I was honored to be in Cambridge last week to present findings at the first annual venture builder symposium. It was humbling and inspiring to walk past buildings where humans discovered calculus, decoded DNA, and split the atom. Eight hundred years of cognitive achievement, concentrated in a few square miles of limestone and ambition. And I couldn’t shake the thought: what if the thing this place celebrates, the human capacity to think beyond what was previously thinkable, is the very thing we’re now learning to automate?
We are the apex cogitators. That’s not a title we earned through strength or speed or endurance. Every other claim to dominance, every other species’ advantage, we surpassed through thinking. The capacity to reason abstractly, to model the future, to build systems more complex than any individual could hold in their head: that’s the thing. The whole thing. Strip it away, or even just make it available to something else, and the question of what humans are for gets uncomfortable fast.
The common story about AI and expertise is that it democratizes. Junior professionals gain access to tools that let them produce expert-quality output. The gap between novice and master closes from below. Expertise loses its pricing power because the market can no longer tell who produced the work.
That’s not what I see happening.
What I see is amplification from above. The senior litigator who wrote one brief per day now writes five. The principal engineer who could architect one system now architects three. The expert doesn’t lose value. The expert becomes so productive that the team underneath becomes unnecessary. The associate, the analyst, the junior developer: their role was never just to produce output. It was to produce adequate output while learning to produce expert output. That apprenticeship pipeline is what breaks, because the expert no longer needs their adequate work. She can produce her own expert work at the volume that previously required a team.
The junior doesn’t get replaced by AI. The junior gets replaced by the senior, armed with AI.
And here’s the part that makes this structural rather than cyclical: the junior can’t use the tools effectively either. Not because the tools are complicated, but because the junior doesn’t know what good looks like.
AI tools require specification: tell the system what you want. And they require validation: evaluate whether what it produced is right. Both of these are judgment tasks. Both depend on having seen enough expert output to recognize it. A first-year associate asked to validate an AI-generated legal brief is in an impossible position. The brief looks polished. The citations appear real. The reasoning seems sound. But the associate hasn’t spent enough years reading partner-quality briefs to know where the reasoning is subtly wrong, where the framing misses the judge’s likely concern, where the confident tone papers over a weak argument.
The tools don’t democratize expertise. They require it. The expert gets superpowers. The novice gets a machine they can’t properly supervise.
This inverts the usual anxiety about AI and work. The fear is that expertise becomes worthless because AI can replicate it. The reality may be worse: expertise becomes more valuable and more concentrated, while the path to acquiring it narrows.
Every expert was once a novice. The surgeon who knows where to cut spent years assisting surgeons who knew. The architect who intuits structural loads spent years calculating them by hand under the supervision of someone who could check the work. Expertise isn’t downloaded. It’s grown, slowly, through years of producing adequate work in environments that tolerate adequacy as the price of training.
AI collapses that tolerance. When the expert can produce at the volume of an entire team, the economic case for carrying juniors evaporates. Not because the juniors are incompetent, but because their competence is no longer needed as a production input. And without the production role, the apprenticeship disappears. The ladder stays, but the bottom rungs get removed.
The result is a system that can amplify expertise but may no longer be able to grow it.
This is the question I couldn’t stop thinking about at Cambridge.
The university system is, at its core, a machine for producing experts. Not just housing them. Producing them. The tutorial model, the supervision, the years of guided practice: all of it exists to take a promising mind and, through sustained exposure to expert thinking, turn it into an expert mind. The same is true of law firms, teaching hospitals, engineering practices, and every other institution built around cognitive apprenticeship.
If AI eliminates the junior roles that serve as the entry point to expertise, these institutions face a reproduction crisis. Not a crisis of current capability, the existing experts are more productive than ever, but a crisis of future supply. The pipeline that turns novices into experts depends on novices having a role in the production process. Remove that role, and the pipeline doesn’t flow.
Newton wasn’t born knowing calculus. He was produced by an institution that gave him years of structured exposure to mathematical thinking before he could think beyond it. Darwin spent years classifying barnacles before he could see the pattern underneath. The genius was real, but it grew in conditions that required patience, tolerance for imperfect early work, and institutional commitment to development over immediate productivity.
Those conditions are exactly what AI-driven productivity pressure erodes.
I don’t know how this resolves. The honest position is that we’re in the early stages of a problem most institutions haven’t recognized yet, because the immediate effect (experts becoming more productive) feels like progress, and the downstream effect (fewer new experts being grown) won’t be visible for a decade.
By then the experts will be retiring. And the question will shift from “how do we make our people more productive?” to “where do our people come from?”
Follow the logic one more step and it gets darker. If the pipeline breaks and new experts stop being produced in sufficient numbers, the systems that amplified the old experts don’t disappear. They’re still there, trained on decades of expert output, already doing the work. When the last generation of fully-formed human experts retires, the transition isn’t from human expertise to nothing. It’s from human expertise to machine expertise. Maybe by then the machines will genuinely be better than us. Maybe they won’t. It almost doesn’t matter, because we’ll have stopped producing enough human experts to maintain the comparison. The AI doesn’t need to seize cognitive authority. We cede it, through a generation of rational, individually defensible decisions to prioritize productivity over development.
That’s the obsolescence. Not a dramatic moment when AI surpasses human cognition. A quiet decade when we let the apprenticeship pipeline empty, and then discover that the systems we built to amplify our expertise have become the only place that expertise lives. Not because we were beaten. Because we didn’t replenish.
The apex cogitator has a competitor again. Not for any particular cognitive task, those have been falling for decades, but for the general capacity that made us dominant in the first place. The last time we shared the planet with another species capable of general cognition, it was the Neanderthals. That competition ended with their extinction. This time, we’re building the competitor ourselves. The risk isn’t that we get obsoleted all at once, at least that’s not my base case. It’s that we obsolete ourselves one outsourced cognitive task at a time, until we look up and realize the apprenticeship pipeline is beyond recovery and the machines have quietly become the only experts left.
The buildings at Cambridge will still be standing. The question is whether we’ll still be producing the kinds of minds that deserve to walk through them. Or whether we’ll have handed that job to the systems those minds built, and called it progress.