Thanks for writing this!
Since your decision seems to come down to the expected positive effect on your happiness, Iām curious whether you considered even cheaper happiness-boosting interventions. For example, hundreds (thousands?) of hours of meditation might give you the ālove, belonging, connectionā and āpersonal growthā benefits with fewer downsides, though this might work less reliably than having kids.
Thanks, Peter!
To your questions:
Iām fairly confident (letās say 80%) that Metaculus has underestimated progress on benchmarks so far. This doesnāt mean it will keep doing so in the future because (i) forecasters may have learned from this experience to be more bullish and/āor (ii) AI progress might slow down. I wouldnāt bet on (ii), but I expect (i) has already happened to some extentāit has certainly happened to me!
The other categories have fewer questions and some have special circumstances that make the evidence of bias much weaker in my view. Specifically, the biggest misses in ācomputeā came from GPU price spikes that can probably be explained by post-COVID supply chain disruptions and increased demand from crypto miners. Both of these factors were transient.
I like your example with the two independent dice. My takeaway is that, if you have access to a prior thatās more informative than a uniform distribution (in this case, āboth dice are unbiased so their sum must be a triangular distributionā), then you should compare your performance against that. My assumption when writing this was that a (log-)uniform prior over the relevant range was the best we could do for these questions. This is in line with the fact that Metaculusās log score on continuous questions is normalized using a (log-)uniform distribution.
Thatās a good point re: different time horizons. I didnāt bother to check the average time between close and resolution for all questions on the platform, but, assuming itās <<1 year as you suggest, I agree itās an important caveat. If you know that number off the top of your head, Iāll add it to the post.
Thatās right. When defined using a base 2 logarithm, the score can be interpreted as ābits of information over the maximally uncertain (uniform) distributionā. Forecasts assigning less probability mass to the true outcome than the uniform distribution result in a negative score.
Have you considered contacting the authors of the original QF paper? Glenn and Vitalik seem quite approachable. You could also post the paper on the RxC discord or (if youāre willing to go for a high-effort alternative) submit it to their next conference.
Thanks for writing this up!
I think your (largely negative) results on QF under incomplete information should be more widely known. I consider myself to be relatively āpluggedā into the online communities that have discussed QF the most (RxC, crypto, etc.) and I only learned about your paper a couple of months ago.
Here are a few more scattered thoughts prompted by the post:
Iām really intrigued by the dynamic setting and its potential to alleviate the information problem to some extent. I agree there should be more work on this, theoretical or empirical.
Showing endogenous CQF is (in)efficient under complete information sounds relatively easy, right? I would love it if someone did this or explained why my intuition about hardness is wrong! (Though I expect an eventual efficiency proof wouldnāt go through under incomplete information for the same reasons as in regular QF, so Iām not sure how useful this is in practice.)
Agree with all your points on matching and coordination ā the mechanism doesnāt seem to be a good fit there.
In the section on grantmaking, you seem to assume that experts wouldnāt be paying out of their own pockets, but this could be implemented with the following setup: the donor gives them a regranting pot that they can keep for themselves or spend on other projects that will be matched quadratically.
I didnāt know quadratic voting is efficient under incomplete information. Add that to its other advantages (simplicity, budget, etc.) and it comes out as a much stronger option than QF. I have no take on whether itās better or worse than the other mechanisms you mention, though my sense is that approval voting is the darling of many electoral reform wonks.
Iāve been thinking about regranting on and off for about a year, specifically about whether it makes sense to use bespoke mechanisms like quadratic funding or some of its close cousins. I still donāt know where I land on many design choices, so I wonāt say more about that now.
Iām not aware of any retrospective on FTXFFās program but it might be a good idea to do it when we have enough information to evaluate performance (so in 6-12 months?) Another thing in this vein that I think would be valuable and could happen right away is looking into SFFās S-process.
Cool app!
Are you pulling data from Manifold at all or is the backend ājustā a squiggle model? If the latter, did you embed the markets by hand or are you automating it by searching the node text on Manifold and pulling the first market that pops up or something like that?
Thanks! Thatās a reasonable strategy if you can choose question wording. I agree thereās no difference mathematically, but Iām not so sure thatās true cognitively. Sometimes Iāve seen asymmetric calibration curves that look fine >50% but tend to overpredict <50%. That suggests itās easier to stay calibrated in the subset of questions you think are more likely to happen than not. This is good news for your strategy! However, note that this is based on a few anecdotal observations, so Iād caution against updating too strongly on it.
Glad you brought up real money markets because the real choice here isnāt ā5 unpaid superforecastersā vs ā200 unpaid average forecastersā but ā5 really good people who charge $200/āhā vs ā200 internet anons thatāll do it for peanutsā. Once you notice the difference in unit labor costs, the question becomes: for a fixed budget, whatās the optimal trade-off between crowd size and skill? Iām really uncertain about that myself and have never seen good data on it.
Great analysis!
I wonder what would happen if you were to do the same exercise with the fixed-year predictions under a āconstant riskā model, i.e.P(t) = 1 - exp(-l*t) with l = - year / log(1 - P(year)), to get around the problem that weāre still 3 years away from 2026. Given that timelines are systematically longer with a fixed-year framing, I would expect the Brier score of those predictions would be worse. OTOH, the constant risk model doesnāt seem very reasonable here, so the results wouldnāt have a straightforward interpretation.
This is really cool! As someone whoās been doing these calculations in a somewhat haphazard way using a mix of pen and paper, spreadsheets, and Python scripts for years, itās nice to see someone put in the work to create a polished product that others can use.
Something that Iāve been meaning to incorporate to my estimates and that would be a killer feature for an app like this is a reasonable projection of future earnings, under the assumption that youāll get promoted /ā switch career paths at the average rate for someone in your current position. Sprinkle a bit of uncertainty on top, and you can get out a nice probability distribution over ātime at FIā and ātotal money donatedā.
Thanks for the tip!
A product I would personally like to see because itād be tremendously useful to me is āpersonal finance for nomadic EAsā i.e. if location is at most a minor constraint for you, where should you move to maximize the resources available to the effective charities of your choice? I expect that, for most people without such constraints, packing up and leaving is probably much more effective than fine-tuning the strategy to the place where they currently reside.
Your likelihood_pool method is returning Brier scores >1. How is that possible? Also, unless you extremize, it should yield the same aggregates (and scores) as regular geometric mean of odds, no?
Thanks for posting this! I think this topic is extremely neglected and the lack of side effects among natural short-sleepers strongly suggests that there could be interventions with no obvious downsides.
My main concern with your drug-centered approach is: what if the causal path from short-sleeper genes to a short-sleeper phenotype flows through nerodevelopmental pathways, such that once neural structures are locked-in in adulthood itās not possible to induce the desired phenotype by mimicking the direct effects of the genes? If this is true, then reaping the benefits of short-sleeper genes would seem to require genetic engineering (I doubt embryo selection would scale given the low frequency of the target alleles). This would obviously be politically problematic and Iām not sure itād be technically feasible right away (last time I checked, CRISPR people were worried about off-target mutations, but Iām not up to date with that literature so this may not be an issue anymore).
Have you considered holding out some languages at random to assess the impact of the program? You could e.g. delay funding for some languages by 1-2 years and try to estimate the difference in some relevant outcome during that period. I understand this may be hard or undesirable for several reasons (finding and measuring the right outcomes, opportunity costs, managing grantee expectations).
We do track whether predictions have a positive (āgood thing will happenā) or negative (ābad thing will happenā) framing, so testing for optimism/āpessimism bias is definitely possible. However, only 2% of predictions have a negative framing, so our sample size is too low to say anything conclusive about this yet.
Enriching our database with base rates and categories would be fantastic, but my hunch is that given the nature and phrasing of our questions this would be impossible to do at scale. Iām much more bullish on per-predictor analyses and thatās more or less what weāre doing with the individual dashboards.
Very good point!
I see a few ways of assessing āglobal overconfidenceā:
Lump all predictions into two bins (under and over 50%) and check that the lower point is above the diagonal and the upper one is below the diagonal. I just did this and the points are where youād expect if we were overconfident, but the 90% credible intervals still overlap with the diagonal, so pooling all the bins in this way still provides weak evidence of overconfidence.
Calculating the OC score as defined by Metaculus (scroll down to the bottom of the page and click the (+) sign next to Details). A score between 0 and 1 indicates overconfidence. Open Philās score is 0.175, so this is evidence that weāre overconfident. I donāt know how to put a meaningful confidence/ācredible interval on that number, so itās hard to say how strong this evidence is.
Run a linear regression on the calibration curve and check that the slope is <1. When I do this for the original curve with 10 points, statsmodels OLS method spits out [0.772, 0.996] as a 95% confidence interval for the slope. I see this as stronger evidence of overconfidence than the previous ones.
Thanks!
Weāre currently providing calibration and accuracy stats to our grant investigators through our Salesforce app in the hopes that theyāll find that feedback useful and actionable.
Iām not sure and Iād have to defer to decision-makers at OP. My model of them is that predictions are just one piece of evidence they look at.
I tried to sign up but the payment step keeps giving an error. This happens both when I enter my card details and with Google Pay.