The Fast Lane Inside the Slow Machine

Tyler Cowen and Dwarkesh Patel spent an hour arguing about how fast AI will change the economy. Cowen's two minutes on healthcare are the most useful part. Read with the rest of the conversation, they point to where AI can actually create an advantage in a slow industry, and where it never will.

Part 1

The claim

Late in a recent Interesting Times conversation, Patel asks Cowen whether he would be surprised if AI companies passed $1 trillion in revenue by 2030. Cowen says yes. He is not bearish on the labs. He is bearish on how fast the rest of the economy can absorb what they build.

His first example is healthcare. He is optimistic about biomedical progress, but notes that "with clinical trials, the FDA, that can take 10 or 20 years" (46:47). He guesses that 30 to 40 percent of the economy will barely move by 2035 (47:07).

He is right about the gates. The path from an Investigational New Drug application to a Biologics License Application is set by regulators, and AI will not shorten it much. Anyone betting on that is betting against the institution. My thesis is different. When every company can rent the same models, the advantage goes to the ones that put AI into the decisions they already own. In healthcare, there are many of those decisions, and most of them sit outside the regulatory gates.

Part 2

Same models, different positions

Start with supply. The best models are available to anyone with an account. A large drug developer, a contract research organization and a startup all rent roughly the same intelligence at roughly the same price. Raw capability is becoming an input, like cloud computing. Inputs rarely decide who wins.

Now demand. Capability climbs fast. What healthcare actually uses climbs in small, late steps.

Most commentary treats that gap as one problem. It is really two. Part of it sits behind regulators and will close on their schedule. Part of it sits inside the company and closes as fast as the company chooses. The strategic question is which part you are working on.

Part 3

Split the machine in two

Look at how a new treatment reaches patients. Some steps are fixed gates: the regulatory filings, the agency reviews, the ethics approvals. These move at the institution's pace. No model changes that.

Other steps are decisions the company makes for itself. Which targets to pursue. How to design the protocol. Which hospitals to open. Where to put recruitment money. When to rescue a study that is falling behind. Which data issues to chase. Nobody outside the company signs off on most of these choices.

The evidence suggests the second group is where time is being lost. A Tufts Center for the Study of Drug Development analysis found that drugs approved from 2014 to 2018 spent about 90 months in clinical trials. The clinical phase grew by more than six months compared with the prior period, while approval time fell (The Pharma Letter). The regulator got faster. The part the industry runs itself got slower.

The FDA has drawn the same line. Its January 2025 draft guidance on AI covers models that produce information supporting regulatory decisions about safety, effectiveness or quality. It explicitly leaves out AI used in drug discovery and for operational efficiencies that do not affect patient safety, product quality or the reliability of study results (FDA). That is a large space. Picking hospitals for a study sits inside it. So does forecasting enrollment, planning supply, and deciding where a study team spends its week.

Part 4

The real constraint is context

If the models are the same, what separates a company that gets value from AI from one that does not? Patel answers this earlier in the conversation, without quite naming it.

His central complaint about today's models is that they start "every single session as if it's their first day on the job" (05:01). They are smart and have no memory of your business. Cowen asks whether companies will "tape everything that happens" and feed it to the AI (41:04). Patel points out that a job is much more than its tasks. An accountant also gives the manager feedback about how the organization should change, "and so jobs are much bigger than people imagine" (34:47).

Put that together and you get the real constraint. A model is only as useful as the context it can reach: the history of which decisions worked, the operating data, the unwritten rules a veteran carries in their head. That context does not come with the subscription. Each company has to build it.

Think of it as three layers. The model sits on top, and everyone has the same one. Under it sits the context layer: clean history, labeled outcomes, and the knowledge that explains why past choices worked. Under that sits the workflow, the actual moment a person makes a decision. AI creates value only when all three connect. A company with a strong context layer and a clear spot in the workflow will get more from an average model than a competitor gets from the best one.

This is a pattern strategists know well: switching costs built from accumulated context. The model can be swapped every quarter. The context layer compounds, and it stays with whoever built it.

Part 5

Where to look first

Not every internal decision is a good place to start. The best candidates share three traits.

Choosing where to run a study fits all three. So does early discovery work, where models can narrow the search space before any human runs an experiment. So does the daily operating rhythm of a trial: spotting sites that are falling behind, moving recruitment effort, and catching data problems before they pile up. None of these needs an agency to approve the method. Each one takes time out of the part of the timeline that has been growing.

The weaker candidates are the ones everyone talks about first: using AI to write the filing faster, or hoping a regulator will accept a model in place of a trial. Those efforts run straight into the gate.

Part 6

What to be skeptical of

Cowen's numbers are guesses, and he presents them that way. "Thirty to forty percent of the economy" is a framing. Nobody can check it. He also works inside a university, which he describes as moving backwards on AI, and that may color his view of other institutions.

The context argument has its own risk. Building a context layer is slow, unglamorous work. Most companies' history is scattered across systems, inconsistent and partly undocumented. A company can spend two years cleaning data and still not change a single decision if nobody owns the workflow at the end. Context without a decision to feed is just storage.

There is also a ceiling. Speeding up internal decisions does not remove the gates. A program that saves a year in operations still waits its turn at the agency. The advantage is real, and it is bounded.

Part 7

What it means

In a world where speed is everything and slowness is the norm, AI is a strategic advantage only when it is pointed at the right place. That place is rarely the regulatory path. It is the long list of internal decisions that nobody outside the company checks, and that have been quietly adding time for a decade.

The companies that win will not be the ones with the best model. Everyone will have that. They will be the ones that built the context layer early, found the decisions it can improve, and changed the workflow so the improvement shows up on the calendar.