@(Anthony) Some quick thoughts on comparative metrics: 1. Scores relative to the community prediction like on [GJOpen challenges](https://www.gjopen.com/leaderboards/challenges/43-coronavirus-outbreak). I'd be very interested in leaderboards for cumulative relative score both normalized and not normalized by number of questions predicted. 2. A mode that attempts to penalize less for inactivity in updating forecasts on questions: can I compare my predictions to the community/others only at the time they were made? 3. Scores relative to other users only o...

An assessment of Kurzweil's predictions for 2019 made in 1999 was just posted on LessWrong.

His predictions were found to be 52% false and 15% mostly false, compared to 32% false and 14% mostly false for predictions made for 2009 in 1999. This is helpful for determining the base rate for the reference class of Kurzweil's long-term predictions.

First I estimate the likelihood of the experts winning on each of the 4 questions: 1.1: 98% (only need 19k more in last day, seems extremely likely) 1.2: 75% (number of weekly deaths only slightly down favors experts) 1.3: 80% (number of weekly deaths increased! favors experts more strongly) 1.4: 20% (expert's median seems very high, weekly deaths would have to increase for them to win which seems unlikely, and weekly deaths decreased this past week) I then calculate the likelihood of experts winning at least 3: .98 * .75 * .8 * .2 + .98 * .75 ...

Updated slightly up on https://www.gracenote.com/virtual-medal-table/

Edit: updated up to 67% based on this forecast realizing that the question is based on number of gold medals.

Edit 2: updated up to 70% based on https://www.topendsports.com/events/summer/me…

— edited by elifland_ought

How many parameters will GPT-4 have?

@Fruo I’m not sure that prior really makes sense given the difference between continuing to be in a civil war and the start of a civil war. One that makes more sense to me is: of the ~240 years in which there wasn’t a civil war ongoing, a civil war started once. This gives a prior of 1/240=~0.4%

Talking with @NunoSempere, @juancambeiro, @misha makes me think this is more of a 5-10%ish risk rather than 1-2%. In particular, a full-fledged civil war seems unlikely but there may be some substantial civilian riots that lead to 500 deaths, e.g. related to the transition of power in 2024/2028. Another factor pushing me slightly up is a [small but significant chance of transformative technology in the next 10 years](https://www.lesswrong.com/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines). While you could imagine tech advances making this less li...

There are no dates on these posts so I'm not exactly sure when they were run, but there are 3 more ads described in blog posts on the first page only [here](https://wellness.consumerfreedom.com/articles/) that were run in a newspaper required for the question. They are: * https://wellness.consumerfreedom.com/ny-times-dog-food-or-fake-meat/ * https://wellness.consumerfreedom.com/warning-plant-based-meat-can-have-this-carcinogen/ * https://wellness.consumerfreedom.com/fake-meat-real-chemicals-campaign/ Assuming the articles are ordered most recent first,...

@alexrjl Why would they predict over-confidently on Metaculus, even if they did so in the surveys? The metaculus predictions have no bearing on the actual results if I understand correctly.

I predicted 20% based on https://projects.fivethirtyeight.com/2020-ele… having an 80% confidence interval for the popular vote spanning approximately 3 points in each direction for Trump and Biden.

I just realized that my conversion to the 20% forecast was based on a basic math error: if Biden under-performs by 2 points, then Trump over-performs by 2 points, and this resolves positively (technically there are a few 3rd-party votes but safe to mostly ignore those).

Oops.

— edited by elifland_ought

I'd like to see more predictions on the effects of social media. I am intuitively pretty sympathetic to a lot of the claims made by sources such as: * [Center for Humane Technology](https://www.humanetech.com/) * [Tristan Harris on the 80,000 hours podcast](https://80000hours.org/podcast/episodes/tristan-harris-changing-incentives-social-media/) * [The Social Dilemma](https://www.thesocialdilemma.com/) * [Digital Minimalism](https://www.goodreads.com/book/show/40672036-digital-minimalism) But, I'm wary of the rigor of the arguments made and would be ex...

[There are currently 593,807 reported deaths](https://www.who.int/docs/default-source/coronaviruse/situation-reports/20200718-covid-19-sitrep-180.pdf?sfvrsn=39b31718_2). With the current rate of ~5k deaths per day, we will end up at ~1.43M. I'm surprised that the community median is significantly below this. [Made an Elicit distribution](https://elicit.ought.org/builder/QPLQTyEb2) to assign very little probability to <600,000, and 17% (intuitive guesstimate) probability to <1.43M as I think the number of deaths per day will rise. [I work on Elicit.]...

I used the SSA actuarial life table and adjusted the estimates slightly downward (since the Queen seems to be in fairly good health and has access to great care) to get a forecast for this. I used the percentiles from those calculations to make my forecast using Elicit. This also helped me put only 2% on her reign ending in the past (< Jun 2020), while the community has 6%.

[Disclaimer: I work on Elicit]

— edited by elifland

I concur with those below in their praise of the track record updates. I particularly like the continuous calibration plots.

As I'm always hungry for more, it would be cool to have a feature where I can (pay tachyons to? or have both people consent to) compare my track record vs. a rival forecaster's track record, on questions that we have both predicted on.

— edited by elifland_ought

@Anthony I think we should wait until the spreadsheet is updated for today (as someone who would greatly benefit from resolving at 931,698).

According to the Wikipedia article, there were 23 in 2020.

GJOpen currently has this at 90%

Predicted using an Elicit distribution based on a few intuitive beliefs regarding the current target launch date of Oct 31, 2021.

[I work on Elicit.]

It's important to note that [the re-election question](https://www.metaculus.com/questions/1100/will-trump-be-reelected-president-in-2020/) doesn't close until 6 PM EST on Election night. Given this, I created an [Elicit distribution](https://elicit.ought.org/builder/sQh12Ohw0) based on factorizing the question into a few parts: - (A) Will the result of an election be a large victory for Biden, a close call, or a large victory for Trump? - (B) Given the Biden large victory possibility in (A), how clear will it be by 6 PM EST election night that it wil...

@Linch thanks for the comment, made me realize I was reading the question incorrectly and predicting when a 10% reduction would happen (which I think has happened already) rather than a reduction to 10%. I wonder if others made the same mistake.

— edited by elifland