- The greater prospect of “let’s have collaborative estimates of the impacts of key longtermist projects” is something I strongly want to see, but I think it’s also *really* difficult to do well.
- This experiment went through a few early strategies. I think the results are clearly mediocre (in that estimates were all over the place, and were wildly inconsistent), but could be a good place to build much better work.
- I see this very much as an MVP, so I’d expect it to have severe limitations. I generally prefer processes of “build a bunch of MPVs and test them out, and see what fails”, then one of “spend a whole lot of time getting it right the first time.”
- The fact that estimates were inconsistent suggests that elicitation is very difficult to do well, but also that there’s a great deal of improvement to be done. So, future work is probably less tractable than expected, but more important.
- I’m still very bullish on relative evalutions, but think that they will require a lot of clever innovations to do well.
- I think that longer-term, it would be promising to have people submit relative evaluations as long Squiggle (or similar) files. I’m unsure how these can best be displayed or organized for specific discussions.
Some thoughts on the estimations:
- I think this is the first most/any of us have really had to estimate the relative value of these kinds of longtermist projects. There’s been very little literature on this before. I think the numbers are correspondingly questionable (including my own).
- Utility elicitation for comparing one item to another that could be negative, in particular, was really poor. I tried some naive Squiggle calculations that clearly weren’t very accurate. I’m not sure what tool would be best here, maybe there’s some custom drawing-with-mouse tool that could work, or people could figure out better quantitative function representations.
- It’s very hard to evaluate these sorts of projects without much more data. Ideally, there would be a lot of data gathering. For example, if a program funds 10 people to do work, we’d ideally have a good table of all of their outputs, and comments from people in the area about how good these comments were. A lot of evaluation work can reduce to “effective systems to gather objective and subjective information from diverse sets of sources.”
- I believe people estimated “how valuable do you think this is” instead of, “how valuable do you think a council would think this is?” The latter should be much more uncertain, and possibly much more important to readers (if done well).
- From what I remember, I think my main disagreement with other evaluators is that some had much narrower ranges than I thought were reasonable. I guess that some of this is part of a learning process.
Some thoughts on the greater project:
- The greater prospect of “let’s have collaborative estimates of the impacts of key longtermist projects” is something I strongly want to see, but I think it’s also *really* difficult to do well.
- This experiment went through a few early strategies. I think the results are clearly mediocre (in that estimates were all over the place, and were wildly inconsistent), but could be a good place to build much better work.
- I see this very much as an MVP, so I’d expect it to have severe limitations. I generally prefer processes of “build a bunch of MPVs and test them out, and see what fails”, then one of “spend a whole lot of time getting it right the first time.”
- The fact that estimates were inconsistent suggests that elicitation is very difficult to do well, but also that there’s a great deal of improvement to be done. So, future work is probably less tractable than expected, but more important.
- I’m still very bullish on relative evalutions, but think that they will require a lot of clever innovations to do well.
- I think that longer-term, it would be promising to have people submit relative evaluations as long Squiggle (or similar) files. I’m unsure how these can best be displayed or organized for specific discussions.
Some thoughts on the estimations:
- I think this is the first most/any of us have really had to estimate the relative value of these kinds of longtermist projects. There’s been very little literature on this before. I think the numbers are correspondingly questionable (including my own).
- Utility elicitation for comparing one item to another that could be negative, in particular, was really poor. I tried some naive Squiggle calculations that clearly weren’t very accurate. I’m not sure what tool would be best here, maybe there’s some custom drawing-with-mouse tool that could work, or people could figure out better quantitative function representations.
- It’s very hard to evaluate these sorts of projects without much more data. Ideally, there would be a lot of data gathering. For example, if a program funds 10 people to do work, we’d ideally have a good table of all of their outputs, and comments from people in the area about how good these comments were. A lot of evaluation work can reduce to “effective systems to gather objective and subjective information from diverse sets of sources.”
- I believe people estimated “how valuable do you think this is” instead of, “how valuable do you think a council would think this is?” The latter should be much more uncertain, and possibly much more important to readers (if done well).
- From what I remember, I think my main disagreement with other evaluators is that some had much narrower ranges than I thought were reasonable. I guess that some of this is part of a learning process.