Alright, we are back to seriously considering AI and economic scarcityAlright, we are back to seriously considering AI and economic scarcity

Talk about this a lot these days (#1514363, #1485162, #1486740); makes sense, everybody else does. And it is the thing that clashes up against the most foundational of economic models (the scarcity component of the marginal revolution)

Here's a long, complicated Econ101.5 story from Brian Albrecht about the economic returns to AI, the shit we all care about these days.

Brian's intro summary:

Two companies whose principal product is access to a model—or access to intelligence—are each valued at nearly a trillion dollars because the market believes that building the smartest machine is how you come to own the future. [...] We are beyond the point where we are asking if AI will be useful. It obviously will be.

"But the economic question is who will capture the returns, the benefits, the rewards?""But the economic question is who will capture the returns, the benefits, the rewards?"

Large reshuffling of economic value in the cards

The valuations suggest the big labs will, and the AI boomers and doomers (not the AI-will-kill-us-all doomers, but the AI-will-destroy-the-economy people) are in agreement on this. The boomers see a fortune waiting to be won by the companies that build the most powerful models. The doomers see the same fortune being taken away from workers, pushing labor’s share of income down.

Is what AI generates scarce? Will it have economic value?Is what AI generates scarce? Will it have economic value?

Infinity has already mostly competed away the scarcity value in music, writing, porn . #798342, #796401

See "The Economics of AI and Infinite Supply"See "The Economics of AI and Infinite Supply"

Deflation in the cost of LLM-inference:

"we have two competing forces. A fixed level of intelligence is becoming abundant, and abundant things are cheap at the margin no matter how useful they may be in total.""we have two competing forces. A fixed level of intelligence is becoming abundant, and abundant things are cheap at the margin no matter how useful they may be in total."

Nobody is paying nearly a trillion dollars because today’s frontier intelligence will remain scarce. Within a year or two, much of it will be cheap. Investors are betting that the leading labs can keep producing (and controlling) something that does remain scarce. Maybe that’s just the moving frontier, or other things like proprietary data, locked-up compute, privileged distribution, or the workflows that make their products habitual.

and Brian gives us an analogy for a recipe and making croissants:

Anyone can buy flour and butter, and anyone can read a recipe card. They cannot easily reproduce the timing and the touch that make the croissant good. The baker earns a premium because the valuable intelligence cannot be separated from the person who has it.

There's some non-obvious, non-replicable (or...?) thing that turns the abundant and cheap ingredients into a scarce and valuable consumer product.

Some of the old premium goes to customers in the form of lower prices. The remainder settles on whichever inputs are still scarce. Maybe there is still something to being on the corner with the foot traffic. Or maybe the oven (not written down) is the key. We could go down the list: the trusted brand, the financing, or the baker whose hands really are faster. Output rises, but the return to knowing the recipe falls. To stay ahead of this process, the baker has to continually come up with new recipes that aren’t copied and ahead of the market.
Replicable ideas allow us to produce more with the same resources, but the resulting wealth does not automatically accrue to the people who once held the knowledge.

"When intelligence becomes abundant, bottlenecks matter more""When intelligence becomes abundant, bottlenecks matter more"

and:

"this is Ricardo in a nutshell. The rent goes to the scarce complement.""this is Ricardo in a nutshell. The rent goes to the scarce complement."

So, what's scarce in an AI-fueled, LLM/server-dominated future?

Attention comes to mind. The SN home page is an auction #1406771

Chips and farms because, as Brian says, "the model still has to run somewhere. It requires chips, electricity, cooling, buildings, land, and time."

Electricity, and particularly the stranded/cheap energy #1466211

That does not mean chips and electricity will remain scarce forever. High returns attract new fabrication plants, generators, transmission lines, and data centers. It means they are among the clearest bottlenecks today. Markets respond. I think we’ve mentioned that in this newsletter.

LOL. yes, neat. And here's a statement you could spend a PhD on (#1515430, #1521292)

If you prevent the recipe from being copied, you turn a nonrival capability into something rivalrous, which can become a priced asset. All of the physical production has this, ASML or chips, for example, offers a physical version of this strategy.

Yes, scarcity: rivalrousness vs not. The entire intellectual property conversation #1516102

What about the workers?What about the workers?

That's what we all care about -- at least us in the commentariat, creating and amassing words for a living. (Machines seem to do this better than us, at at least hint/promise to, so yes we're afraid.)

Brian teaches us that workers sell an output, and it's critical to identify what that is:

A worker sells an output, whether it be a diagnosis, a legal brief, a repaired machine, a lesson, a design, a relationship, a judgment someone trusts enough to act on. She does not sell “tasks” in isolation.

Either: the worker captures more of the gains... #1485162

It becomes an input into the worker’s own production. It makes research cheaper, drafting faster, scheduling easier, or diagnosis more accurate. In that case, AI feeds her work. Inputs become cheaper. If the rest of what she supplies remains scarce, her return can rise. She didn’t do anything more, but her productivity increased, and she captured (at least some of ) the gains.

... or it replaces her, or basically the massively expanded supply competes away whatever value-add she once had:

all of her competitors (and new competitors) are also using AI. Maybe some of that is much more “pure AI” and we can think of the AI as the direct competitors. The knowledge that you had is no longer scarce.

"we have a horse race between AI as an input and AI as a competitor. It’s not obvious which of those forces wins.""we have a horse race between AI as an input and AI as a competitor. It’s not obvious which of those forces wins."

If AI lets many more people provide an adequate human service in a particular market, the premium attached to any one provider can fall even though the service remains human. This is the same ownership question we have been talking about all along. For workers, ownership often shows up as portability. Can the scarce human input walk away? Can it reach customers without the firm, platform, brand, license, or copyright wrapped around it?

tl;dr:

LLMs are here, they massively shifts around the economic rents (=what's scarce) in the production processes we're used to. It's unclear who will benefit and who will carry the cost, where the burden will fall etc.

Right now, we see that the scarce thing is physical, like chips, megawatts, and land. Sometimes it is generated or manufactured, like data, distribution, a copyright, a brand. And sometimes it is a person whose judgment or relationships cannot be separated from the service she sells.

"Scarcity depends on property rights. What is AI making abundant, what remains scarce, and who owns it? That is where the returns to AI go.""Scarcity depends on property rights. What is AI making abundant, what remains scarce, and who owns it? That is where the returns to AI go."