Adaptive Prosperity: Rethinking Economic Success in the Age of AI

The piece I started with was Forecasting the Economic Effects of AI by Ezra Karger, Philip Tetlock, and collaborators. What drew me in was not just its topic, but its framing. The paper asks a straightforward question: what might AI do to the economy? That is a serious question, and one worth asking. But while reading it, I found myself nudged toward a different one: what should an AI economy optimize for in the first place?

That difference matters. A paper can be perfectly reasonable in asking about expected GDP gains, labor-market disruption, or productivity effects. But once we begin thinking about the design of an AI economy, GDP alone starts to look like a narrow compass for a landscape that is much wider, messier, and more human than a single aggregate number can capture.

GDP is a durable metric because it is measurable, comparable, and often correlated with useful things: infrastructure, health, education, and state capacity. It is not a bad metric. The problem is that it is too often treated as “the metric”.

That becomes a mistake in an AI economy. Two countries could add the same amount to GDP while having very different structures underneath that growth. In one case, a handful of firms capture most of the value. In the other, millions of firms and workers make small gains that compound across the economy. On paper, the two outcomes may look similar. In practice, they are worlds apart.

This is why I think the deeper objective is not just growth, but adaptive prosperity.

An economy should not only be judged by how much it produces. It should also be judged by how well it can reinvent itself.

That sounds abstract, but it is a very practical idea. Modern engineering rarely optimizes for throughput alone. Distributed systems are designed for resilience, redundancy, fault tolerance, and graceful degradation. A system that is slightly less efficient but far more adaptable often wins in the long run.

Economies are similar. If the next twenty years bring repeated waves of technological change, the winners will not necessarily be the economies that predict the future best. They will be the ones that make it easiest for people and organizations to respond to it.

That means the real question is not just how powerful AI becomes. It is how cheaply and broadly that power can be adapted.

A lot of AI discussion centers on frontier capability. Which model scores higher? Which lab leads? Which benchmark moved?

Those are useful questions, but they are not the whole story.

History keeps showing that technological capability and economic transformation are different things. Electricity existed long before factories were redesigned around it. The internet existed long before it became commerce infrastructure. Cloud computing became valuable when entrepreneurs stopped treating it like rented servers and started building entirely new products on top of it.

AI seems likely to follow the same path. The frontier will move first. The broad economic effects will arrive only when organizations redesign workflows, retrain workers, change products, and discover new business models.

The bottleneck is not raw intelligence.

The bottleneck is adaptation.

I do not think democratizing AI means every graduate student, civil servant, or small business should try to host a massive frontier model on their own machines. That would be technically wasteful and economically unnatural.

We do not expect every household to generate electricity or every firm to manufacture its own semiconductors. Some layers centralize because scale matters.

AI should probably work the same way.

A handful of organizations will train the largest frontier models. That is fine. The more interesting economic question is what happens above that layer.

Imagine an ERP company. It does not need to train Kimi K3 from scratch. It needs to understand procurement, invoicing, approvals, compliance, support, and user flow. If it can take a powerful foundation model and adapt it through SFT, distillation, retrieval, routing, and workflow integration, it might cut completion time by 60% or 70%.

That would be a huge economic gain.

The company did not improve frontier benchmarks.

It improved value creation.

That is the layer that matters.

I think the most useful way to understand the AI economy is as a stack.

Most of the economic surplus is probably created in the middle layers, not at the frontier itself.

That matters because broad prosperity is rarely created by the core platform alone. It is created by the enormous number of things people do once a platform becomes cheap enough to use.

A good analogy here is Jio.

Jio did not invent mobile internet. It reduced the cost of access to it. That single move made it economically viable for millions of people and businesses to experiment. The result was not merely more internet usage. It was a wave of complementary innovation: digital payments, creator platforms, logistics, SaaS, e-commerce, telemedicine, edtech, and many more forms of internet-led business.

Jio is useful because it shows how democratization can work in practice.

It does not mean handing every user the full stack.

It means lowering the cost of experimentation so that the system can discover value in places nobody predicted.

That is exactly what AI may need.

Open-weight models are often discussed as a hedge against vendor lock-in. That is true, but it is only part of the story.

Their bigger contribution may be that they lower the cost of specialization.

If a company can safely adapt an open model to its own workflow, then thousands of specialized models can emerge. One for procurement. One for underwriting. One for literature search. One for clinical navigation. One for compliance. One for customer support.

No frontier lab can predict all of those use cases. And it should not have to.

Innovation in a healthy AI economy should be decentralized. The frontier should exist, but the edge of the platform should be where most experimentation happens.

One reason AI matters so much is that it saves time, not just money.

If an accountant saves three hours a week, GDP may not immediately register a dramatic change. But the accountant feels the change. Multiply that across millions of workers and the effect becomes enormous: more experimentation, more learning, faster shipping, better service, and more room for new ideas.

Time saved is not a trivial convenience. It is economic potential waiting to be redirected.

The best AI systems will not just answer questions. They will reduce friction in the exact places where work gets stuck.

I would add labor retraining to the macroeconomic objective function.

That is not just a social policy concern. It is a structural economic one.

If workers can transition quickly into new roles, firms can no longer use scarcity of skill as a moat. They must compete through better products, stronger execution, better customer experience, and genuine innovation. In other words, retraining weakens the kind of moats that come from lack of competition or lack of skilled labor.

That seems like true value creation.

The deeper idea is that labor should not be treated as a fixed input. Human capital is renewable. Economies should be judged partly by how efficiently they convert obsolete skills into productive new ones.

AI can help here by lowering the cost of personalized learning and continuous training. The future may reward workers who can reinvent themselves repeatedly, not only those who mastered one trade and stayed there.

This leads to a larger conclusion.

GDP is only one dimension of economic success. It does not tell us who captured the value, how resilient the system is, how much experimentation is happening, or whether people are being continuously re-skilled.

An AI economy should be judged on a broader set of objectives:

GDP still matters. But it should be one outcome among several, not the only score that counts.

If I had to name the framework, I would call it adaptive prosperity.

The argument is simple:

AI will not create prosperity just because it makes models smarter. It will create prosperity when it makes adaptation cheaper.

That means the societies that benefit most will not necessarily be the ones that build the smartest frontier models. They will be the ones that enable the widest number of people and organizations to turn general intelligence into domain-specific value.

The real economic advantage of AI may therefore lie less in who invents intelligence and more in who makes it easiest to build with it.

That is the lesson I took from the paper I stumbled across. It began as a forecast about the economic effects of AI. It ended, at least for me, as a question about the kind of economy we want to build around AI.