AI Safety and Philosophy
Esther Huang
AI reliability, from infrastructure to behavior.
I'm building Typicity around AI characters people can come to know over time. There is something I keep returning to: a character should be changed by what happens, yet remain recognizable through those changes.
A choice that once felt right may come to feel wrong. Sometimes the story has given the character a reason to change; sometimes the model has simply drifted. I'm interested in how we tell the difference, and how we build systems that can help us see it.
Typicity
AI characters people can get to know over time.
Building Typicity
Part of the work is learning to notice when a character has drifted. A new model may carry the conversation differently, even when the character's description has stayed the same. I'm working on ways to make those changes easier to see and evaluate.
Previous projectsWriting and notes from along the way.
Notes on how AI behaves. Some questions lead into philosophy.
Recent writing
Building in public