Blog
Let the model vote, not summarise
Wind Tunnel uses an LLM to deepen a question rather than answer it: a draft goes to a hundred AI personas and comes back as a map of positions. The step that makes the map possible is turning a hundred free-text reactions into a vote matrix, with the model as the voter.
Saying yes with the danger in mind
I read four open models with Anthropic’s Jacobian lens. Under social pressure, their replies agreed with risky plans while a word for the danger ranked near the top inside them — in up to 15 of 21 cases.