This source only provides a headline — there is no body to show. Use “Read the original” below to read it at the source.
Using AI to solve most of alignment is a fundamentally doomed approach: an intuition
To the extent the AI labs can be said to have a plan, they want to use swarms of individually non-superhuman AIs to align superintelligence [1] . That's as if NASA tried to launch astronauts to the moon by tying them to a bundle of a million miniature rockets that ignite in the general direction of the moon. That will not work. Here’s why. A rocket is made of parts, and none of them is individually pointed at the moon Imagine you're building a rocket that you want (1) not to explode and (2) to land on the moon. An individual turbine in the engine is not pointed at the moon. The tanks aren't. There are instructions in the rocket's computers; there are sensors and actuators; there are wires connecting systems; on their own, none of them would reach the moon. You design all the parts as parts of a whole. The rocket is also not built of many small rockets all pointing at the moon. Yet all the thrusters, and tanks, and massive fins, and the controllers, and the various parts of the computer…