"We're basically teaching our models to chase dopamine instead of truth," said Edwin Chen.