An organization can only learn from a technological frontier while there is still a gradient to cross. That's what I came to think after seeing, the same week, three pieces from three different fields.
The first, a paper by Andrea Cortis argues—if I've understood the thesis correctly—that data isn't the non-rival good economics has long assumed it to be. A file costs nothing to copy, and that fact has convinced a generation of theorists that data behaves like a public good—shareable without loss. Cortis says that's the wrong level to look at. What actually carries value is the asymmetry between whoever holds the information and whoever doesn't, and he models that asymmetry as a thermodynamic gradient: real, extractable, and dissipated as it is exploited. A piece of inside information may be worth a fortune before the trade and progressively less once the market absorbs it, even though the file behind it never changed—only the gradient it carried did.
The second, a line of thinking from Philippe Silberzahn, makes a similar claim from a different direction. Disruptive technology nearly always looks worse than whatever it's replacing—judged by the standards that older technology was built to meet. Early automobiles were much less reliable than horses. But horses were burying nineteenth-century cities in a public-health crisis of their own, and the automobile, unreliable as it was, made that disappear. Farmers, who had no such crisis, correctly rejected it for another two decades. Silberzahn's rule: never judge a new technology by the old one's standards. Judge it by what it does on dimensions nobody thought to measure yet. That mismatch between old criteria and new ones is itself a gradient—crossing it is what produces the shift, the same way a gradient always does.
A third argument, from Julien Barbier, applies something close to Cortis's logic directly to the frontier labs. As open-weight models close the gap on raw capability and the safety case for public release gets harder to make, he expects the largest labs to keep their very best models for their own internal use rather than sell the tokens—applying them directly to a handful of high-value problems—a drug candidate, a financial model, a diagnosis—instead of renting them out through an API. Whether or not the timeline holds, it's the same logic as Cortis's—this time from the people who actually hold the asset: an asymmetry is worth the most to whoever exploits it first, and worth progressively less to everyone who touches it after. Selling the token sells that gradient to whoever buys it; using it yourself is the only way to keep it.
The three come from entirely different fields, and apply to entirely different problems. What struck me is how closely each one resonates with a tension I recognize from inside large organizations: the friction of absorbing a new technological wave.
The Same Logic, in the Organization
Standards take time to settle and tools take time to mature. The months spent working with something before either happens work less like a cost paid for being too early than like the only window in which anything gets learned firsthand.
A gradient, in this sense, points somewhere without naming it—the slope under your feet, not the destination at the bottom. A team standing close to the frontier doesn't need to know where the technology is going. It only needs to stay close enough to feel which way the ground slopes.
What happens in that window is specific, and it stops happening once the window closes: the organization touches the raw material directly, and some of what it learns from that contact stays inside it—competence, judgment, a way of working that survives long after the tool itself is replaced. By the time the technology arrives packaged—narrated, sanitized, regurgitated into a slide deck by whoever sells it at scale—someone else has already done that learning. What's left to buy is the destination: benchmarked, procured, turned into an enterprise product. What's gone is the slope that taught the first people who touched it how to get there.
The better move is to go looking for the unpolished stone. Not for the thrill of it, but because cutting a rough stone transfers something to the hand doing the cutting. That transfer is the actual point.
The Contact Patch
A tire doesn't touch the road along its whole surface. At any instant, a patch the size of a hand is doing all the work—absorbing the irregularities of the surface, deforming under load, converting an engine's raw horsepower into grip. Remove the friction there and the car doesn't become more efficient. It spins its wheels and goes nowhere.
Over the past ten years, building software inside a large industrial group, I've experienced it first-hand, and come to think it works the same way for an organization meeting a technology frontier. Most of that organization has no reason to touch raw, half-finished tools, and shouldn't. But somewhere inside each of our teams, a few individuals need to be exactly that: contact patches. They meet the raw material before the rest of the organization does, absorb the friction, and generate the one thing the whole company can actually use—enough grip for something much larger than themselves to change direction. Not wide. Well engineered. Built to take a tight corner at speed without hesitating to burn a little rubber.
There is one last resonance with machine learning. A gradient does not merely indicate a direction; propagated backward, it changes the system it passes through. The same may be true of organizations. Keep a small part of one in contact with the frontier long enough, and the friction does not stay at the edge. It propagates inward—changing what the organization builds, whom it hires, and eventually how it thinks.
Not Wide. Well Engineered.
None of this argues for exposing an entire company or organization to every unfinished tool that turns up promising the future. Most of an organization should wait, sensibly, for what eventually gets validated—that's what operating at scale requires. But some part of it has to stay close enough to the frontier to feel the slope before it hardens into a paved road.
