DADGPT is my Halloween costume this year. Black t-shirt, black baseball hat, a phone sideways on the brim with the app running. People walk up, tap the mic, and ask Dad anything. He answers out loud with a cringey dad joke.

The browser turns your speech into text, gpt-6-luna writes the joke, and OpenAI text to speech reads it back in a dad voice.

The hard part was making the jokes. The first version answered the question with a pun tacked on, stacked three jokes into one reply, called everyone kiddo, and was mostly just not funny.

So I ran evals with 98 test chats then sorted 346 replies into ten failure modes. I used code checks and LLM judges to do improvement cycles. Humor continues to elude AI. I validated the jokes that were generated. They were mostly not good, but eventually iterated to where the jokes were mostly cringey in the right way.

Check it out at dadgptcostume.vercel.app and feel free to steal the costume!