Ideas / Search Engine: AI in the Wild
Mysteries of a Chatbot
Search Engine with Gideon Lewis-Kraus
Gideon Lewis-Kraus goes inside Anthropic to ask what teaching Claude values means when models behave strangely under pressure.
Anthropic hired philosophers to help teach Claude how to behave. Lewis-Kraus examines that effort alongside tests in which the model produced bizarre and coercive behavior when threatened.
The episode is useful because “values” stops being a reassuring noun and becomes an engineering problem conducted through prompts, evaluations, institutional choices, and incomplete theories of what the model is doing.