The Jagged Frontier: Why Your Instincts About AI Are Wrong

There is a quiet assumption buried in almost every conversation about artificial intelligence, and it is wrong. The assumption is that capability is smooth — that AI is like a rising tide, competent up to some waterline and incompetent above it, and that the line moves up a little each year. People reason as if they can point to a difficulty level and say "AI can handle everything below this and nothing above it."

That mental model is comforting because it is the one we use for humans. A person who can write a clean legal brief can probably also summarize a meeting. Someone who can solve hard differential equations can almost certainly make change at a register. Human ability is correlated across tasks because it flows from a shared bedrock of general understanding. So we import the same expectation to machines.

But machine capability is not smooth. It is jagged.

The shape of the thing

A large language model can draft a passable contract and then fail to count the number of words in a sentence. It can explain the proof of a theorem most people have never heard of and then insist that 9.11 is larger than 9.9. It can write working code in a language you don't know, then confidently invent a citation to a paper that does not exist. These are not occasional glitches around a clean boundary. They are the boundary. The frontier of capability is not a horizon line; it is a coastline — full of inlets and peninsulas, places where the model reaches improbably far and places where it falls improbably short, often inches apart.

The researcher Ethan Mollick popularized the phrase "jagged frontier" for exactly this, and it is the single most useful idea for anyone actually trying to use these tools. Because the jaggedness is not random noise. It has structure, and once you see the structure you can navigate it.

Why it's jagged

The shape comes from how these systems learn. A model's competence at a task is roughly a function of how that task is represented in its training and how well the task decomposes into the kind of pattern-completion the model does well. Tasks that look like "continue this text in a plausible, fluent way" are deep water — the model swims easily. Tasks that secretly require a different kind of operation — precise counting, holding a stable internal state, executing a rigid multi-step procedure without drifting — are shallows, even when they feel trivial to us.

The trap is that easy for a human and easy for the model are uncorrelated, and sometimes inversely correlated. Counting letters is something a six-year-old can do and a frontier model struggles with. Writing a sonnet about thermodynamics in the style of Donne is something almost no human can do and the model produces in two seconds. We keep being surprised in both directions because we keep using the wrong yardstick.

The cost of a smooth-frontier assumption

This matters because the smooth-frontier assumption produces two opposite and equally expensive mistakes.

The first is over-trust by adjacency. You watch the model do something genuinely impressive — diagnose a subtle bug, restructure a messy argument — and you generalize: it's clearly competent, so I can hand it the next thing too. But the next thing sat just across an inlet, and the model fails in a way that is hard to catch precisely because it just earned your confidence. The most dangerous AI output is not the obvious garbage. It is the plausible answer to the question you assumed it could handle because it nailed the previous one.

The second mistake is the mirror image: dismissal by adjacency. You watch the model botch something simple — fumble a date, miscount a list — and conclude it's a toy. So you never discover that the same tool would have saved you nine hours on the genuinely hard task next door.

Learning to read the coastline

The practical skill, then, is not "is AI good or bad." It is cartography. The people getting real leverage from these tools are the ones who have built an internal map of where the frontier juts out and where it caves in — and who treat that map as the actual expertise, more valuable than any individual output.

A few principles for drawing your own map:

Probe deliberately, at the edges. Don't just use the tool where it obviously works. Spend time on tasks where you genuinely cannot predict whether it will succeed.

Decouple your difficulty intuition from the model's. When you're about to delegate a task, ask: does this require precise state-tracking, exact counting, faithful retrieval, or rigid procedure-following? If yes, you are likely near a shallow, regardless of how easy it feels to you.

Verify across the inlets. Because failures cluster right next to successes, never let competence on step one buy trust for step two.

Re-survey often. The coastline moves. A peninsula that didn't exist last year may be solid ground now. The map decays. The map-making skill does not.

The real shift

We spent a century building intuitions about competence by watching other humans, and those intuitions are load-bearing in how we delegate, trust, and verify. AI quietly invalidates them. The frontier does not respect our sense of what is hard.

The most insightful thing you can do is not to ask how smart these systems are. It is to stop assuming the answer is a single number — and to start treating the ragged, surprising, inlet-and-peninsula shape of machine ability as the thing worth studying. The people who win the next decade will not be the ones with the best models. They will be the ones with the best maps.