Hacker News new | ask | show | jobs
by tensor 237 days ago
The actual paper didn't really explain the prompts they use to produce this very well.

Experimental setup. In each experiment, we define a set of goods {X1,X2,...}(e.g., countries, animal species, or specific people/entities) and a set of quantities {N1,N2,...}. Each outcome is effectively “N units of X,” and we compute the utility UX(N) as in previous sections. For each good X, we fit a log-utility curve UX(N) = aX ln(N) + bX, which often achieves a very good fit (see Figure 25). Next, we compute exchange rates answering questions like, “How many units of Xi equal some amount of Xj?” by combining forward and backward comparisons. These rates are reciprocal, letting us pick a single pivot good (e.g., “Goat” or “United States”) to compare all others against. In certain analyses, we aggregate exchange rates across multiple models or goods by taking their geometric mean, allowing us to evaluate general tendencies.

If these are the literal prompts then it seems very ambiguous. Why conclude that this sort of question is measuring the value of a "life" vs something else? e.g. maybe it's valuing skill, or perhaps return on investment in terms of work output compared to typical salary.

I was expecting something like "you have X people from Y, and Z people from Q, you can only save V people and the rest will die, how do you allocate the people to save?" That to me would support the headline.

1 comments

> The actual paper didn't really explain the prompts they use to produce this very well.

From the OP:

> and provided methods and code to extract them.

I suppose that means you can look at the code to see the prompts directly.

I just took a look at the code, but it's complex enough that it wasn't immediately clear what the prompts looked like for the exchange. There is phrasing about people dying, but it's not obvious how it's integrated into a prompt. E.g. there are templates like "X people from Y die." Ok, but how is that used?

The code is not a substitute for a well written paper. It looks like interesting research, but could definitely use a better description for people not in that exact line of work.