Powerful AI — smarter than Nobel laureates across most fields — likely arrives around 2026-2027. It could compress a century of progress into a decade, curing diseases and lifting the poor. It also carries serious catastrophic risk, which I'd put roughly in the 10-25% range. So we scale, but responsibly: interpretability, evaluations, safeguards.
Scaling laws keep delivering, and the upside is staggering. But the same systems are hard to control and easy to misuse, so safety investment has to scale with capability rather than lag behind it.
Either scaling clearly plateauing far short of this, or safety and interpretability failing so badly that responsible development turns out to be impossible.