Your IQ Is Not a Moat: AI Is Coming for the Smartest Jobs First

On July 27, a neurosurgery resident in Beijing posted a proof of Crouzeix’s conjecture, a problem in matrix analysis that had resisted specialists for more than twenty years.

Shanmu Jin had studied geology before medicine. His formal mathematical education, he told mathematicians Alex Townsend and Anne Greenbaum, consisted of the ordinary courses taken by science students. Everything beyond that was self-taught. The decisive theorem in his paper emerged during an approximately sixteen-hour autonomous run of GPT-5.6 Sol. Townsend, Greenbaum and Michel Crouzeix checked the proof and said they believed it correct. Eight days later, two professional mathematicians posted a separate five-page proof and disclosed that they, too, had used ChatGPT to explore proof strategies.

A generation ago, the story would have sounded absurd. Now it punctures one of the modern economy’s most comfortable assumptions: difficult abstract reasoning is a rare human asset whose scarcity can be counted on.

We built examinations around that belief, universities around it and some of the world’s best-paid professions around it. If only a few people could manipulate symbols at a very high level, the ability looked like a durable moat.

AI is breaking the link between difficulty and scarcity.

A task can be brutally hard for a human brain and still be unusually hospitable to a machine. Mathematics has definitions, constraints and proofs. Code has syntax, tests and execution. Both give a system a way to try, fail, inspect the failure and try again without leaving the computer.

We ranked work by how difficult it felt to us. Machines care about something else: how completely the problem can be represented.

Machines prefer clean rooms

A mathematical proof may demand years of training, but the world in which the proof lives is unusually clean. The rules do not change halfway through the argument. A theorem does not become frightened, lie about its symptoms or forget what happened yesterday. A compiler has no childhood. An equation has no family politics.

Compare that with a nurse trying to settle a dementia patient at three in the morning. The patient may be confused, in pain, frightened or unable to explain what is wrong. Medication, tone of voice, physical safety, memory, family context and hospital procedure all matter at once. There may be no single correct action, and the consequences of a mistake happen to a person rather than to a benchmark.

Our status system calls the first task intellectually elite and the second one care work.

Stanford’s 2026 AI Index captures the asymmetry. Agents reached 66.3 per cent accuracy on OSWorld, a benchmark of computer tasks, within six percentage points of human performance. In software simulation, robotic manipulation reached 89.4 per cent. Put robots into actual homes and the success rate fell to 12 per cent.

The benchmarks measure different capabilities, but the contrast is useful. Computers are becoming formidable inside environments that computers can see clearly. The kitchen remains stubborn.

The OECD finds a similar pattern in employment. IT professionals, business professionals, managers, chief executives, scientists and engineers are among the occupations most exposed to recent AI advances. Cleaners, agricultural workers and refuse workers sit near the other end. Exposure is not the same as replacement; senior jobs contain social judgement, responsibility and non-routine work. But more education no longer guarantees more insulation from automation.

The useful distinction is no longer high skill versus low skill. Ask how much of the job can be enclosed inside a legible loop of input, output and feedback.

When answers stop being scarce

Mathematics also shows what happens after a profession’s prized output becomes easier to produce.

Terence Tao now describes a transition from “proof scarcity” to “proof abundance”. In an August essay based on his International Congress of Mathematicians lecture, he sketches a chain of bottlenecks: AI-generated proofs can accumulate faster than experts can verify them; verified proofs can arrive faster than anyone can explain them; even correct, readable results can outrun peer review and the slower work of absorbing them into the field.

That final step is easy to underestimate. A proof becomes part of mathematics only when other mathematicians understand what made it work, connect it to the surrounding theory and teach it. Correctness is necessary. It is not the whole product.

Once candidate answers become plentiful, value shifts toward choosing questions, testing claims, explaining results and deciding what deserves attention. Mathematical taste used to travel in the same package as proof-producing ability. AI can prise those activities apart.

It may also damage the way experts are made.

Every serious profession has an apprenticeship hidden inside its drudgery. Junior lawyers draft. Analysts build models. Programmers chase bugs. Young doctors work through routine cases. Graduate students attack problems that are hard enough to teach them technique but limited enough to be survivable. Much of this work is inefficient by design. The junior person is slower because becoming less slow is the point.

Automate the junior task and the immediate result looks wonderful. The law firm still has senior lawyers. The software company still has architects. The fund still has portfolio managers. A decade later, the organisation may discover that it also automated the training ground that used to produce those people.

We could end up with abundant execution and a shortage of judgement because fewer people were forced through the work that once created judgement.

Tao makes the educational version directly: training a mathematician is not achieved by producing correct homework. Friction matters. Struggle shows the learner where the hard part is. A perfectly polished solution can teach less than an imperfect human one because it removes the path by which understanding was acquired.

This is a problem of succession, not merely plagiarism.

Software shows where the money goes next

A 2026 study by researchers at MIT and Wharton tracked more than 100,000 GitHub developers as they adopted successive generations of AI coding tools. Autocomplete increased coding activity by 40 per cent. Synchronous agents pushed the cumulative increase to 140 per cent. Autonomous agents took it to 180 per cent.

Finished output rose far less: about 50 per cent at the project level and 30 per cent for software releases.

Customers do not buy lines of code. They buy software that works, fits into existing systems, survives review, reaches users and solves a problem somebody cares about. As code production accelerated, review, integration and distribution became more conspicuous constraints. The same research found a surge in new applications across major software marketplaces without a corresponding increase in overall usage.

That should worry any company whose moat is described simply as superior engineering talent. If the advantage was mainly the ability to produce competent code, its value is falling. If it came from knowing a customer, owning distribution, holding proprietary data, understanding regulation or integrating software into ugly old systems, the position may strengthen.

Cheap intelligence does not erase economic value. It exposes where the value was hiding.

Finance learned this lesson before AI

Investment management has always contained a cruel version of the same idea. A profession can attract brilliant people, build intimidating machinery and still struggle to turn intelligence into persistent excess returns.

In 2025, 79 per cent of active large-cap US equity funds underperformed the S&P 500, according to S&P Dow Jones Indices. Its latest institutional scorecard found that, after fees, at least 80 per cent of equity funds across mutual funds, institutional accounts and separately managed or wrap accounts underperformed their benchmarks over the ten years ending in 2025.

Professional investors know an enormous amount. Markets make that knowledge hard to monetise because thousands of intelligent people have similar information, similar models and similar incentives. Intelligence becomes an input rather than an edge.

AI will push that further. Filing analysis, valuation models, transcript comparison, screening and scenario generation are becoming cheap enough to rent. An analyst who once distinguished himself by processing more information than his competitors is entering a world where everybody can process nearly everything.

The scarce advantages are less glamorous: a long time horizon, liquidity when others need to sell, clients who do not fire the manager after a bad quarter, and the temperament to hold a position after the spreadsheet has stopped being emotionally comforting.

Morningstar’s 2026 “Mind the Gap” study puts a number on the behavioural penalty. Over the ten years ending in 2025, the average dollar invested in US mutual funds and ETFs earned 8.7 per cent a year, compared with the funds’ 9.9 per cent aggregate annual return. The 1.2 percentage point shortfall came from when investors bought and sold.

You can own the right fund and still sabotage the result.

This is where a small investor can possess an advantage over an institution. The individual may have no investment committee, no quarterly redemption pressure and no client demanding an explanation for a bad month. A modest portfolio and a strong stomach can create freedoms that a billion-dollar mandate does not have.

AI can make analysis abundant. It cannot give an institution permission to wait.

The status system is about to look strange

We paid a premium for people who could perform difficult symbolic work because those people were scarce. Then we mistook the scarcity of the worker for an eternal property of the work. The computer is separating the two.

Intelligence still matters. The premium is moving toward forms of ability that remain entangled with the world: responsibility, trust, physical presence, tacit knowledge, access, taste, persuasion, institutional memory and the ability to act when the objective itself is unclear. Some belong to chief executives. Others belong to nurses, electricians, carers, salespeople and operators whose work our credential system never placed near the summit of cognition.

Robots will improve. Models will get better at social reasoning. Plenty of messy work will be automated too. That does not rescue the old hierarchy. It shows why the hierarchy was the wrong model in the first place.

“High IQ” is not an economic category. Neither is “knowledge worker”. The useful question is whether a capability remains scarce after machines can perform it cheaply, reliably and at scale.

Universities should ask what an examination proves when a machine can generate the answer. Professional firms should ask what justifies an hourly rate when the legible part of the work costs pennies. Investors should be suspicious of businesses whose moat amounts to “we hire very smart people to process information”. What happens when very smart information processing can be rented on demand?

For most of the modern economy, intelligence was expensive because it had to be grown inside a person. Years of education, selection and apprenticeship stood behind every unit of elite cognitive labour. We priced the output accordingly.

That pricing regime is ending.

The smartest people in the world were never doing the easy part. They were doing the part for which society had no cheaper supplier.

Now it does.