I’m estimating sequential experiment depth. could be wrong but I expect fast saturating returns from superintelligence that leave uncertainty about untested physical processes annoyingly high. still lower enough to allow demigod-like bioweapons that kill us immediately if we screw up alignment, but we’re discussing drastically more intense meddling with a large complex adaptive system. I think there will be slow experiments that are needed to resolve uncertainty the physics simulator can’t fully resolve even after you’ve got a cell sim, because scaling approximate cell sim up to full body sim results in a lot of accumulated uncertainty. it might take as little as 3 months, but even for a saturated asi on the entire Google datacenter network, I’d be shocked if it takes minutes or days with only targeted experiments to send to labs. there’s too much emergent behavior to map.
I’m estimating sequential experiment depth. could be wrong but I expect fast saturating returns from superintelligence that leave uncertainty about untested physical processes annoyingly high. still lower enough to allow demigod-like bioweapons that kill us immediately if we screw up alignment, but we’re discussing drastically more intense meddling with a large complex adaptive system. I think there will be slow experiments that are needed to resolve uncertainty the physics simulator can’t fully resolve even after you’ve got a cell sim, because scaling approximate cell sim up to full body sim results in a lot of accumulated uncertainty. it might take as little as 3 months, but even for a saturated asi on the entire Google datacenter network, I’d be shocked if it takes minutes or days with only targeted experiments to send to labs. there’s too much emergent behavior to map.