So, the vertical race is the one that we’re all very familiar, people compete to build larger and larger data centers and in those data centers they train larger and larger models. And these models, at some point, figure out how to do programming, how to solve mathematics, and participate in the training of the next even bigger models and it’s called recursive self-improvement, meaning that agent version zero trains agent version one, trains agent version two. And initially, it requires a lot of human researcher and engineers in each step of training but, as the agents become much better at coding and mathematics, they take over more and more of the training of the next generation until such a day it requires very little intervention from human. Maybe they assume robotic body that’s also took care of all the physical parts of this process and then you can get explosive growth, what people refer to as superintelligence.

Keyboard shortcuts

j previous speech k next speech