Earlier AI capabilities — research and coordination — could help us to navigate later safety problems. And the highest risk probably comes in the era before strong superintelligence.