About phosphine... Seems like it may have been a 12th-order polynomial fit producing spurious results. > [Re-analysis of the 267-GHz ALMA observations of Venus: No statistically significant detection of phosphine](https://arxiv.org/abs/2010.09761) > Context: ALMA observations of Venus at 267 GHz have been presented in the literature that show the apparent presence of phosphine (PH3) in its atmosphere. Phosphine has currently no evident production routes on the planet's surface or in its atmosphere. > Aims: The aim of this work is to assess the statist...

@Nostradamnus As far as I know, part of Metaculus mission is to collect users track record on short term question in order to use that record as a proxy for assigning credibility in long term predictions.

But most importantly, it's just an important question to ask :) .

— edited by isinlor

[Seroprevalence in Dheli seems to be around 20%](https://www.hindustantimes.com/delhi-news/over-23-of-delhi-s-population-affected-by-covid-19-shows-serosurvey-results/story-Sq1XXdPkfmDdggyNwGdlSK.html) based on [tests with 92% sensitivity and 98% specificity](https://twitter.com/DrJarlov/status/1285546389821325321/photo/1). Dheli reported overall 3628 deaths and has population of 19 million. $$ \frac{3600\ deaths}{0.2*19*1000^2\ infected} * 100\% \approx 0.1\%\ IFR $$ If these results hold under scrutiny this will significantly decrease my prediction....
@(Jgalt) wrote: > Given the tremendous cost and time overruns (already more than a decade late and billions of dollars over budget) I suspect NASA will have been extremely thorough in ensuring that the JWST is as likely as possible to succeed before they finally launch it. A critical failure after this much effort has been expended would be a public relations disaster for the agency and seriously damage public confidence in their competence. I agree 100% and this is also why I think this question is very important. And NASA made mistakes in calculating...
Poland is a good reference point. From European Commission: [First results of Poland’s Family 500+ programme released](https://ec.europa.eu/social/main.jsp?langId=en&catId=1246&newsId=9104&furtherNews=yes): > Under the programme parents can receive a tax-free benefit of PLN 500 (about EUR 120) per month for the second and any consecutive children until they reach the age of 18. The additional support is roughly 12% of the average gross wage in Poland in 2016. > Families are also eligible to receive the benefit for their first child if the family incom...
@Roko @Jgalt @alexrjl @Sylvain I've compiled a spreadsheet from ranking with points and resolved question count per user. https://docs.google.com/spreadsheets/d/1Fy7uqy_KgFAZhL_2br_O9_sV2vzki6Vzhzx-xvJ1_b4/edit?usp=sharing Seems like a ranking based on points per resolved question would look a lot different and would benefit a lot people with low number of resolved questions. Notable outliers with high current ranking: traviswfisher, datscilly, AndrewMcKnight, DanielG - their ranking would decrease only by 6 to 9 positions. Bernhard, tragedyofthecomme...

@Linch See this Twitter thread:

Our daily update is published. We’ve now tracked 5.2 million tests, up 300k from yesterday, another huge day.

Note that we can only track tests that a state reports.


After a time when it seemed testing numbers had settled down, there is a lot of change in the statistics coming out of the states right now. Testing is ramping; new systems are coming online; old reporting mistakes are being discovered; numbers are being restated.

I have a feeling that the community confidence comes from wrong reading of the NYC.gov data. My understanding of how this disease works in other places just gives me a low priori on sudden drop. It's easy to think that hospitalization will drop like a stone taking NYC.fov graph at face value. But I think there is more like 30% chance that it will dip below 200 in the next 9 days. And if it does, it will be mostly because of weekends drops and/or bigger than 7 day delay in reporting. Besides random fluctuations there doesn't seem to be clear downward tre...

@casens I don't believe it's possible even in principle to solve control problem. It's an ill posed problem, and it's highly unlikely that there could be any consensus around solution to that problem. But I think it's highly likely that AGI will be beneficial.

For example, corporations are super human entities that are generally beneficial. But corporate law is not a problem that can be solved. So I'm at 3% for this question.

— edited by isinlor

@Reprisal The question states:

Currently, with 53 seats, the Republican Party occupies a majority of the 100 seats in the Senate.

The 53 seats count does not include VP, even tough he is a republican.

Therefore, it's clear that the question can not count VP in 51 seats count needed for positive resolution and be internally consistent.

— edited by isinlor

@(Anthony) BTW - I suspect that there may be a small systematic problem with open questions. If I'm very uncertain about open question then quite often most likely outcome I can foresee is that I will loose points, so I don't predict at all. I remember looking at this question specifically and that I wanted to predict a lot wider tails - just because I don't really understand what nonfarm payrolls are, how they are computed etc. - but after looking at points distribution I've decided to not make any prediction at all. I'm curious if there are more pred...
I'm getting quite confident that Russia will not invade Ukraine. At least not Ukraine controlled territory and not right now. Russia will recognize independence of the separatists, then separatists will ask to join Russia or ask for Russian army presence. The army around Ukraine is in case Ukraine decides to respond to the attempt at annexation. In case Ukraine responds to annexation with force, Russian army will proceed to force Ukraine to surrender with siege of Kyiv. If Putin thinks he needs this whole show around recognizing the separatist, he wo...

In "Community stats":

The Metaculus prediction is 0.6%

Cool, Metaculus can go below 1% :) .

— edited by isinlor

Cool! I'm kind of warming up to the new design. I also see that some of my comments were addressed :) . There is one remaining thing that jumps at me - the choice of font for the question content. IMO it's really unappealing font that seems to conflict with the new design. Serif and rounded corners just don't go together. I would really, really want to see some non-serif font. I just randomly switched the font to Ubuntu and it looks so much better. You can compare the two here: https://imgur.com/a/o2kXqAS Additionally, I have a question: Was the ab...

The question hinges on the definition of "superhuman performance".

What does it mean to be superhuman? Does it mean better than any human in their respective field of expertise or just better than an average human across all field of expertise?

One is genuinely extremely hard. Another one is not that far away from what Google or GPT-3 is able to do if we think about question answering.

— edited by isinlor

It could be interesting to see how many points one could get using only naive strategies and by naive I mean anything that can be programmed. The simplest strategy would be to use as input only the community predictions. Some more complex solution could try to do some NLP and web scraping outside Metaculus. I'm curious what results would we get if we flooded Metaculus with predictions from naive strategies. Where in the ranking would the best naive strategy be? How different would be the community prediction if we flooded Metaculus with naive strategies...

Has anyone an idea what's happening in Japan? The number of confirmed cases there seems to be growing linearly!

Take a look at this dash board. Go to panel 3 (the bar at the bottom) and select Japan from the left panel. Almost exactly linear!

Unless it's has something to do with testing, it's a glimpse of hope that taking hold of this disease is possible outside China too.

NVIDIA is experimenting with architectures to train [GPT-3 with 1 trillion parameters](https://arxiv.org/abs/2104.04473). > Let us consider the GPT-3 model with P =175 billion parameters as an example. This model was trained on T = 300 billion tokens. On n = 1024 A100 GPUs using batch-size 1536, we achieve X = 140 teraFLOP/s per GPU. As a result, the time required to train this model can be estimated as: (...) just over a month. For the 1 trillion parameter model, assume that we need about 450 billion tokens for end-to-end training. With 3072 A100 GPUs...