What Stays Scarce After AI?
AI looks a lot like cars in 1915 — plus the relational economy, the regulation that broke nuclear power, and whether we're actually in a meaning crisis.
“In times of change, learners inherit the earth, while the learned find themselves beautifully equipped to deal with a world that no longer exists.” — Eric Hoffer
Forward Deployed, Episode 5: Aligning Agents
with Noah Brier
I chatted Noah Brier on Forward Deployed to dig into how AI is reshaping organizational structure and what knowledge work actually looks like day to day in a post-agent world.
We both suspect agentic systems look more like companies than software. Agents execute well when problems are well specified but can aim at the wrong problem. This is the same problem people often have so why wouldn’t we lean on the near century of organizational theory we have on how to solve them?
We get into some of our favorite concepts and how they apply to AI including Pace layers (Stuart Brand), OODA loops (Boyd), transaction cost economics (Coase), and Cynefin (Snowden).
What Will Be Scarce
by Alex Imas
Over the course of a few years, Starbucks automated more and more of their coffee-making process. This is a very logical, business minded thing to do: margins were getting tighter and so they looked for costs savings. The costs savings worked, but they walked the whole thing back.
It turns out people don’t just go to a coffee shop for the best coffee at the lowest price. Those are bundled into a larger experience. The CEO concluded that handwritten notes on cups, ceramic cups, and “the return of great seats” were what drove satisfaction. I wholeheartedly endorse this. I am not much of a coffee connoisseur but I am very much a comfy coffee shop connoisseur.
It turns out that having a one minute conversation with the barista and them writing your name on the cup was actually very important to the value of Starbucks’ stock.
This is an example of a broader phenomenon that suggests something about how economies may be reshaped by productivity gains from AI. As people get wealthier, they don’t just buy more of the same types of stuff. They reallocate spending towards new types of goods.
2022 BLS data shows households in the top income quintile spent 4.3x as much in total as the bottom quintile, but the ratio is much larger in categories with a strong relational component: in-person dining, entertainment, and education.
My observation talking with high-net-worth investors is that as people get wealthier, the share of conversation about investing keeps shrinking. Part of this status signaling (only really rich people can afford not to think about money), but part of it is that the marginal amount of attention and money is better rewarded thinking about parenting, marriage, and finding purpose.
Imas argues that as AI grows the share of the economy devoted to these ‘relational’ services is going to grow. Economic forces once moved 40% of the American workforce off farms and into factories and can now move workers out of automatable commodity production and into sectors where humans are part of the product.
Once basic needs and capital allocation are handled, the binding constraint isn’t more money. It’s living well. And living well is largely a relational problem: “No man is a failure who has friends.”
The Architecture of Complexity
by Herbert A. Simon (1962)
There is a tendency (of which I have been guilty) to fantasize about “equal” or “fair” systems where all participants are equal. I was at one point, enamored with concepts like holocracy. Simon’s 1962 paper argues that these approaches don’t work. Complex systems in nature, biology, society, and engineering overwhelmingly take the form of hierarchies and there is good reason for this.
Consider the story of the watchmakers:
Two watchmakers, Hora and Tempus, both build watches of about 1,000 parts. Tempus assembles each watch as a single integrated process — if interrupted (say, to answer a customer call), the partly-assembled watch falls apart and must be restarted from elements. Hora, by contrast, builds his watches from stable subassemblies of roughly ten parts each, then combines ten subassemblies into a larger subassembly, and ten of those into the whole watch. When interrupted, Hora loses only the current subassembly.
Simon’s arithmetic shows that with a 1% chance of interruption per step, Tempus takes roughly four thousand times longer than Hora to complete a watch. The general result: evolution of complex forms depends critically on the existence of stable intermediate forms, and time-to-assemble grows with the logarithm of the number of elements rather than linearly.
This made me think of Saras Sarasvathy’s notion of effectuation: the idea that entrepreneurs build from the means at hand (what you know, who you know, what you can already do) rather than working backward from a target. You can build a system, a business, say, much faster by combining fleshed-out subsystems, rather than starting from scratch.
AI, Tractors, and the Productivity Paradox
by Sachin
From 1903 to about 1950, farmers used their cars in ways manufacturers never intended. They blocked up the hind axle and used the power of the engine to run things like corn shellers, saws, hay balers, and cream separators.
One rancher used a Cadillac to shear sheep and a Maine farmer used his car for so many jobs that he got into a dispute with the tax assessor who couldn’t decide whether to classify it as a pleasure vehicle or agricultural machinery.
The car gave people access to a new and powerful general purpose technology, the internal combustion engine, and they extended its use into all sorts of tasks that it hadn’t been explicitly designed for.
Eventually, organizations like Ford figured out how to take all the scattered approaches and designs and embed them in a single organization at scale to make things like cars and tractors. Effectively, companies productized the various ‘kit’ use cases which ‘indie hackers’ had been doing in bespoke and sub-scale ways.
Many general purpose technologies went through similar ‘kit’ phases that are poorly understood because so little was written about them and no institutions were built around them.
Sachin’s argument is that AI is in the same place cars were around 1915. Robert Solow’s 1987 line about seeing the computer age everywhere except in the productivity statistics is back but this isn’t unusual. Gains from a general-purpose technology lag investment by a decade or more.
My favorite example: In 1969, Neiman Marcus ran an ad in their Christmas catalog for a Honeywell personal computer. It was priced at $10,600, about $85,000 in today’s money, and weighed over 100 pounds.
For that you got the computer, a cookbook, an apron, and a two-week programming course. The tagline was perfect: “If she can only cook as well as Honeywell can compute.”
We are very much in the “kit phase” of AI with custom GPTs, niche agent scaffolds, individual tinkerers wiring LLMs into their own workflows. It’s not clear what the ‘productized’ version of these tools will look like.
Sachin takes a Coasean angle on that which is very much in line with how I’ve thought about it. LLMs erode the firm’s advantage market transaction costs (finding specialists, evaluating contractors, writing enforceable specs) are collapsing fast.
I have been thinking about it as a constellation of technologies. The internet collapsed discovery costs. LLMs are collapsing coordination and contracting costs. If crypto and stablecoins eventually collapse enforcement costs, you get more, smaller firms. Is there some productized layer on which all these run?
The Nuclear Renaissance
Works in Progress Podcast
In the 1950s and 60s, there was a huge push across the developed world to build nuclear power. The reason was simple: nuclear was the cheapest electricity anyone had ever produced. In present day costs, it was about $500 per kilowatt of capacity, three cents a kilowatt-hour, and a nuclear plant could be built in roughly four years.
The reason it was cheap was simple.
You could use the same turbines boiling the same water as a coal plant, just heated a different way. So the whole supply chain for that was already well-established and functional.
The cost savings came from the fact that uranium is roughly 300,000 times more energy-dense than coal, so fuel is a trivial 2% of the cost of running a nuclear plant whereas fuel is about half the cost of running a coal plant. It costs a lot more money to dig up mountains of coal and ship it to a power plant than it does to produce uranium.
Ergo, nuclear power should be about half the price of coal generated power. For the first decade, it was. Then, something happened.
The popular narrative that I had understood was that nuclear production declined in response to events like Three Mile Island and Chernobyl. The reality was those were part of the picture but the bigger problem was nuclear got a lot more expensive.
The most recent American plant, Vogtle in Georgia, came in around $16,000 per kilowatt and took 27 years to plug in. It was thirty times more expensive and seven times slower than plants built half a century before.
Nothing about the engineering got harder. What happened?
The cost explosion was regulatory and it traces to one mechanism: ALARA which stands for “as low as reasonably achievable,” a safety regulation adopted in the 1970s.
ALARA obliges an operator to spend on any safety improvement its profits can absorb. This means any additional profit has to be invested in additional safety even if it’s not a material risk.
That means as oil prices rose, and two oil shocks hit right as ALARA came in, nuclear’s profit got eaten by new compliance costs. As energy security has become more of an issue, there’s a move back towards nuclear in some places but the compliance and regulatory piece will have to get worked out.
Meaning Lost, or Muddled by Metaphysics
with Andrew Conner and David Chapman
Is the “meaning crisis,” the modern sense that we’ve lost the sources of meaning that religion and tradition used to supply, actually a loss at all?
The cognitive scientist John Vervaeke says yes. His account runs a 2,500-year arc, from a world where people were embedded in a meaningful cosmos to one where Descartes and Kant left us sealed inside our own heads. In his view, we lost something real, and the repair is to rebuild it deliberately, through practices that train the kinds of knowing that argument alone can’t reach.
Vervaeke even names the toolkit: DIME — dialogical, imaginal, mindful, embodied.
Dialogical is real conversation over debate: Socratic dialogue, or “circling,” where a few people attend out loud to what’s happening between them. Imaginal is working with image, myth, and symbol — active imagination, meditative reading, internalizing a wise figure until you can ask what they’d do. Mindful is the meditation family: sitting with the breath, watching attention move. Embodied is practice that lives in the body — tai chi, yoga, martial arts. The bet is that meaning comes back through doing, not through believing the right things.
I have done some version of all these practices and found them helpful. Chapman’s move is to reframe and argue that we didn’t, in fact, lose anything. Rather, we inherited a wrong idea, traceable to Plato: that for a meaning to be real, it has to be perfectly definite.
Hold that assumption and you get whipsawed between two errors. Eternalism says meanings are fixed and absolute; nihilism says that because nothing is fixed, nothing means anything. Eternalism more or less worked at a societal level until post-modernism showed up and spun everyone out into nihilism.
Both philosophies are problematic in that Platonic way: they refuse the obvious middle, where meaning is patterned and fuzzy at once. It’s real but not in a sharply defined ‘Platonic’ way.
If you’re active on any of these platforms and want a daily(-ish) thought between newsletters, you can follow me on Substack Notes, Twitter/X, or LinkedIn.
A humanities nerd would point out that the idea of the world having "lost enchantment" is pretty common, popping up in various guises since the 18th century (Romanticism, Eliot's "objective correlative"). I'd blame Kant too, but he's more a symptom than a cause: the Early Modern mania for classifying, describing, and separating, key to the scientific method, tends to strip away *overdetermination*, the idea that a thing is multiple things at once. (The trivial example: a rose as a symbol of love. Beautiful; doesn't last long; can pierce you if you touch it.)
This also cut apart the Great Chain of Being, the medieval/Aristotelian attempt to classify based on observed or occult similarities between apparently distinct things. For instance, the sheep is an animal wool; cotton is a plant wool; and there was presumably a sea wool that someone would haul up in a net sometime, even if it hadn't been hauled up yet. There were sea-bishops, the marine equivalent of terrestrial bishops. (These seem to have been various skates, which sort of look like a mitered face when seen from underneath.) And so forth and so on. These correspondences, most of which were hidden, needed to be ferreted out by the scholar...or the magician, whose power came from his ability to, say, focus the power of the Sun via copper goblets of white wine, bright clothing, talismans made of gold, etc., to achieve effects within the Sun's dominion.
It's interesting, and I must say a little amusing, to see a cognitive scientist jump into this. One has to admire his bravery if nothing else.
Please allow me my two nickels.
1. AI currently feels similar to where computer hobbyist's were in the late 1970's; under defined mission and a needy toy.
2. The real growth of AI is waiting for another great marriage like the one IBM and Microsoft had. My personal view is it will be a transformation of the "data center"; the data center is like the oil lamp prior to Edison, nice but dim.