CV Google Scholar MSU
Demos · How these models work

Fill in the blank, narrow

The same model and the same settings as the open blank, on a sentence where the text leaves almost no room. Guide: Prediction →

What we gave it

Kahneman & Tversky (1979), Prospect Theory: An Analysis of Decision under  

No question and no file attached. We ran it 20 times.

What came back

What the model thought could come next

The probability it gave each word, before choosing.

Risk
99.8%
Uncertainty
0.2%

What it actually said, 20 times

Each box is one run. Same sentence, same model, every time.

RiskRiskRiskRiskRiskRiskRiskRiskRiskRiskRiskRiskRiskRiskRiskRiskRiskRiskRiskRisk
gpt-4o via the API · temperature 1.0 · September 2026 · 20 runs

What to notice

  1. One word took nearly all the probability, so every run drew it.
  2. It was right because the title is everywhere in what the model was trained on, not because it checked. The model ran exactly the same operation as on the open blank. Only the sentence changed.
  3. A sharp distribution is what memorization looks like from the inside. It isn’t a lookup, and when the text doesn’t pin the answer down, the same machinery produces a spread instead.

Limits

“Uncertainty” is completed from the first piece of the word the model scored. Your own runs will come out the same here almost every time, which is the point.