The same AI that can pass a bar exam will confidently insist that 9.11 is bigger than 9.9. This isn't a bug being patched — it's a fundamental shape to what AI can do, and researchers call it the jagged frontier.
We expect intelligence to be smooth: if something is smart enough for a hard task, surely it can handle the easy ones. AI breaks that assumption completely, and understanding the break is the single most useful thing for actually working with these tools.
Intelligence with holes in it
With a human, capability is a reasonable predictor — someone who can write a legal brief can also count. With AI, capability is spiky. It shows superhuman fluency on some genuinely hard tasks and then fails at something a child finds trivial, with no warning and total confidence in both cases.
Researchers named this the jagged frontier: the boundary of AI ability isn't a smooth line but a jagged edge, with peaks of brilliance right next to valleys of surprising incompetence. Two tasks that look equally hard to you can sit on opposite sides of that edge.
Why the valleys exist
The failures aren't random once you know where to look. A model that can't compare 9.11 and 9.9 isn't reasoning about quantity — it's pattern-matching over text, and "9.11 > 9.9" looks plausible if you've seen a lot of version numbers and dates. It never did the arithmetic; it predicted likely-looking text. Its weakness at counting letters or digits traces back to the fact that it sees chunked tokens, not individual characters. The valleys line up with tasks that need precise symbolic manipulation rather than fluent pattern completion — a distinction I dig into more across my writing.
How to actually work with a jagged tool
Once you accept the shape, your whole approach changes:
- Never assume competence transfers. "It nailed that hard thing, so I can trust it on this easy thing" is exactly the wrong inference. Verify per task, not per impression.
- Confidence is not a signal. The model sounds equally sure when it's right and when it's catastrophically wrong. Its tone tells you nothing about which side of the frontier you're on.
- Use it on the peaks, guard the valleys. Lean on it for fluent, generative, pattern-rich work; put checks around anything needing exact calculation, counting, or rigorous logic.
The people who get the most from AI aren't the ones who trust it most or least — they're the ones who've mapped its jagged edge and know which side of it a given task falls on.
The real lesson
The jagged frontier is a reminder that AI isn't a smaller version of human intelligence — it's a different kind, brilliant and blinkered in unfamiliar places. Treat it like a person and its failures blindside you. Treat it like what it is — a spiky, powerful, uneven tool — and you can rely on it exactly where it's reliable. More of how I reason about AI's real capabilities at www.divyakush.com.
Related reading
- Tokenization: why an LLM can't count the R's in strawberry — the mechanism behind many of those valleys.
Top comments (0)