Ah, gotcha. Yeah, it was meant mostly as an aside and one that strengthened my praise for Vavilov Day (as indicating that this is appealing even to someone who dislikes most rationalist holidays), but I suppose the dislike was too controversial and/or too flippant.
I may write a post of my own describing why I don't like rationalist holidays/think they can do better, but I think that post would itself likely be extremely controversial so I'd have to approach it carefully.
During their brief masking-optional pilot, the school reported that “smiling is more contagious than covid-19,” and a survey of students found that 70 percent said the policy improved their experience, including their ability to learn.
That only 70% of kids said that getting rid of an annoying thing was good from their point of view is to me surprisingly low, and an update in the opposite direction vs what was seemingly intended by those who shared that result.
Then again, none of these statistics ever mean much at all without looking at the survey instrument and such.craigmichael on Covid 1/27/22: Let My People Go
In my model, it made sense take precautions to buy time to get vaccines and treatments. We have those now, so the cost-benefit calculation changes. Further wearing a mask is not going to help people avail themselves of them, or encourage institutions to approve existing treatments or accelerate new ones.
I'm grappling with the NNT numbers mentioned above. It's very much worth it to me to pay $10 for fluvoxamine to say cut the duration of omicron (and perhaps probability of long covid, however small it is) by, say, ~30% (that's my estimate). It's even worth it for me at that price to buy enough to not have to watch friends and family suffer longer if they choose to avail themselves.craigmichael on Covid 1/27/22: Let My People Go
Peterson’s statement to open Rogan that ‘climate is everything therefore your models are useless’ as a fully general argument against anyone ever knowing anything.
I appreciate you drawing attention to this. It does fall pretty flat the moment you start to think about it. It's trivially easy to imagine a very poor model of climate and then imagine small tweaks to make it better, and then iteratively improve this overtime. Like we would all imagine and be familiar with. It is a very anti-Bayesian statement from someone who should know better.chris_leong on The Nature of Counterfactuals
I think it's crucial to note that it's not the presence of the unicorns world that makes trouble here, it's the absence of all the other ones here. So what you're gesturing at here is I think the need for a kind of plenitude in the possibilities one believes in.
I would express this as it not making sense to include some worlds without others.
I think you are underestimating how much I think falls into these categories. I suspect (although I do not know) that much of what you would call being dishonest to oneself I would categorize into a) or b).
(General PSA: although choosing a career that encourages you to develop your natural tendencies can be a good thing, it also has downsides. Being someone who is on the less trusting side of things at the best of times and works in embedded hardware with an eye toward security... I am rather acutely aware of how much information leakage there is from e.g. the phone in your pocket. Typical English writing speed is ~13 WPM. English text has ~9.83 bits of entropy / word. That's only, what, 2.1 bits / second? That's tiny.)
(I don't tend to like the label, mainly because of the connotations, but the best description might be 'functionally paranoid'. I'm the sort of person who reads the IT policy at work, notes that it allows searches of personal devices, and then never brings a data storage device of any sort to work as a result.)
Good point about cognitohazards. I'd say: Beware self-fulfilling prophecies.
Could you please elaborate?
https://eprint.iacr.org/2021/1064.pdf ("whoops, brightness is a sidechannel attack to recover audio because power supplies aren't perfect". Not typically directly applicable for a phone, but still interesting.)
https://github.com/ggerganov/kbd-audio/discussions/31 ("whoops, you can deduce keystrokes from audio recordings" (which probably means you can also do the same with written text...))
https://www.researchgate.net/publication/346954384_A_Taxonomy_of_WiFi_Sensing_CSI_vs_Passive_WiFi_Radar ("whoops, wifi can be used as a passive radar")
https://www.hindawi.com/journals/scn/2017/7576307/ ("whoops, accelerometer data can often be used to determine location because people don't travel along random routes")
https://en.wikipedia.org/wiki/Words_per_minute#Handwriting (Actually, that's copying speed, but close enough for a WAG.)
https://www.sciencedirect.com/science/article/pii/S0019995864903262 (May be different for this particular sort of sentence than typical English, but again, WAG.)
Or rather, modern data storage is huge, and modern data transmission is quick. It doesn't take much correlation at all to inadvertently leak a bit or two per second when you're transmitting millions (or billions) of bits a second. (And that's assuming P(malice)=0...)
This sort of calculation is also why I don't put too much weight in encryption schemes with longer-lived keys actually being secure. If you have a 512b key that you keep around for a month, you only need to leak a bit an hour to leak the key... and data storage is cheap enough that just because someone didn't go through the work today to figure out how to recover the info, doesn't mean they can't pull it up a year or decade from now when they have better or more automated tooling. There's a side question as to if anyone cares enough to do so... but another side question of if that question matters if the collection and tooling are automated.
I've recently gained a better appreciation for how astonishingly good this work is at linear perspective, which had only come about in European art in the prior century. Many things about this painting are good (and some bad to my eye, like the messy color scheme) but those hexagonal details on the curved arches in perspective is 100% Raphael showing off.
An aside, but linear perspective is the most rational part of art, in the older philosophical sense of rational; it's pretty much the only major part of classical art which descends from first principles rather than having an empirical basis.tag on Coping with Undecidability
But (for convenience?), mathematicians/computer scientists use Turing machines/Lambda Calculus/(your favorite programming language) for most purposes instead to model things, even though they have these really peculiar undecidability properties.
FSMs also have undecidability: for each one there an infinity of programmes they can't even run.chris-maloof on ELK First Round Contest Winners
Really pleased to see that so many good proposals have come in! And kudos to unselected applicants for taking on the challenge as well.
And thanks to ARC for the contest. I've just spent a lot of time fumbling down (what seem to be) blind alleys, which may or may not say something about my fitness for the field, but I've also gotten my head around a few basics of modern ML in the process.
I hope the contest is achieving its goals for ARC too.tlw on Use Normal Predictions
(If this makes no sense, then ignore it): Using an arbitrary distribution for predictions, then use its CDF (Universality of the Uniform) to convert to U(0,1), and then transform to z-score using the inverse CDF (percentile point function) of the Unit Normal. Finally use this as zi in when calculating your calibration.
This glosses over an important issue: namely, how do you find (and score once you do find it) which distribution is correct?
(This issue occurs surprisingly often in my experience - one person thinks that a distribution has a thin tail, and another thinks that the distribution has a fat tail, but there's little enough data that both estimates are plausible. Of course, they give very different predictions for long-tail events.)
(Also, P(0 votes for Biden) is arguably significantly more likely than P(1 vote for Biden), in which case even the example chosen suffers from this issue of choosing the wrong distribution.)