Import AI 445: Bostrom on when to race, and a math benchmark AI can't beat yet

Two items in Import AI 445 are worth pulling out. The first is Nick Bostrom's framing of superintelligence timing. He argues the choice is not safety versus danger but one risky path against another, since the status quo also kills people, especially in poorer countries. His phrase is "swift to harbor, slow to berth": move fast toward capability, then consider pausing only at the critical stages, because a pause that comes too early invites bad regulation and leaves society without advanced defenses.

The second is a cleaner empirical signal. Researchers built a genuinely held-out proof benchmark: ten unpublished problems from active research across algebraic combinatorics, spectral graph theory, and topology, chosen so the answers are not online. Neither GPT-5.2 Pro nor Gemini 3.0 DeepThink solved them. The test targets the creative leaps mathematicians make on problems that have no published trail to pattern-match. Read the full issue on Import AI.

Why it matters

If you track AI progress, the held-out math benchmark is the more useful instrument than the timeline debate. Saturated public benchmarks tell you less every month; a test built specifically so models cannot have seen it is the kind of measure that still means something.