I'm tired of these posts; LLMs are good for happy-path demos, that's it. And even then, their success rate depends on the prompter already knowing the answer!
Literally any out-of-distribution project in which I used LLMs lead to catastrophic failure. The models can't "see" stuff outside their training data.
You don't need to know the answer, just be able to recognize it. There are many situations where you can judge the output a lot easier than you can produce it, and LLMs are quite useful for those.
As a former Theoretical Physicist, this result is remarkable. I myself I tried to use AI for calculations in Perturbative Quantum Field Theory and I was impressed. I agree with the authors: it looks like the future of Theoretical Physics would be more in verification and consistency checking of AI-assisted results rather than in manual calculations.
4 comments
[ 3.6 ms ] story [ 15.8 ms ] threadLiterally any out-of-distribution project in which I used LLMs lead to catastrophic failure. The models can't "see" stuff outside their training data.