The Spearhead Problem

There’s a quiet tension at the heart of our current AI moment that few seem willing to name directly: the people increasingly relying on large language models to perform intellectual work are often precisely those who lack the domain expertise to recognize where these tools fail.

This isn’t a criticism—it’s a structural inevitability. If you possessed deep expertise, you’d have less need for the shortcut. And so the tool gets adopted fastest by those least equipped to audit its output.

You encounter the boundaries of LLMs in specific places. In coding, they appear when you face problems of genuine intrinsic complexity—the kind that demand novel abstractions, not recombinations of patterns the model has seen before. In research, they emerge at the frontier, where by definition the answers don’t yet exist in any training corpus. These are the moments that require expertise, and they’re precisely the moments the tool cannot help you.

What concerns me is the generational trajectory this implies. Why endure the slow, often painful process of acquiring deep competence when a tool offers you a shortcut to adequate solutions? The incentive gradient has shifted. And incentives, given enough time, reshape populations.

In the short term, this is arguably good. LLMs raise the floor. They democratize access to median expertise—the competent doctor’s intuition, the experienced programmer’s pattern recognition. Society’s mean capability rises, and real problems get solved that wouldn’t have been solved otherwise.

But expertise was never about the mean. The mean moves because a small number of people push far beyond it—individuals who spent years developing the refined judgment to see what others couldn’t, to ask questions that hadn’t been formulated. These are the spearheads that, across generations, drag the entire distribution forward.

If we systematically disincentivize the development of that depth—if we make the arduous path to mastery feel economically and socially irrational—we may find ourselves a generation from now with a higher floor but a dangerously lowered ceiling.

The tools will plateau. The frontier will still require humans who can see past the current boundary. And we may discover, too late, that we’ve forgotten how to produce them.