't see it. Closing that gap, eliciting a half-formed observation from the person living it and turning it into something a model can act on, is the work. That conversation, of empathy and understanding and accountability, is a human-shaped problem.
It is my view that the complementary demand for software is actually much more expansive than it might first appear. For example, my work is greatly informed by great sci-fi and fantasy works and as my income increases, so does my consumption of literature and cinema.
Lump of Labor Fallacy
I just bought a home and there is this little urge in me to believe that if I can just get through my honey-do list then I could relax forever. But, this is obviously false because that list keeps getting longer, infinitely.
The root of this fallacy is in zero-sum thinking. This type of thinking is as pervasive as it is corrosive. The idea that there is a fixed pie and if my neighbor gets a slice then that is less pie to me. Zero-sum thinking is an evolutionary outcome of scarcity. The bison was only so large to share with the tribe.
But zero-sum thinking is a heuristic that is blind to growth. As more is possible, more will be wanted and if we have the opportunity to reallocate resources to other things, we will (if I did not have to do the dishes every day then maybe I would put some higher maintenance plants in my flower bed).
Compute Constraint
Until now we have been looking at demand, now let's look at the supply because this is the last and final nail in the coffin of the subsumption of labor by AI: compute is and will always be a scarce resource just like labor is. It may be cheap. It may be cheaper per unit of work than people (though I actually doubt this will be true of certain models- you may be paying for scale and speed) but there will still be demand that AI cannot fulfill because latent aggregate demand is essentially infinite.
The cost of compute in the long run is actually very difficult to track. My guess is that while the cost for units of work (the actual tokens) drops over time, the purchasing of compute units increases. This is again due to the increased demand due to Jevons paradox. I will discuss this further in a later essay; Java coders be warned, it's coming for your baby.
So to put it together, not only are there problems AI cannot take care of, but the complementary work produced because the increased productivity of AI will increase the demand for those people-shaped problems, and there are some problems where humans will have a comparative advantage in some domains due to the compute cost rising as compute is allocated towards highest value problems first, leaving a long tail of economically valuable tasks for humans to perform.
"In the long run, we are all dead"
This is not to say there will not be a transition period of challenge. The pressures of efficient capital allocation will push leaders in and out of SV towards automation at status quo levels of output. And if they are your McKinsey Comes to Town types, that means layoff rather than creative reallocation.
But it does not have to be this way. IKEA reallocated their customer service team to design consultancy and created a billion euros in net new value for the firm. If more leaders of this caliber are calling the shots, we will see productivity translate into GDP gains into increases in the velocity of money into more GDP gains into wage gains.
If we continue with uninspired, one-track-mind leaders making these calls then we will see long periods of extreme inequality under laissez-faire capitalism and in that way permanent could be true for the lifetime of these grindsetters.
But even if that is the case, the reaction is inappropriate. The failure rate of these start-ups is enormous and if we take the labs seriously, then this is going to be much more of a winner takes all than the tech-bros are willing to admit.
But I think both are inherently incorrect. Standard Oil was supposed to be the winner forever, but that market is now healthily competitive- whether that happens through governmental intervention or market innovation (e.g. KV cache innovations from Deepseek).
Another Call for Multidisciplinary Approaches
I believe this myopic view of the world demonstrates why we should not abdicate the governance of AI to Silicon Valley. They are engineers and consultants.
Most of these builders in the Valley are CS grads from Stanford, UCLA, MIT, etc. Truly exceptional and intelligent. But they cannot shape our world. We need philosophers, classical work experts, librarians, and, yes, economists, to help us if we are going to build the version of the future we actually want.
The future I want is balanced. I want to write and code. Teach and build. Work hard and relax. To work at home, and in the office, and in the community. Give me leaders who believe in that.