Lately I’ve been diving into the concept of jagged intelligence — and it’s reshaped how I think about where AI is actually headed.
Jagged intelligence describes how AI can be brilliant at complex tasks — passing medical exams, solving advanced math problems, writing production code — yet still stumble on things that seem simple and intuitive to us. The capability profile isn’t a smooth curve. It’s a jagged edge: peaks of superhuman performance next to valleys of basic failure.
That’s a stark contrast to humans, where capability tends to be more uniform. Someone who can reason through a complex business decision usually doesn’t also struggle to count letters in a word.
What fascinates me is that this jaggedness isn’t just an early-stage quirk. It’s persistent — even as models get larger and more capable. The peaks get higher, the valleys move around, but the shape stays jagged.
That raises two questions worth sitting with:
Can we ever build AI that’s as consistently reliable as human intuition? Or is jaggedness an inherent property of how these systems learn — pattern matching at scale rather than building grounded understanding?
What does this mean for how we use AI in real workflows? If the capability surface is unpredictable, the trust model has to account for it. You can’t just hand off a task and assume competence the way you would with a person. You have to verify in the spots where the jaggedness shows up.
For now, the safest bet is to treat AI as brilliant but uneven. Pair it with humans in the gaps. Don’t assume reliability just because it nailed the last hard task.