The Jagged Frontier Is About The Problem, Not The Model

When people ask me about what AI can and can't do, the questions are often about the models: How smart is it? How far is it from AGI? However, I find it's often more helpful to think about the problem we're asking the AI to solve first.

Some problems are convergent—they have essentially one right answer, and it's the same answer for everyone. Math problems, bug fixes, and coding challenges are convergent. There's "the" answer.

Other problems are divergent—they have many valid answers that depend on context, audience, and purpose. Business strategies, poems, and design work are divergent. There are countless valid possibilities, and which one is "right" depends on your specific situation.

Regardless of which type of problem you give it, AI always produces some output. Ask for a business strategy and you'll get one. Ask for a poem and you'll get one. The question isn't whether AI can generate output. It's whether that output aligns with your actual intent.

For convergent problems, your intent is mostly captured in the question itself. "Fix this bug" or "solve this equation" contains what you need. The output matches your intent. More training data - with everyone wanting the same answer - works to make the model clearer and clearer on what that right answer is.

For divergent problems, your intent requires context the AI doesn't have. "Develop our Q3 strategy" will produce a strategy, but likely generic consultant-speak drawn from training data patterns. Without knowing your company's specific capabilities, constraints, market position, and goals, AI can't distinguish which of a thousand valid strategies is right for you, no matter how much training data or compute it has. It has to use the unique context in each request in different ways by learning much subtler meta-patterns from much weaker signals of success and failure.

As AI models get smarter, the bottleneck shifts from "can it do this task?" to "does it understand what I specifically want?" For convergent problems, the AI knows what you want immediately, and you can just get out of its way. For divergent problems - and most real problems are divergent ones - your value is increasingly in providing that context—specifying the audience, constraints, and criteria that transform a generic output into your specific solution.