tl;dr:

  • Expertise is learned structure.
  • Intelligence is what survives prediction.
  • The world trains its own experts.
  • Every domain is a language.
  • Understanding is the ability to expect.
  • A collection of specialists still needs a way to become a generalist.
  • Recursive improvement requires recursive verification.

Timestral forecasts time series. TabPFN works with tables. Liquid’s D1 specializes in decisions. Their training differs, but together they raise a question: how much intelligence can emerge from learning to predict?

And what exactly are we protecting when we insist that a model predicts but doesn’t understand?

Take a mechanic. He anticipates failures, diagnoses unfamiliar noises, and knows what replacing a component will change. If he explains engines beautifully but cannot repair them, I question his understanding. If he reliably repairs unfamiliar engines without articulating every rule, I call him an expert.

Understanding, then, might be a model of what follows from what. Explanation exposes those relationships. Prediction tests them.

“It only predicts” describes an objective, not the capabilities acquired to satisfy it. Predicting language benefits from learning grammar, facts, and relationships. Forecasting benefits from learning trends, cycles, and uncertainty. A simple objective can reward elaborate internal structure.

But prediction can also reward shortcuts. A hospital model might associate a particular scanner with illness without learning much about disease. Change the equipment, and its apparent expertise disappears.

Understanding should survive more than familiar conditions. Can the model handle an unfamiliar case, anticipate an intervention, or recognize that the question itself is wrong? Intelligence is what survives prediction: expectations meet reality, failed assumptions are revised, and useful structure remains.

Is human expertise fundamentally different?

Experience supplies examples. Education organizes them. Practice exposes errors. Experiments distinguish explanations that observation alone cannot separate. Evolved biases and culture shape how we learn, and we actively choose what to investigate. “Exposure to data” understates that process, but placing understanding categorically beyond learning seems equally unsatisfying.

Perhaps domain expertise is what a general learning process acquires by encountering a domain’s structure.

In that sense, every domain is a language. Tables encode relationships, time series encode change, and decisions connect conditions, choices, and consequences. Prediction makes this learning scalable because observations can supply both questions and answers. More varied experience can reward richer models, provided the objective tests the capabilities we actually want.

The practical consequence is small, reusable specialists assembled into larger systems. Does that lead to general intelligence?

Possibly. But a general system must also frame unfamiliar problems, select or build tools, reconcile assumptions, and notice when its specialists are answering the wrong question. Those abilities could themselves be learned. Assembling specialists doesn’t establish that they have been.

Recursive self-improvement adds another requirement: the system must verify that its changes improve capabilities beyond the tests it optimized against. Recursive improvement requires recursive verification.

Small specialists don’t prove AGI is inevitable. They make “domain expertise” harder to defend as something separate from general learning.

If a system can learn what follows from what, transfer that knowledge, and revise it when reality disagrees, what useful definition of understanding would exclude it?