There is a pattern where AI safety researchers have historically been skeptical that current AI systems pose near-term risks, but recent models' capabilities (superhuman performance on math and coding) have surprised them, providing some evidence against worst-case scenarios about alignment but not strongly disproving concerns.