Melanie Mitchell: AI Intelligence Is Jagged and Unmappable

The most unsettling thing about today's AI systems isn't that they fail. It's that you can't predict where they'll fail. In the Yale Review, Melanie Mitchell, a scientist at the Santa Fe Institute, has a name for this: "jagged intelligence." The landscape of AI capabilities is profoundly uneven: these systems demonstrate excellent abilities on certain problems while producing surprising failures on other, seemingly similar ones.

This cuts against a basic assumption humans bring to competence. When a person is skilled in one area, you can often predict they'll handle adjacent tasks. AI systems don't work that way. Ilya Sutskever, cofounder of OpenAI, is quoted in the essay saying that models "generalize dramatically worse than people. It's a very fundamental thing."

Mitchell traces part of the jaggedness to something the field has largely treated as irrelevant: the absence of a body, a self, and intrinsic drives. An LLM has no conception of itself as a "self" and no self-generated desires or motivations. Most in the AI field have assumed that embodiment and engagement with the world are beside the point when training machines to think. The success of language-trained models encouraged a belief that intelligence could be sifted away from "the messy world of bodies, emotions, and caring."

The honest caveat is that the essay is more diagnosis than prescription. Mitchell writes that researchers, herself included, are "still struggling to design effective evaluation methods" and to "smooth out the jagged terrain of these systems' skills." What the piece doesn't give you is a map of which domains are most at risk or a path to fixing the jaggedness. The question Mitchell raises, how society collectively decides what AI should be used for, is exactly the kind of question the technology's current momentum makes harder, not easier, to answer deliberately.