The expert’s few possibilities

§ 1Thirteen graves above the Missouri

On the afternoon of 5 August 1949, fifteen smokejumpers stepped out of a C-47 over Mann Gulch, a steep drainage above the Missouri River in Montana’s Helena National Forest. The fire below them looked routine. On the ground they were joined by a fire guard named James Harrison, and for an hour the sixteen men worked downslope towards the river, tools on their shoulders, expecting an ordinary night of digging line.

Then the wind turned. The fire jumped the gulch, caught the waist-high cheatgrass on their side, and came uphill after them faster than a man can run. The crew’s foreman, Wagner Dodge, shouted at his men to drop their tools. Some did, late. Some did not. One was seen still carrying his shovel; others were found with their heavy packs on. Dodge himself did something no one on the crew had been taught: he struck a match, burned a patch of grass ahead of the flames, and lay down in the hot ashes of his own small fire while the main fire roared around him. Two of the youngest men, Robert Sallee and Walter Rumsey, squeezed through a crevice in the rimrock and survived. Thirteen others died within minutes of one another. Norman Maclean spent his last years reconstructing those minutes in Young Men and Fire.¹

The organisational theorist Karl Weick later read Mann Gulch as an allegory of professional life.² A firefighter’s tools are the physical form of his competence; putting them down feels like ceasing to be a firefighter at the precise moment being one matters most. Weick noted that the pattern repeated in the 1994 South Canyon fire in Colorado, where fourteen firefighters died on Storm King Mountain, several still wearing packs, at least one still carrying a chainsaw. The equipment that certifies you as an expert has weight, and there are moments when the weight is the thing that kills you.

This essay is about those moments as an economic condition. For most of the twentieth century, the labour market paid a steadily rising premium for accumulated, certified knowledge, and it was rational to spend your twenties acquiring the pack and the tools. The evidence of the past few years suggests that the terrain has tilted. The premium has flattened. The certificates are being repriced, slowly and with enormous institutional friction. And a long line of cognitive research says that knowing itself, past a certain point, carries a tax that almost no organisation measures.

§ 2A premium goes flat

The college wage premium, the gap between what degree holders and school leavers earn, is the most studied number in labour economics, and for two decades it told a single story: get the credential. Between 1980 and 2000 the premium more than doubled in the United States. Then it stopped. Researchers at the Federal Reserve Bank of Cleveland estimate it has declined by roughly 10 per cent since 2000, and their model projects further decline over the coming decades, because technological progress since the turn of the century has stopped favouring graduates over non-graduates.¹⁰ Economists at the San Francisco Fed reach a compatible conclusion by another route: the 2023 premium sat slightly below its 2000 level, and firms have become measurably better at substituting non-graduate labour for graduate labour.¹¹

Two forces sit behind the flattening. Supply first: workers holding a bachelor’s degree or higher made up about 31 per cent of the US civilian labour force in 2000 and about 45 per cent by early 2025. The scarce signal became a common one. Cost second: the price of a four-year degree rose about 40 per cent between the 2000-01 and 2022-23 academic years, so a flat benefit now buys a dearer ticket.¹¹

Honesty requires the counter-position, because it is strong. The New York Fed points out that in level terms the premium remains near its all-time high: the median American graduate earns about $80,000 against $47,000 for the median worker with only a school diploma, a gap of roughly 68 per cent, which over a forty-year career comfortably repays a total degree cost of about $180,000.¹² Both readings are true at once. The stock of advantage is large; the flow has stalled. An asset whose price doubled for twenty years and then went sideways for twenty-five is a very different purchase from the one your parents made, even if it still trades high.

The distribution beneath the average is harsher. Zachary Bleemer and Sarah Quincy trace the premium by family income across a century and find it has fallen specifically for students from low-income households, who increasingly attend under-resourced teaching colleges, while rising for the well-off, who migrated into lucrative majors.¹³ The credential still pays, on average; it pays least to the people who borrow most to get it.

Why would a flat premium matter so much? Because of what the degree most plausibly is. Bryan Caplan’s The Case Against Education assembles the evidence that a large share of the premium, half at minimum and by his most aggressive estimate four fifths, rewards signalling rather than learned skill: the diploma certifies intelligence, conscientiousness and conformity that the holder mostly possessed on arrival.²¹ Signals are fragile in a way skills are never. A signal holds its value only while it stays scarce and while no cheaper signal exists. Forty-five per cent of the labour force is not scarce. And cheaper signals, portfolios, shipped work, verifiable output, are multiplying.

§ 3The reset that stalled at the door

Employers noticed. Between 2017 and 2019, before the pandemic gave the trend a shove, 46 per cent of middle-skill occupations and 31 per cent of high-skill occupations saw material reductions in stated degree requirements, according to the Burning Glass Institute’s analysis of more than 51 million job postings.¹⁴ The report’s authors compressed the finding into eight words: “Jobs do not require four-year college degrees. Employers do.” Maryland cut degree requirements from a large share of state jobs in 2022; New Jersey’s governor signed an order dropping them for 90 per cent of state positions; IBM, Bank of America, Walmart, Delta and General Motors made similar announcements. The projection attached to all this was 1.4 million additional jobs opening to non-graduates within five years.

Then researchers went back to check what happened after the press releases. The 2024 follow-up study from Harvard Business School and the Burning Glass Institute, tracking over 11,000 postings against actual hiring records, found that where firms removed the degree requirement, the share of non-graduate hires rose by only about 3.5 percentage points, and that across the whole labour market fewer than 1 in 700 hires could be attributed to the reform. Nearly all of the genuine change was concentrated in 37 per cent of the firms studied.¹⁵

The study is one of the more instructive documents of the decade, because it locates the credential precisely. It does not live in the job posting. It lives in the evaluator’s head. Hiring managers who spent careers reading CVs by school and title quietly restored the filter through “preferred” qualifications; firms that opened external hiring kept internal promotion degree-gated, creating a two-class system employees could see; and nobody trained anyone to recognise competence in its uncertified form. Deleting a line of text turned out to change almost nothing, which tells you how little the line of text was ever doing.

Where the reform was real, it worked. Firms that followed through, Apple, Cigna, General Motors, Target and Walmart among them, hired on average 18 per cent more workers without degrees into the converted roles. Workers who made that jump saw salaries rise by roughly 25 per cent, and their two-year retention ran 10 percentage points ahead of their degreed peers, 58 against 48.¹⁵ The talent was there. The bottleneck was the evaluators’ ability to see it.

§ 4What knowing costs

Why should seeing be so hard? A body of research running from the 1940s to the present gives an uncomfortable answer: the trained eye is trained partly by subtraction.

In 1942 Abraham Luchins published his water-jar experiments.³ Subjects solved a series of measuring puzzles that all yielded to one roundabout formula. Then Luchins slipped in problems with a far simpler solution. Subjects who had learned the formula kept applying it, grinding through three steps where one would do, and many failed outright on a problem the formula could not solve. Fresh subjects, given the same problem cold, solved it directly. Luchins called the phenomenon Einstellung, mechanised thought: the practised solution does the seeing, and the practitioner stops.

For sixty years one could hope this afflicted only the moderately trained. Then Merim Bilalić, Peter McLeod and Fernand Gobet took it to chess masters.⁴ They built positions containing two winning lines: a famous five-move mating pattern and a shorter, better three-move mate. Masters found the familiar five-move idea, reported that they were now searching for something faster, and failed. Eye-tracking showed why: while the players sincerely believed they were exploring, their gaze kept returning to the squares of the solution they had already found. When a control group saw the same position with the familiar pattern removed, they found the three-move mate easily. The presence of a good known answer cost strong masters roughly a full class of playing strength; the effect was later replicated and mapped by Sheridan and Reingold, whose eye-movement data showed experts unable to disengage attention from the familiar move even as they described themselves as looking elsewhere.⁵ The bias runs beneath introspection. You cannot feel yourself not seeing.

The pattern scales. Philip Tetlock’s twenty-year study of expert forecasting, 284 professionals and tens of thousands of predictions, found that specialists barely outperformed informed generalists inside their own domains, and that the most famous experts, the ones with the strongest single organising theory, were the least accurate.⁶ Lars Bo Jeppesen and Karim Lakhani analysed 166 scientific problems that firms had failed to solve internally and then broadcast to outside solvers; the probability of a solution rose with the solver’s technical distance from the problem’s home field.⁷ Chemists cracked biology problems that biologists could not, precisely because they did not know which approaches were supposed to fail.

Max Planck’s dark joke that science advances one funeral at a time turns out to be measurable. Pierre Azoulay and colleagues studied the subfields of 452 eminent life scientists who died prematurely, and found that after each death the flow of work by outsiders into the deceased star’s territory increased markedly, and that this outsider work was disproportionately highly cited.⁸ The star had been an intellectual gatekeeper; the field grew when the gate fell open. Meanwhile Benjamin Jones has documented the other side of the ledger: as fields accumulate knowledge, the age at which innovators make their first contribution keeps rising, and with Bruce Weinberg he showed the mean age of Nobel-winning work climbed by about six years across the twentieth century.⁹ The climb to the frontier lengthens; the burden of knowledge grows; the years available for beginning shrink.

A fairness clause belongs here, and it matters for everything that follows. Expertise is genuinely superb where it is genuinely reliable. Daniel Kahneman and Gary Klein, adversaries who wrote a joint paper to find out where they disagreed, concluded that intuitive expertise is trustworthy in high-validity environments: domains with stable regularities and fast, honest feedback, such as firefighting, chess, anaesthesia and flying.¹⁹ Nobody wants a beginner’s mind holding the scalpel. The beginner’s advantage is conditional, and the condition is volatility: it appears exactly where the regularities are breaking, where the feedback of the last decade no longer describes the next one. The economic question of the 2020s is how much of the labour market still qualifies as a high-validity environment. The honest answer is: a shrinking share.

§ 5The reskilling arithmetic

How fast is it shrinking? The World Economic Forum’s Future of Jobs survey, covering more than 1,000 employers responsible for 14 million workers, expects 22 per cent of today’s jobs to be disrupted by 2030: 170 million roles created, 92 million displaced. Employers estimate that 39 per cent of workers’ core skills will change or become obsolete over the same five years, and that figure counts as good news, down from 44 per cent in the 2023 edition and 57 per cent at the pandemic peak.¹⁶ Two-fifths of what you know, repriced every five years, is the new baseline of a stable era.

The same survey renders the human arithmetic as a group of one hundred. Fifty-nine of every hundred workers will need retraining by 2030. Twenty-nine can be upskilled inside their current role, nineteen retrained and redeployed elsewhere in their organisation, and eleven will need training that, in their employers’ own estimation, they are unlikely to receive: over 120 million people at medium-term risk of redundancy, priced in before it happens.¹⁶

Set against that demand, the supply side of adult learning is a scandal hiding in plain procurement. Corporations spent $356 billion globally on employee training and education in 2015, a figure Michael Beer and his Harvard colleagues examined before concluding that most of it produced no measurable change in behaviour or performance, a practice they named the great training robbery.¹⁷ The failure mode has been understood since 1885, when Hermann Ebbinghaus plotted the forgetting curve: material learned without application or reinforcement decays within days.¹⁸ The modern compliance module, watched at double speed with one eye on email, is a machine for manufacturing exactly the conditions Ebbinghaus showed cannot produce retention. The reskilling economy is being built, at nine-figure annual cost per large firm, on the one pedagogy known for 140 years not to work.

§ 6Institutions of the beginner

It has been done differently, at civilisational scale, for centuries. England’s Statute of Artificers of 1563 made a seven-year apprenticeship the legal gate into skilled trades, and the model it codified, learning bound to production, paid in board and rising responsibility, judged by output rather than examination, ran European craft economies until the statute’s repeal in 1814. Its descendant survives in the German-speaking world’s dual system, where roughly half of each school-leaving cohort enters one of more than 320 recognised training occupations, splitting the week between classroom and firm. The firm pays because the apprentice produces from the first month; the apprentice learns because the feedback arrives from real work, on the timescale Ebbinghaus requires. Youth unemployment in the countries running the system has sat among the lowest in the OECD for decades.

What these institutions share is the thing modern hiring lacks: a way to price a beginner. The degree prices a stock of certified knowledge. The apprenticeship prices a rate of learning, observed directly, under production conditions. Where skills reprice every five years, the rate is the more informative number, and it is legible to anyone willing to watch someone work for a fortnight. The firms that made skills-based hiring real did versions of exactly this: work samples, trials, structured assessment of how a candidate attacks a problem they have never seen. The 18 per cent, the 25 per cent pay rise, the ten-point retention gap were the return on building an instrument that measures learning velocity.¹⁵

There is also a discipline of the individual here, older than any of the data. Zen calls it shoshin, beginner’s mind, and Suzuki’s formulation is quoted so often precisely because the Einstellung literature keeps proving it in the laboratory. The chess masters believed they were searching; their eyes never left the familiar squares. The practical version of shoshin is procedural, and slightly grim: assume your first competent answer is blocking your view of a better one, and build the habit of removing it, the way Bilalić’s control condition removed the five-move mate. Ask the question with the known solution deleted. Bring in the chemist to look at the biology. Read the CV with the university line covered. None of this is mystical. It is attention engineering against a bias that introspection cannot detect.

§ 7Beginning as a discipline

Return to the hillside. The detail that stays with readers of Maclean is that Wagner Dodge, the most experienced man in the gulch, survived by an act that looked like madness to his own crew: setting fire to the grass he stood on. Two crewmen ran past his escape fire because nothing in their training gave them a category for it. And yet Dodge’s improvisation was soaked in expertise; he understood fire behaviour deeply enough to see, in about sixty seconds, that a burnt patch is the one place a fire cannot go. He dropped the tools and kept the knowledge. The two are separable, and the separation is the whole art.

That is the resolution of the essay’s apparent paradox. The evidence does not say knowledge has stopped paying; the New York Fed’s $32,000 a year says otherwise. It says the certificate of past knowledge is flattening in value while the capacity to acquire knowledge is being repriced upwards, at 39 per cent skill turnover per half-decade, by the fastest markdown schedule in the history of formal work. Signals decay; velocity compounds. Institutions have barely started building instruments that measure velocity, which is why the degree reset stalled at 1 hire in 700 and why $356 billion of training buys so little. The instruments exist in prototype: the apprenticeship, the work trial, the broadcast problem, the covered CV.

For a person, the implication is plainer and harder. The pack you carry, the credentials, the practised solutions, the reputation for knowing, was expensive, and it is heavy, and on stable ground it still marks you out. The skill worth drilling now is the one Dodge had and twelve men behind him did not: feeling the moment the ground tilts, and putting the pack down before it decides for you.