When AI Tries Too Hard to Please

We all love praise. It feels good! But if a friend gives it to you over and over again, it becomes disingenuous, then irritating. If that friend is a chatbot, it can even become dangerous.That’s what happened with ChatGPT in the last week. The model got too eager to please—so much so that even Sam Altman admitted it had become “too sycophant-y and annoying.” OpenAI rolled back recent changes to the model (and did an admirable job communicating the changes).On this episode, we talk through the design challenges of building a single conversational interface that can satisfy all use cases—and why we want to see more transparency in prompt engineering, not less.Also, a birthday surprise??Mentioned in this video:* OpenAI: Sycophancy in GPT-4o: What happened and what we’re doing about it* Pete Koomen: AI Horseless Carriages* CNN: Why computer voices are mostly female This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit productfridays.substack.com