Generative Agent Simulations of 1,000 People

This paper presents several fascinating insights:

‍

1. A voice-enabled GPT-4 conducted two-hour interviews with 1,052 individuals.

2. GPT-4 agents were provided with the transcripts and prompted to simulate these people.

3. The agents were then given surveys and tasks, achieving 85% accuracy in simulating the interviewees.

‍

This is an earlier post where our parity study yielded the same results. This was a while back, we're past this now :-)

‍

‍