Two recent pieces caught my attention. They arrive from different angles but land on the same conclusion: intelligence is not the binding constraint on real-world progress.

Ernie Tedeschi at Stripe Economics looked at the productivity data. At the task level, AI is crushing it. Customer service agents are 14% more productive. Writers are 40% faster. Consultants finish 12% more tasks in 25% less time. Knowledge workers save two hours a week on email. These are real numbers from peer-reviewed studies, and most of them are based on older LLMs. The current generation should be even better.

So where is the macro productivity boom?

It’s complicated. US labor productivity is running at about 2.5% — solidly above the 1.6% average of the last two decades. But when you look at total factor productivity (the “pure technology” part economists care about), it’s basically flat. The San Francisco Fed puts the probability of a high-TFP-growth period at under 20%.

What’s actually driving the labor productivity numbers? Companies running existing capital harder. Longer factory runs. More GPU cluster utilization. More hotel room occupancy. That’s real output, but it’s not the same as “AI is making everyone more productive.” It’s more like “everyone is scrambling to build AI infrastructure and pushing everything else harder to keep up.”

Here’s the part that should make you uncomfortable: when you control for pre-existing trends, the correlation between AI adoption and sector-level productivity growth drops to essentially zero. The sectors that adopted AI fastest were already the high-productivity sectors. The data can’t distinguish between “AI is helping” and “these sectors were going to be productive anyway.”

Ruxandra Teslo’s piece explains why this makes sense. She’s a medical researcher who keeps running into the same wall: the bottleneck is never the science. It’s the system around it.

Take clinical trials. Seven years. Over a billion dollars per drug. And trials aren’t a formality — they generate the human data that would actually make AI models better. But even when you have the data, you can’t always use it. Companies trying to build better biomarkers have waited a year for the NIH to release imaging datasets. The FDA’s validation of Bone Mineral Density as a surrogate endpoint took twelve years. Twelve. Years. For what was basically a regression analysis on data that already existed.

Or look at Eroom’s Law — the number of new drugs approved per dollar of R&D has been falling for decades, even as our scientific tools got exponentially better. More intelligence, worse returns. That should tell you something about where the actual bottleneck sits.

The patent system makes it worse. Patents protect molecules, not biology. Whoever validates that a novel target is worth drugging bears all the risk and captures none of the reward. Once the first drug posts clinical data, competitors pile on with their own molecules against the same validated target. So AI-driven biotech companies — even the ones valued at billions — end up optimizing molecules against the same handful of proven targets. The system rewards fast-following, not exploration.

This is the inner-loop/outer-loop problem. AI coding tools increase commits and code volume. But the gains shrink at every production stage — from pull requests, to projects, to shipped releases — because review, integration, testing, and deployment are still human bottlenecks. AI speeds the writing. It doesn’t speed the shipping.

The same pattern appears everywhere. We’ve had the technology to build better housing for decades. Housing is more expensive than ever because it’s a political will problem, not a capability problem. Customer service chatbots are still the worst part of most support experiences. GDP is up, but not dramatically. Entry-level jobs are largely intact despite years of predictions about mass displacement.

Tyler Cowen said this at a Progress Conference in 2023: the capabilities would arrive faster than the changes they produce. GDP growth would disappoint relative to hype. There would be no massive job displacement. He was right. And the people who saw it coming were the ones with a better feel for friction — for how slowly and unevenly a technology actually diffuses into the world.

The people who didn’t see it coming? A particular scene in San Francisco that has become something close to a monoculture. Within that circle, it’s genuinely low-status to suggest AGI might not be wholly transformative. The people who were early and right about AI made fortunes, and being right about the big thing hardened into an assumption that they’re right about everything. Questioning the most extreme claims gets you marked as dim, insufficiently “AGI-pilled.”

The result is self-censorship and counterfeit contrarianism. Smart people inside AI labs genuinely believe their future knowledge is privileged. Everyone else nods along. Meanwhile, the real problems — housing, de-skilling, fraying social fabric, people too incurious to read — pile up unattended.

Here’s what I think is the right frame: AI is a powerful tool for accelerating tasks that are already well-defined and bounded. It is not yet a tool for dissolving the institutional, regulatory, and economic friction that determines whether task-level gains become real-world outcomes. Those two things are very different. Confusing them leads to misplaced priorities.

In medicine, regulatory reform gets a fraction of the attention lavished on AI-enabled drug discovery — even though the reform is what would actually move the needle. In tech, we’re fixating on AGI while mundane but critical problems go unfixed. In the economy, we’re watching utilization masquerade as productivity transformation.

The gains are real. The transformation is slower than anyone in the room wants to admit. And the binding constraints are almost never about intelligence.

Sources: