I believe in two things, perhaps now even more firmly than ever: a.) SFT is still the cheapest way to get what you want; b.) subtle forms of memorization, and the inherent, inevitable tension between the model's parametric knowledge and the conditioning context is worthy of further study.
