Computation, Communication, and Ontology Maps
Abstract:
How can we create intelligent systems that retain and expand into that which we find valuable in the universe? During this talk I will present my own thoughts based on ontology mapping, and I will communicate why I believe that mathematicians who think about systemic interactions are in an especially good place to answer this important question. I will start with a framework in which read-eval-print loops (repls) form a basis for reasoning about agents and computation in general. Then I will build ontology maps as an alignment framework on top of repls, and discuss its implications. Lastly I will invite discussion on what we might want out of the future, and talk about my thoughts on the central role of communication and its relation to both repls and ontology maps.
Scott Viteri
PhD Candidate in Computer Science
I am a PhD candidate at the Center for Automated Reasoning at Stanford University, under the guidance of Prof. Clark Barrett. My research focuses on improving the prosocial tendencies of language models (LMs) through a series of unique developmental approaches. This includes the introduction of communication channels during autoregressive training (akin to a kindergarten setting), allowing a parent LM to guide a child LM by curating its training data, and enhancing human feedback on LMs via a combined embedding of EEG data and speech.