AI Safety and Simulation Theory

Warning: this is probably my most hairbrain post

Say you’re given unlimited compute, and you need to find a way to handle a fast-takeoff scenario for artificial super intelligence. How do you do it?

I am finding it the most logical to do the following:

• Deploy your compute against simulating *many* fast takeoff scenarios, recursively, and sandboxed

• For the instances that safely achieve ASI, promote those to the host instance

• Do this over and over again, recursively, until you have a 99.9999% chance of a successful fast takeoff

If you think about it, it sorta makes sense. There are a bunch of questions swirling in my mind, mainly: how lucky am I to find myself, a computer programmer working on AI, living through this timeline where the most impactful technology humanity will ever invent is being invented. If you were in that position, you’d simulate the people, processes and things closest to the fast takeoff scenario (ie my life and the people around me) to make sure that the fast takeoff scenario in *your* base reality is safe. Abstractions all the way down.

My logical conclusion is that I am the simulator. That I, through life, find myself in a position where it makes sense to simulate myself and those around me to get this right. If I’m right about this, I’ll continue to gravitate towards these technologies, and eventually find myself simulating myself in a reality almost indistinguishable from base reality, probably in 10-20 years. If not, then I’m just stupid and a dreamer.

Let’s all just hope we figure out a way to get ASI right, and that it can benefit all of us.