Such problems are mostly solvable in principle by anyone (human or AI) who understands the laws of physics/mathematics can interpret the abstract twists in the question, can model and compute the result. We are built to interpret these problems through practice and extreme mental gymnastics which is not plausible for an AI to recreate.

BUT — and here’s the dagger — interpreting is not computing. AI doesn’t truly understand a question like a human does. It recognizes patterns in symbols, but it doesn’t form mental models of physical reality.

tf can AI do under stress? It doesn't even know how to handle it. Neither can it form analogies to unrelated concepts On the other hand humans see the soul of a problem rather than the syntax

Example: A JEE Advanced physics problem might seem to be about optics, but it's secretly a problem about relative motion. A top student gets this from a flash of intuition. An AI? It often chokes unless that specific twist is already in its training data. I can send the question here if you want. It was in this year's paper. You will be shocked to see the connection in the concepts.