Extremely technical obscure questions related to what they in particular have recently read.
Eg questions on “efficient” ways to reduce the dimensionality of a very large weight matrix. Answer isn’t even LORA rank decomposition it’s randomly removing rows and columns. Why can this work? Some obscure mathematics field I’d not heard of even with a masters in physics and PhD in ML. Very frustrating imo, but in fairness it does reflect what they were looking for.