Scott Viteri

Ontology maps

I am interested in coming up with semantics preserving maps between the internal state of agents, in part because you could use such a map to lift preferences from one agent onto another. However, there are various choices one might make along the path of designing and implementing such a map, and I will use this talk to bring up the pros and cons of taking these different choices. Each option that I will present is implementable as a differentiable loss function which can be used to train a neural network translating between the states of two state machines. I will talk about how this story fits into my view of the alignment problem, and how I want people to collectively become superintelligent instead of dying.

Keywords

• Agents
• Ontology