The Talent Question

TAM-CMN.11 · The Common Mind · The Approximate Mind

A machine learning engineer named Kavitha graduated from IIT Madras three years ago, near the top of her class, with research in small-model optimization that her advisor said was among the best the department had produced. She had two offers before she defended her thesis. One was from a frontier AI lab in San Francisco, salary in dollars, visa sponsored, housing stipend, stock options that would vest into an amount she had never held in her hands. The other was from a research institute in Bangalore, salary in rupees, mission to build Indian-language models for public health, housing her own problem to solve. She took San Francisco.

She was not wrong to take it. The salary differential was not a small multiple. It was a life-changing one, and she has loans and a family that watched her work her way through an education they could not have afforded themselves. The rational thing to do with an offer like that is take it, and she did, and the country that trained her watched her leave the way it has watched thousands of engineers leave before her, with pride and loss in the same breath.

The talent question is not whether people like Kavitha exist. India produces them in numbers that would be the envy of any country on earth. The question is where they go, and whether the places that need them most can make staying rational.

The frontier recruits from a global pool, and it recruits the best, because the best is what frontier research requires and because the compensation the frontier offers is set against a revenue base that a public research institute in Bangalore cannot match. This is not a new pattern. The brain drain from the Global South to the Global North has run for decades across medicine, engineering, science, and academia, and it runs through the same mechanism every time: the center can pay more because the center captures more, and the periphery trains the talent the center employs. The AI version of this drain is faster and more concentrated than previous rounds, because the talent is rarer, the compensation gap is wider, and the window during which foundational decisions are being made is shorter.

What the floor needs is not the same talent the frontier needs, and this distinction matters more than the compensation gap. The frontier needs researchers who can push the boundary of what is possible: larger models, new architectures, capabilities that did not exist last year. The floor needs engineers who can deploy, fine-tune, maintain, and govern models in specific contexts for specific populations. It needs the person who can take an open-weight model, adapt it to Odia or Twi or Darija, train it on local health data, and keep it running on hardware that overheats in April. It needs Priya from the storage closet in Bhubaneswar, and it needs hundreds of Priyas, and the talent pipeline that produces them is different from the pipeline that produces frontier researchers.

This is not to say the floor’s talent is less skilled. It is differently skilled, and the skills it requires are undervalued precisely because they are applied rather than theoretical, local rather than universal, operational rather than published. The person who can keep a district health model running through a monsoon, patch it when the data format changes, retrain it when a new drug enters the formulary, and explain its outputs to a doctor who does not trust machines is doing work that is at least as demanding as frontier research and receives a fraction of the recognition. The floor cannot be built without this work, and the work cannot be sustained without people who choose to do it, and the people who might choose it are being pulled toward the frontier by every incentive the market knows how to produce.

There is also a category of talent the frontier does not produce at all, because it has no use for it: the person who holds both the technical knowledge and the domain judgment needed to know whether a model is right for the wrong reasons. A frontier researcher can build a model that predicts treatment outcomes. A floor practitioner needs to know that the model’s predictions are systematically wrong for patients whose glucose readings follow seasonal labor patterns, and that this wrongness is not a bug but a gap in what the model was shown. This is the epistemic judgment the previous essay described, the ability to see where the system’s confidence exceeds its evidence, and it lives at the intersection of AI competence and deep domain knowledge. No training program currently produces this person on purpose. The floor needs them by the hundreds.

India is the case where the talent exists and the question is allocation. India has enough engineers, enough researchers, enough institutional depth to build both a frontier champion and a public floor, if the talent is distributed between them rather than drained entirely to one. The question is not capability but will: whether the country values the floor enough to create career paths that make building it a rational choice for someone as talented as Kavitha. This does not require matching San Francisco salaries. It requires making the work legible, respected, and funded well enough that choosing it is not a sacrifice but a different kind of ambition.

The harder case is the country that does not have India’s depth. A country with a small engineering base, a few hundred capable ML practitioners rather than tens of thousands, cannot lose even a handful to the frontier without hollowing out its capacity to build anything at home. For these countries the talent question is existential, and the answer is not domestic. It is coalition.

A regional coalition that pools talent the way it pools data can distribute the burden of building and maintaining the floor across multiple countries, each contributing what it has, none required to sustain the full stack alone. A West African coalition does not need every member country to produce frontier-class ML researchers. It needs enough, collectively, to build and maintain the domain models the region shares, and it needs institutional structures that make contributing to the coalition’s floor a career rather than a detour.

In practice this looks like a regional AI institute with permanent positions, competitive enough to hold good people, embedded in the universities and health systems and agricultural ministries of the member countries. It looks like a rotation program where a Ghanaian engineer spends two years working on the coalition’s agricultural model and returns home with skills the national system needs. It looks like shared infrastructure maintained by a shared team, so that no single country bears the full cost of the storage closet and the Priya who keeps it running. The institution is not glamorous. It does not publish in the journals the frontier reads. It builds and maintains the thing that works, and the career it offers is the career of the person who keeps the lights on rather than the person who invents the next bulb.

The coalition does not solve the brain drain. San Francisco will still call. But it provides a destination that is closer to home, professionally meaningful, and embedded in a mission that the frontier, for all its intellectual excitement, does not offer: building something that serves the people you grew up with, in the language your mother speaks.

I wonder whether the talent pipeline’s deepest failure is not that it loses people to the frontier, but that it never teaches them that the floor is work worth doing.

The kind of talent the floor needs most is not the kind that publishes papers. It is the kind that maintains systems: the administrator who notices the model’s predictions have drifted before the doctors do, the linguist who catches the mistranslation that a monolingual engineer would miss, the domain expert who knows that the training data is missing a population the model claims to cover. These are people whose expertise is not in AI at all, but in the thing AI is being applied to, and their judgment is what stands between the floor and the kind of confident failure that CMN-09 described. Retaining them is not an engineering problem. It is a question of whether the institutions that build the floor value domain knowledge and epistemic judgment as much as they value the ability to write code.

Kavitha, in San Francisco, is training a model that will be very good at a great many things for a great many people who look, in the ways that matter to her employer, like the people who already pay. She is brilliant at this work, and the work is real, and she does not regret the choice she made. But there is a version of Kavitha’s career that runs through Bangalore instead, building the Odia health model that serves the clinics where her grandmother waited in line, and that version does not exist yet, not because the work is impossible but because no one has made it a career that a person at the top of her class would choose without apology.