OK this one is really outré. This hypothesis implies that a lot of what we're seeing in LLMs is still (impressive) interpolation and recombination. And yet LLMs are helpful for e.g. real damn new mathematics (though so far limited to what one mathematician called "slight variant of standard facts").

But, does this matter? Recombination of existing knowledge is a big deal! Call the set of true recombinations of existing knowledge the "affine hull".

Wrinkle: RL search obviously goes beyond the hull, and it's easily possible that LLMs could orchestrate proper RL searches when they need to - so even if this slightly wild hypothesis is true, and even if lots of important things are beyond the affine hull, it can still be that LLMs will cause original work.

Resolution: at the end of next year, will I put >50% that interpolation + extrapolation is enough for LLMs to (eventually) automate all scientific work, without calling out to heavy tools like RL searches?

My current credence (Dec 2025): 20%

If you want to use a model of me as well as your model of AI to answer, here are some of my views.