Three great flows have defined the modern world: capital, goods, information.

Each followed the same arc: demand existed, rails got built, and a broker formed on top.

• SWIFT was the rail for capital, and Stripe built the broker on top - now $150T crosses borders annually in seconds
• The shipping container did the same thing for goods, and Amazon followed: a factory in Shenzhen can now sell directly to a doorstep in Ohio
• The internet, Google, and ChatGPT put the sum of human knowledge in the pockets of 5.5 billion people

One flow, though, never went through this arc. The most important one.

People.

A broken payment rail surfaces in hours. A broken talent pipeline surfaces in 2045.

Constantinople fell in 1453. Byzantine scholars scattered across Italy carrying texts that Western Europe had lost for a thousand years. Florence absorbed the critical mass, and within a generation, those scholars seeded the Renaissance. The cause wasn’t Medici money or Italian genius; it was a convergence event.

Five centuries later, Germany won a third of all Nobel Prizes in science. Scientists converged on Göttingen from across the world. Each had a piece the others were missing: Heisenberg had the intuition to abandon classical orbits but didn’t know his arrays were matrices, whilst Born recognized the mathematics, and Jordan formalized both. The output was modern physics itself - quantum mechanics, nuclear theory, and everything that followed. Then Hitler took power. 2,600 scientists left Germany in the first year alone, and they reconvened at Los Alamos. That convergence produced the weapon that ended the war.

These weren’t flukes. Talent follows a power law.

The most consequential in any field are <1% of all practitioners, scattered across geographies by the accident of birth. Dispersed, talent produces local optima; concentrated, it produces global optima.

Patrick Collison left Tipperary at sixteen - a science prize, an MIT offer - and routed into the Bay Area ecosystem. Elon Musk arrived from South Africa via Queen’s and UPenn, and Vinod Khosla came from IIT Delhi through Carnegie Mellon to Stanford GSB. Convergence happened inside the institutions when they arrived, but there was no systematic infrastructure that placed any of them. Each arrived through their own very narrow channel.

These are the rare victors. For every success case who found the opening, there are millions of capable people who never knew it could exist. [1] This is symmetrical blindness: The student can’t see the university; the university can’t see the student. And every convergence event it prevents is permanent.

Capital | Goods

Every flow that scaled followed the same arc. Demand existed, a common standard - the rails - created volume, and then an intelligent broker formed on top. SWIFT gave Capital a common syntax - cross-border payments went from hand-written Telex to cheap, machine-readable messages - and global financial assets went from $250 billion to $70 trillion. Likewise for Goods, the creation of the shipping container gave goods a standard unit - uniform dimensions, transferable between ship, truck, and train - and world exports went from $312 billion to $25 trillion. In both cases, costs dropped, volume accumulated, and intelligent brokers - Bloomberg, Refinitiv, Moody’s; Alibaba, Amazon, Flexport - were built on top.

The last flow is left behind, because the destination has to be built by the person moving through them.

The fourth flow has the rails. Student visas have existed for centuries; now tests have gone digital, the Common App arrived, and international student volume has tripled from 2.1M to 6.9M in the last 2 decades - all through 9/11, 2008, and even COVID with flights grounded and embassies closed. But the Broker never formed, because there are no specifiable preferences.

A dollar doesn’t have preferences. A shipping container doesn’t decide where it wants to go. The first three flows are decided with clear criteria: return profile, delivery window, margin. Most AI agent use-cases also work the same way; you tell a shopping agent what you want, and you tell a travel agent where you’re going. But the fourth flow has a person in it, and a person has human agency.

A sixteen-year-old choosing between universities does not know her own criteria set yet. She is, in a real sense, still deciding who she is going to be, and that is a preference construction problem. No Broker can help if there is no preference yet, and for the last 25 years the Jevons payoff never came.

So without a digital Broker, what happens is the industry goes old-school - a human fills the gap. A counsellor evaluates ten destination countries, each with different visa regimes, tuition structures, post-graduation work rights, admission criterias, and employment outcomes that largely don’t exist as data. The recommendation is whichever school feels like the best fit. Multi-variable optimization, from memory - which is to say, a gut call.

But every gut call evaporates; every subsequent student starts afresh, ad infinitum.

The Broker

When a serious Broker arrives, 2 things will happen:

1. Every recommendation, every outcome will become a decision trace, and previous gut calls will become data points.

A student can discover that someone six years ahead of her, with the same background, became a designer at a frontier AI lab. She sees 10 universities that placed students like her, and 4 of them give her an 80% shot at a scholarship that narrowly fits her parents’ budget. In addition, her visa probability stops being a binary gate at the end of the process, and becomes a continuously assessed variable that shapes the portfolio from the start.

As millions of these decision traces form, a map of human possibility takes shape. Preference construction can now happen systematically - full latitude to explore, before exploiting. Monte Carlo would be happy.

More importantly, the student now knows the full set of options available to her and which universities want her. She is now an asset that institutions compete for. She has found her opening. [2]

2. Then, agents will form on both sides of the market.

The university’s agent carries its real specs - rubrics, quotas, cohort needs - probably compressed eventually to a singular Assessment_Rubric.md file. A sixteen-year-old with unspecifiable preferences won’t have her own, so someone will have to build the agent for her, trained on the data of thousands who came before. The future is A2A: the student’s agent queries fifty institutions simultaneously, whilst the university’s agent evaluates her and comes back with an offer. The application, as a discrete event, disappears, replaced by continuous matching, with agents negotiating on both sides.

For the first time, talent begins to converge intentionally. And the last flow completes its arc.

The Last Flow

Talent is scattered by the accident of birth. Every convergence in this essay was the product of narrow channels - rational people picking the best move they could see from where they stood.

Constantinople. Florence. Göttingen. Palo Alto. Widen the channels, maximize the frequency and intensity of convergence, and you accelerate the rate at which humanity produces breakthroughs.

Capital has its protocols. Goods have their containers. Information has its networks. The last flow will have its broker.

Thanks to Charlie Songhurst for encouraging me to bring this idea into existence. To Thanh Hoang and Hayden Vu for helping me sharpen the idea. To Charlie Maynard, Ophelia Cai, and Gina Gotthilf, for reviewing drafts of this.

Footnotes

[1] Unstructured mass migration - who gets in, on what terms, at what social cost - is a policy problem. This is about something narrower: talented people have always crossed borders to find each other, but the layer that matches them to opportunities and routes them efficiently doesn't exist. That's an engineering problem - and no one has built it, because the political debate around migration swallowed the engineering question whole.

[2] There has been vigorous discussion around whether the university - the destination this channel feeds - is itself being replaced. The Thiel Fellowship pulls twenty fellows to San Francisco each year to build companies instead of attending lectures; Y Combinator concentrates hundreds of founders in San Francisco in the same room; Sora Schools and Outsmart College are building AI-first classrooms; Multiverse is reviving apprenticeships. But the common ground here is that all of them are physical convergence events. The lecture hall may be dying, but the gathering function is not. Talent development requires convergence, and convergence still requires physical proximity. The cross-border channel that feeds these gatherings will still need to exist, and the decision of whether to move through it is still the friction. Make it cheap, clear, and early enough to act on, and the latent demand starts to crystallise.