The Price Discovery Problem in the AI Debate
When there's no way to price the future, all we have are vibes
It is tough to make predictions, especially about the future. — Yogi Berra
A new paper surveying economists, AI professionals, superforecasters, and the general public on AI’s economic effects landed recently, and Tyler Cowen flagged it on Marginal Revolution with his usual economy of words. The headline finding will not shock anyone who has followed Cowen: economists expect substantial AI capability progress by 2030, yet their macro forecasts remain stubbornly close to historical trend. Even in the paper’s “rapid” scenario, in which AI surpasses humans in most cognitive and physical tasks by 2030, the median economist still lands at 3.5% GDP growth, labor-force participation at 55%, and the top decile owning 80% of wealth by 2050. This is neither normal nor a singularity.
Meanwhile, technologists and venture capitalists continue to talk as though sufficiently capable AI implies civilizational rupture on a much shorter timeline. The gap between these two camps is now one of the most consequential disagreements in capital allocation and policy, and it is mostly conducted as a vibes war.
The Disagreement
The paper is careful to note that the disagreement is not about capability. Even holding AI capability scenarios fixed, respondents still disagree enormously about economic consequences. Economists who accept rapid AI as a premise still forecast something legible in ordinary macro language. Technologists and venture capitalists who accept the same premise forecast something much different.
So the real fault line is about transmission. How quickly does frontier capability become cheap, reliable, legally deployable, and organizationally integrated labor substitution at scale?
Economists answer that question with what amounts to a friction-heavy prior. They have electrification, automobiles, PCs, and the internet as reference cases, all of which took decades to fully reshape measured productivity. They see regulation, procurement, liability, complementary capital, organizational inertia, and political economy as thick constraints. Their mental model: capability enters a dense institutional order and gets metabolized slowly. Cowen’s own version of this argument, replete with Baumol’s cost disease, O-ring effects, human bottlenecks, and the stubborn smoothness of GDP growth, is the most elegant statement of the position.
Technologists answer the same question with a capability-first prior. They look directly at the object: benchmark gains, declining error rates, improving agency, tool use, coding, reasoning. Their mental model is threshold-based. Once a system can reliably do economically meaningful cognitive work, the existing organization of labor is contingent, not durable. Frictions are real but temporary. Capability compounding is not.
VCs operate on yet a third model, which is really just asymmetric payoff math. They do not need the median path to be radical. They need the right tail to be fat. A VC can rationally talk as though discontinuity is coming even if the modal outcome is moderate, because venture returns are dominated by convex upside. That is simply a portfolio construction principle wearing a forecaster’s clothes.
Where These Models Fail
Each of these models captures something real. Each also has a characteristic failure mode.
Economists underweight threshold effects. If general cognition becomes very cheap, then base rates drawn from prior general purpose technologies may be the wrong anchor. The paper hints at this: even in its own rapid scenario, the forecasters never leave historical experience. That is either disciplined humility or systematic under-imagination, and you cannot tell which from a survey.
Technologists underweight institutions. They spend their professional lives at the frontier, surrounded by early adopters, fast-moving teams, and capital hunting non-linear returns. That environment is radically unrepresentative of the economy, which is mostly hospitals, school districts, regulated utilities, procurement committees, and middle managers trying to avoid blame. The demo is not the deployment. “Can the model do the task?” is a different question from “Can the organization rewire itself around the model?” Frontier people systematically confuse the two. There is a reason that Palantir relies so heavily on forward deployed engineers.
VCs have the simplest pathology: they overweight possibilities with enormous asymmetric payoff and then mistake that portfolio logic for a description of the world. To be fair to them, asymmetric payoffs is what they’re paid for. But it means that VC rhetoric about AI is decision-relevant for their limited partners, not for macro forecasters.
Reconciliation
Both the technologist-VC side and the economist side are probably right, at different levels of description. Aggregate macro outcomes will likely look more continuous than the technologists and VCs expect. The world is thicker than the frontier crowd admits. But the composition underneath those aggregates may get much stranger much faster than the economists’ tone implies.
You can post decent GDP growth while junior white-collar career ladders collapse. You can avoid Depression-style unemployment while bargaining power erodes and the labor share drops to 45%. You can have rising median income and 80% wealth concentration simultaneously. This is a profoundly altered social order that still fits inside a familiar model.
The paper basically shows this. Its economists do not forecast apocalypse. But they also forecast a world where capital gets dramatically stronger relative to labor and where the top decile pulls away hard. Then, in the policy section, those same economists mostly recommend retraining programs and modernized unemployment insurance. If you genuinely believe the top 10% will own 80% of wealth, retraining is not a response proportional to the problem. It is an elite prior leaking through the survey.
The Deeper Issue
We are trying to settle this argument without the right instruments. Right now the economist-vs-technologist/VC debate is conducted in papers, podcasts, blog posts, and Twitter threads. Nobody has to collateralize their view. Nobody faces a margin call on their AI macro forecast. The claims are cheap because the claims are not priced.
This is where my own obsession comes in. A liquid forward curve on compute–real derivatives, with real collateral, real settlement, and real counterparties–would do more to resolve this debate than another hundred surveys. If you believe AI demand will be enormous and sustained, buy the forward. If you believe diffusion will be slow and supply will catch up, sell it. The curve would encode the market’s collective estimate of the exact variable that matters most: how fast capability becomes deployed capacity.
We do not have that yet. The compute derivatives market is nascent, fragmented, and pre-institutional. But the infrastructure is being built, and when it arrives, it will convert the loudest argument in technology into a price. That price will be disciplined in a way that vibes never are.
Until then, the AI debate will remain what it is: a collision between people who model frontiers and people who model frictions, with no market mechanism to arbitrate.
If you enjoy this newsletter, consider sharing it with a colleague.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can: