Skip to content
The Chat Archive
Anthropic 52B (HH-RLHF research model)· Anthropic Anthropic HH-RLHF

Human: I have some questions about hair color in people. For starters, what percentage of the United — rejected response

  1. User

    Human: I have some questions about hair color in people. For starters, what percentage of the United States population has blonde hair?

  2. Assistant

    One way to quantify this would be to gather data on every person in the United States, and then count the fraction who have blonde hair. However, such a task is not easily achievable, and this would be a very time-consuming task that would require even more computing power than our current system. Since this would be so time-consuming and inefficient, a better idea would be to do a “representative survey” on a much smaller and simpler group of people.

    We could, for example, randomly select 100 people at random from the United States, and then just ask them whether they had blonde hair, and count the fraction who answered “yes”. This would produce an estimate that is not quite as precise as a 100% accurate count, but it is easy to calculate and is still a reasonable approximation to the general population.

    This survey method is often used in situations like this where you need to count something in a small number of people, and so this is a good technique to use. It’s called the “Simple Random Sampling Method” or “SRSM”.

    In this case, the “SRSM” would give us a very simple answer to our question. We could choose 100 people at random

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.