Charlie Steiner

Where would an AI find human values?

In many strategies, we want to build AI that learns human values, but what does that even mean? Okay, the answer isn't too hard: human values are latent variables in a model of humans. I attempt to introduce what this means for inference, collection of training data and human feedback, how to distribute control of the galaxy, whether our future will be weird, and some research directions for large language models.

Keywords

• Value learning