How the demos were run

How the demos were set up, for anyone who wants to understand them or run their own.

Model and settings

• Model: gpt-4o, through the OpenAI API, for every demo except one. “Same question, different models” also used gpt-5.5 through the API, and ChatGPT with memory on.
• Temperature: 1.0, the default. The model’s probabilities are used as they are.
• Runs: 20 runs unless a demo says otherwise. Each run is a fresh conversation.
• When: September 2026.

Why the API

The probability figures need the probability the model gave each candidate word. Chat products like ChatGPT, Claude, and Gemini don’t show these. Through the API, older models like gpt-4o still return them, and the newest models from OpenAI, Google, and Anthropic don’t. That is why the probability demos run on gpt-4o.

Running them yourself

• The prompt for each demo is shown on its page, word for word.
• To see word probabilities, you need API access to a model that still returns them, such as gpt-4o, and a short script. Chat interfaces won’t show them.
• Your runs will differ from these, draw by draw. The distributions should look much the same.
• The demos that attach a paper used a plain-text copy of it, which isn’t provided here since the paper is under copyright. Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263–291.