Intelligence Is Leaving Smart People Behind
Over the past two years, LLMs have been commoditizing depth, and with it the definition of intelligence: the thing that decides who gets the hard problems and whose objection kills a decision. And the smartest people under the old definition will be the last to notice.
I. Understanding becomes quality control
Until about three years ago, understanding was the bottleneck of all knowledge work. If you wanted to do anything with a codebase, a legal domain, a body of literature, you personally had to load it into your brain. The marginal return on depth was enormous and stayed enormous nearly all the way up, because there was no substitute: the person who understood the system at ninety-five percent was categorically more useful than the person at sixty. So depth didn't just pay. It became the operational definition of intelligence. Interviews measured it, promotion ladders rewarded it, and "lacks depth" became the polite professional way of saying "not that smart." A generation of intelligent people built their identities inside that definition.
Then LLMs changed the shape of the curve. The model now holds most of the understanding; your job shifts to holding context (what are we trying to do, what does "done" look like, what would be wrong) and operating from first principles at a high level. The marginal value of the sixtieth-through-ninety-fifth percentile of understanding collapses, because the model supplies it on demand. What does not collapse is the first chunk, since you need enough understanding to specify the problem and verify the output, and the last sliver, the frontier-level insight the model doesn't have.
The failure mode of "barely understand," taken literally, is Gell-Mann amnesia at industrial scale: you cannot verify what you cannot evaluate, and models are fluent enough to be wrong invisibly. So the post-LLM optimum isn't "understand nothing." It's "understand enough to specify and check, then stop." Comprehension moves from being the product to being the quality-control department.
II. The new frontier
If depth stops paying, what does? When a production input gets cheap, the rational move is to consume much more of it, and the binding constraint moves elsewhere. Cognition just got cheap. So the new frontier is not how deeply one person can understand one thing. It is how many threads of shallow-but-verified work one person can keep coherent at once. Parallelism, speed, output.
The subtle point: the new optimum still sits above zero depth, and the frontier moved out in both dimensions. You can now go deeper than any pre-LLM specialist if you want to, because the model amortizes the grind. But the slope of returns favors sliding rightward. The person who insists on the old operating point isn't wrong, exactly. They're producing one artisanal unit while their competitor ships nine.
III. Cognition's complements
Now the economic punchline: intelligence is being commoditized, so the scarce input becomes something else. This is Econ 101. When one factor of production becomes abundant and cheap, its price falls and the value migrates to its complements, the things you need alongside it that didn't get cheaper. Fifteen years ago, "very smart person who can absorb a domain quickly" was scarce and commanded a premium. If a model does a large fraction of that for twenty dollars a month, the premium doesn't vanish; it flows to whatever cognition-on-tap still requires. Which is, roughly: knowing what to want, deciding without permission, tolerating the discomfort of shipping something checkable, and iterating fast when it's wrong. My word for this bundle is agency.
There's a historical rhyme. When calculation was expensive, "computer" was a prestigious human job title. Spreadsheets didn't make numerate people worthless. They made the complement, judgment about which numbers to run, the whole game.
IV. The identity trap
Picture the best engineer you know. Fifteen years of accumulated depth, the person who gets paged when the database does something impossible, the one who reads compiler source code for fun. Now tell him his depth is being commoditized. What do you think happens? What are the chances he says "huh, interesting, I guess I'll restructure my entire professional identity by Thursday"?
Of course not. He generates counterarguments. Excellent ones, actually, because generating excellent arguments is the thing he is best at. And the smarter he is, the longer he can keep it up.
This isn't speculation about smart people being secretly petty. It's one of the better-replicated findings in psychology. In Dan Kahan's motivated numeracy experiments, subjects got a tricky statistics problem about whether a skin cream works. The mathier the subject, the better they did. Then Kahan swapped the labels so the identical numbers were about gun control, and suddenly the mathiest subjects were the most polarized of anyone, because they used their skill to torture the numbers until they confessed to whatever their politics needed. Stanovich found the same shape with myside bias: the tendency to grade evidence more kindly when it favors your side is roughly uncorrelated with IQ. Smart people don't have less of it. They're just better at it.
Intelligence is an engine, not a steering wheel. It takes you wherever you were already pointed, just faster, and with better footnotes.
Now point the engine at a career. "Depth is being commoditized" sounds like a claim about labor markets. To our engineer it's a claim about him, and his brain files it with the insults. He becomes one of Kahan's subjects defending the tribe's position, except the tribe is "people whose depth made them special" and the position is "depth still matters most."
And the whole workplace is built to agree with him. For as long as anyone can remember, shallow was how you caught the frauds: someone hand-waves past the details in a design review, everyone exchanges a look, and the look means not that smart. That made sense when shallow meant unverified. It stops making sense when the details live one prompt away and the hand-waver, if she has the right reflexes, ships more verified work across five domains than the expert covers in his one. But every workflow around her still runs on the old signal. Trust still routes to the person who already knows things, and the deepest voice in the room still gets the last word. The definition of intelligence moved; the detectors didn't. Careers are currently being made, and quietly ended, in the gap.
Because the deepest change here isn't how much depth survives. It's that depth stopped being something you own. A decade to build, a moat once built. Now you rent it for the weekend. A moat you can rent is not a moat.
V. The curve won't sit still
What happens over the next few model generations, as the systems get better and cheaper? Those are two different forces deforming the curve in two different ways. Better models compress the left hump: the verification depth you need shrinks each generation, because the model catches more of its own errors. Cheaper models raise the parallelism ceiling: when a thread of competent cognition costs cents, the binding constraint on how many threads you run stops being money and becomes your working memory, meaning how many contexts you can hold, switch between, and keep coherent. Which is exactly the trait this essay is calling the new intelligence. For now.
The chart says two things. First, owned depth isn't just worth less at every point on the curve. It's worth less in total. The value isn't moving somewhere else along the axis; it's leaving.
Second, and sneakier: the point where you can stop double-checking the model keeps moving, and nothing tells you when it moves. Last year you had to check its SQL. This year you don't, but you probably still do, out of habit. Meanwhile someone else stopped checking a year too early and shipped the bug. Both of you are miscalibrated, in opposite directions, and neither of you got any signal about it. So the real skill of this era is neither depth nor breadth. It is re-asking, every few months, "what can I trust the model with now?" Almost nobody asks this explicitly, because recalibrating trust feels like nothing, while building depth felt like virtue.
VI. The comb-shaped engineer
So what does the ideal engineer look like on this curve? The old archetype was T-shaped: one deep spike, thin horizontal bar. The new one is closer to a comb: many medium tines across domains, no permanent spike, but the ability to grow a temporary spike anywhere on demand and then retract it. Plus a cluster of meta-skills the old world barely priced.
Stare at the axis that stays high on both profiles: verification taste. Put simply, it is the residue depth leaves behind. Not remembering how the B-tree is implemented, but the trained flinch when an answer is too smooth, the instinct for where a system will lie to you. The failure the depth-holdouts correctly fear isn't shallowness per se; it's shallowness without the flinch. The people who get called out for glossing over things aren't failing because they lack depth. Rather, it is because they lack the taste that depth used to be the only way to buy. The new game is acquiring the taste without paying the old tuition, mostly by shipping a lot, being wrong fast, and doing the autopsy every single time.
VII. Who to work with
How do you optimize for this, and who do you work with? The intuitive answer is: people intelligent enough to recognize the new intelligence. Open-minded, high-agency, ask the right questions. Correct, but unfalsifiable as stated, because everyone believes they are all three. The observable versions:
Measure their ego latency. Show someone they're wrong about something and start a stopwatch. The interval between "presented with disconfirming evidence" and "genuinely updated, no residual litigation" is the single best proxy for everything in this essay. Depth-era stars often have terrible ego latency precisely because being right was their identity. But some deep people have superb latency. They get the new game and carry the flinch.
Watch them drive a model, not write code. The old test of a mind was the whiteboard. The new test is: hand them a model and a gnarly, underspecified problem, and watch. Do they interrogate outputs or accept them wholesale? Do they reframe the problem when the first framing stalls? Do they fork three approaches in parallel or grind one serially? Do they know when to stop trusting the machine and drop down a level?
Test question quality under ignorance. Drop them into a domain they've never touched and see what they ask. The new-intelligence people ask questions that compress the domain, like "what's the thing everyone here knows that outsiders get wrong?"
VIII. The meta-trait
One last turn of the screw. The breadth people have exactly the same identity stake in breadth that the depth people had in depth, and the models are improving fastest at precisely the things this essay just praised: holding huge context, switching domains, running threads in parallel. The trait being crowned here may simply be next on the conveyor belt. So the durable version of the thesis sits one level up:
The meta-trait is low ego latency about which traits matter.
Look at who actually made it from depth-world to breadth-world. They weren't the broadest people in depth-world. They were the depth people who updated fastest when the ground moved, the ones whose response to disconfirming evidence about their own moat was measured in weeks, not model generations. The same filter will run again on breadth-world, and the ones who make it through won't be the best jugglers. They'll be the jugglers least attached to juggling.
Intelligence turned out to be a definition rather than a possession, definitions move, and the people the old definition flattered most are the ones with the most to lose by letting it. So yes: embrace the new cohort of intelligence, or get left behind. But hold the cohort loosely, because the trait you're betting on is also on the conveyor belt. The thing that compounds isn't any position on the curve. It's the speed at which you're willing to abandon your position on the curve. Which is, admittedly, a deeply inconvenient trait to build an identity around, since building an identity around it defeats it. That might be exactly why it works.