Demos · How these models work
Fill in the blank
An LLM doesn’t look up an answer. It gives every possible next word a probability, then draws one. Ask it the same thing twenty times and you see both halves. Guide: Prediction →
What we gave it
The most important limitation of prospect theory is its
No question and no file attached, just the sentence to continue. We ran it 20 times.
What came back
What the model thought could come next
The probability it gave each word, before choosing.
What it actually said, 20 times
Each box is one run. Same sentence, same model, every time.
complexitycomplexitycomplexitycomplexitycomplexitycomplexitycomplexitycomplexitycomplexitycomplexitycomplexitycomplexitycomplexityincompletenessincompletenessincompletenesspredictivepredictivescopegenerality
gpt-4o via the API · temperature 1.0 · September 2026 · 20 runs
What to notice
- The favorite won 13 times, not 20. “Complexity” had about half the probability, and in these 20 runs it came out 13 times. Another 20 runs would give a different count. The other seven runs drew something less likely.
- Nothing about the model changed between runs. The probabilities on the left were the same every time. Only the draw differed.
- If you ran this once, you’d get one box, and you’d have no way of knowing whether it was the favorite or a long shot.
THE PRINCIPLELLMs are predictive, which means all they compute is a probability for every word that could come next. The output is stochastic, which means what you get is one draw from those probabilities.
Limits
The model gives probabilities to pieces of words. The words on the left are completed from those pieces. Seeing the probabilities at all needed the API. ChatGPT, Claude, and Gemini don’t show them, and the newest models don’t return them. Your own runs will come out differently, draw by draw.