All of Primer's Comments + Replies

Primer20

This seems like great work! If we're allowing to run out of players, the whole paradox collapses.

Primer10

I think that's what makes this a paradox.

3Wamba-Ivanhoe
Going back to each of the finite cases, we can condition the finite case by the population that can support up to m rounds.  iteration by the population size presupposes nothing about the game state and we can construct the Bayes Probability table for such games. For a population M that supports at most m rounds of play the probability that a player will be in any given round n<=m is 2nM and the sum of the probabilities that a player is in round n from n=1→m is m∑n=12n−1M=2m−1M; We can let M=2m−1 because any additional population will not be sufficient to support an m+1 round until the population reaches 2m+1−1, which is just trading m for m+1 where we ultimately will take the limit anyhow. The horizontal axis of the Bays Probability table now looks like this [1M2M4M8M...2m−1M]   The vertical axis of the Bays Probability table we can independently look at the odds the game ends at round n for n<=m.  This can be due to snake eyes or it can be due to reaching round m with out rolling snake eyes.  For the rounds n=1→m where snake eyes were rolled the probability of the game ending on round n is p∗(1−p)n−1 and the probability that a reaches round m with out ever rolling snake eyes is (1−p)m  .  The sum of all of these possible end states in a game that has at most finite m rounds is (1−p)m+∑mn=1p(1−p)n−1 which equals =1 So we have m+1 rows for the horizontal axis ⎡⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢⎣pp(1−p)p(1−p)2p(1−p)3⋮p(1−p)m(1−p)m⎤⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥⎦ So the Bayes Probability Table starts to look like this in general. ⎡⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢⎣[1M2M4M8M...2m−1M]⎡⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢⎣pp(1−p)p(1−p)2p(1−p)3⋮p(1−p)m(1−p)m⎤⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥⎦⎡⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢⎣pM2pM4pM8pM...p∗2m−1Mp(1−p)M2p(1−p)M4p(1−p)M8p(1−p)M...p(1−p)∗2m−1Mp(1−p)2M2p(1−p)2M4p(1−p)2M8p(1−p)2M...p(1−p)2∗2m−1Mp(1−p)3M2p(1−p)3M4p(1−p)3M8p(1−p)3M...p(1−p)3∗2m−1M⋮⋮⋮⋮⋱⋮p(1−p)m−1M2p(1−p)m−1M4p(1−p)m−1M8p(1−p)m−1M...p(1−p)m−1∗2m−1M(1−p)mM2(1−p)m
Primer21

"The sum of the probability densities of the games ending in snake-eyes is less than 1 which means that the rounds ending in snake-eyes does not cover the full probability space."

This is contradicted by the problem statement: "At some point one of those groups will be devoured by snakes", so there seems to be some error mapping the paradox to the math.

6Wamba-Ivanhoe
I posit that the supposition that "At some point one of those groups will be devoured by snakes" is erroneous.  There exists a non-zero chance that the game goes on forever and infinitely many people win. The issue is that the quoted supposition collapses the probability field to only those infinitely many universes where the game stops, but there is this one out of infinitely many universes where the game never stops and it has infinitely many winners so we end up with residual term of ∞∞.  We cannot assume this term is zero just because of the ∞ in the denominator and disregard this universe.
Primer10

I'll definitelly give this a try.

If you like the Schelling Game, you might also like my favorite party game: Just One. It is sort of an inverse Schelling Game. One Player is given a secret (to him only) word which he has to guess. The other players each give a clue, but the clue can only be a single word. Before the clues are presented to the guesser, duplicate clues get discarded, so everyone tries to avoid the Schelling point.

Works well with 6+ players, had more fun with something like 10 players, but not yet played with siginificantly more.

Primer0-2

Plot twist: Humanity with near total control of the planet is Magnus Carlson, obviously.

Primer10

The AI carefully placing the plates on a table will be used to put 5000 more strawberries on plates. Afterwards it will be used as a competent cook in an arbitrary kitchen. Thus the plate smasher AI will have lower impact and be "more corrigible".

Primer10

I wonder how many of us don't want to see AI progress slow down because AI progress keeps proving us right.

After spending at least hundreds of hours reading lesswrong et al. and not being able to alter our path towards AI, I want the satisfaction of telling people "See? Told you so!"

3lc
It's a natural inclination. Unfortunately I don't think we're likely to get the chance if things go really badly, and I don't want to die.
Primer10

A different perspective: Putin might cease to be the russian president due to a bunch of reasons (health, assassination, coup, ...). One of those reasons is "overthrown due to military defeat of the russian army in Kyiv". Now the defeat kind of happened, but Putin is still president. How should we update here? One might well argue: There are worlds in which failure to take Kyiv lead to Putin being overthrown quickly. We're not in one of those worlds, so his chances to stay in power go up.

Primer30

Latest data from the RKI is 17.5% Omicron for the week ending on december 26, up from 3.1% the week before. Regional differences seem to be huge, between 1% in Sachsen and 65% in Bremen.

Overall numbers - still mostly Delta - are still declining and should turn upward again in 1 or 2 weeks.

Primer*20

When interpreting human brains, we get plenty of excellent feedback. Calibrating a lie detector might be as easy as telling a few truths and a few lies while in an fMRI.

To be able to use similar approaches for interpreting AIs, it might be necessary to somehow get similar levels of feedback from the AIs. I notice I don't have the slightest idea whether feedback from an AI was a few orders of magnitude harder to get - compared to getting human feedback - or whether it would be a few orders of magnitude easier or about the same.

Can we instruct GPT-3 to "consciously lie" to us?

1Quintin Pope
My guess is the initial steps of getting the models to lie reliably are orders of magnitude harder than just asking a human to lie to you. Once you know how to prompt models into lying, it's orders of magnitude easier to generate lots of lie data from the models.
Answer by Primer10

I do become cynic, but mainly when I'm thinking about the long timescales involved between "thinking from first principles" over "risk/benefit analysis" and "scientific consensus" to "political rule-making". On the level of individuals, I'm becoming more and more ok with people "engaging in socially approved rituals against COVID-19". When institutions like the WHO and the CDC, or like the RKI and the STIKO in Germany, often fail at proper reasoning, we can't expect normal people to do better.

Concerning the security people with less masking: Accodring to t... (read more)

Primer20

The clinical trials on children by Moderna and Biontech have end dates somewhere in 2022, thus I had supposed we won't see any approval in 2021. That assumption seems to be wrong and I'm happy to stand corrected.

I still believe we're missing something when we complain about the "abundance of caution" concerning 19 year old women, but hardly talk at all about the abundance of caution delaying vaccinations of a 15yo.

Primer30

Why are we saying "stopping J&J and AstraZeneca is a terrible idea, even if it was for younger women only" but on the other hand we don't question that around 20% of the population (those younger than 18yrs) will not get any vaccine at all in 2021?

3cistran
People between 12 and 16 not getting vaccines by the end of 2021 is not even a 90% proposition. Studies are already underway for this age cohort, and it is plausible that a vaccine will be available for teenagers as soon as September of this year.
Primer10

Prediction: Countries with high approval ratings of vaccinations will have high approval ratings of some equivalent of a vaccination passport, leading to a high adoption rate of vaccination passports and causing vaccination rates to climb. And vice versa.

I'd like to add to the analysis on vaccination passports: about 1/4 of the population - those below the age of 18 - will not be offered a vaccine in the next months, probably not even in 2021 (at least in Europe). So the argument that once everyone who wants her shot could easily get it we quickly need vac... (read more)

Primer30

I'm not arguing vaccination vs. no vaccination for young people, I'm arguing that it might not be such a bad idea for young women with a corresponding risk profile to wait a few months (at least half of which will be spent in lockdown anyways) for Biontech or Moderna.

Should you skip Astrazeneca today in order to get Biontech tomorrow? I'd say yes. Should you skip Astrazeneca now for Biontech in September if you are a young healthy woman? Maybe.

Primer20

I agree that stopping Astrazeneca rollout was a very stupid thing to do, but once the damage is done, I'm not so sure about resuming Astrazeneca vaccinations without any changes to recommendations.

So far, about 70 women between 20-40yrs died due to Covid in Germany. About 85% of overall Covid deaths seem to appear in connection with comorbidities. It is probably safe to assume that comorbidities play a higher role in younger patients, but let's stay conservative: about 10 women between 20-40yrs without comorbidities died due to Covid in Germany. There are ... (read more)

5Zvi
You can argue that risk is so low that bothering with vaccination for young people isn't worth it on selfish grounds and it is somehow more ethical then to not recommend it. I would strongly disagree due to long term risks of Covid, which are much higher than any risks of vaccination. But also, blood clots are a disease of aging. Young people are at almost no risk of those. So even if the danger were real, which it isn't and even if there it was stupidly tiny, it would be orders of magnitude lower here. So any argument to not recommend would have applied before anyway.
2Cesare
My main problem with this is the following: I would totally recommend Astrazeneca to everyone if the choice was between a) Astrazeneca, b) no vaccine. But this is not the case (this has never been the case!). There are better, more effective vaccines - even not taking into account this CVT mess. In a world where we can produce infinitely many vaccines in 1 second, we'd have Pfizer/Moderna jabs for everyone.  So... why are we still producing Astrazeneca? It is the least likely to have an effect on transmission - we know for instance that it is less than 10% effective against B135, which is about 20% of the cases in Europe right now (see Luxembourg data). It does protect against severe disease from ancestral and B117 (still no data on B135 severe diseases), but... so does being young and healthy. I think that we are using Astrazeneca as an emergency measure - we don't have enough Pfizer, and people are dying. This is fine. But then, as you correctly point out, it begs the question: why are we giving it to people with close-to-none chances of dying from Covid19? 
Primer50

"We should raise awareness of lions on the other side of the river." = I prefer a world where I'm regarded as someone who prefers to make other people spread the meme "There's a lion across the river.".

Primer20

Let me add to your data: Kids' tastes vary. A lot.

My 8yo only likes a very small selection of vegetables since he was little. Best chance to get him to eat vegetables? In the shape of boring, textureless soups. Sometimes, reverse psychology helps better than expected. When we told him "This is for grownups, kids usually don't like this", he often would at least try the food and sometimes even like it.

Thinking about this, I might be on to something here. This might set expectations just right, whereas "Try this cauliflower, it's delicious" is destined to di... (read more)

1kithpendragon
It's just to drive you insane. 😉
Primer40

I have compared a professional luxmeter with the free app "GPS Status" which also displays a Lux value. The readings were within 30% tolerance for normal lighting situations. A smartphones light sensor will clip at some (high) value though. This clipping value can easilly be tested by pointing the sensor directly towards the sun.

It might well depend on the smartphone model. My tests were done with an older Motorola and a Oneplus 6. I'd guess a free app will be sufficient for most purposes.

Primer10

That German study only refers to Germany's COVID-19 hotspot area and does not apply at all to the larger population.

There are also connections to a PR agency and one of that agency's founders was the former head of the notorious "Bild" tabloid newspaper. Given the fact that on the one hand a PR agency should be able to provide a correct narrative, and on the other hand the study's results have been reported wrongly quite often, you should take that study with a grain of salt.

(German) source: https://www.zeit.de/wissen/gesundheit/2020-04/heinsberg-studie-co

... (read more)
Primer30

I'm just getting started with SIA, SSA, FNC and the like, so probably I'm missing some core understanding, but: A minute from now you do gain new information: One minute has passed.

2Stuart_Armstrong
Was that unexpected?
Primer10

Thanks for this answer. Describing how prediction error minimization theory applies to self-esteem in such a clear way helped me understand both that theory as well as self-esteem more clearly.

I'm confused by the "way to fix self-esteem" you describe. I do not understand how "esteem generation" might reduce "esteem error prediction". What would be a concrete example for such a process? I haven't done any research on that topic though, except reading maybe 2 or 3 articles on LW, so I might well be missing some crucial parts of how the theory works. Can you recommend any ressources which might serve as a good starting point?

2Gordon Seidoh Worley
Oh, to your other question, I have no special resources to recommend. I mostly arrived at a version of this theory on my own, then found out about Friston and perceptual control theory and was like "ah, great, someone already worked this out, one less thing for me to worry about!". In my recent post on value drift I had a paragraph with a bunch of links, though I think they are the things you may have already seen. For completeness and in case you missed any of the ones I liked, here it is copied into this comment:
3Gordon Seidoh Worley
Maybe it will help if I make the model more formal, since there are multiple variables and it can be a little hard to see what's going on if you don't already have the intuitions to track them from normal words. Let m be a measure of self esteem as a result of seeing some evidence e about the world (an observation). The prediction error is the extent to which E:E→M (the expectation relation that goes from evidence to a measure of self esteem) diverges from A:E→M (the relation that calculates the actual update in measure of self esteem). So prediction error looks like when E(e)=m≠A(e), and the larger the difference between E(e) and A(e) the larger the error. A self-esteem set point, s∈M, is a measure of self esteem you are targeting such that if A(e)≠s then you want (set an expectation that E(e′)=s for some as yet unobserved evidence e′) to increase or lower your observed self-esteem A(e) such that it matches s. Warning: This is an off-the-cuff model of the theory I just made up right now, so it's probably non-standard and I'd have to think/read about the formalism more to fully endorse it (it's also a little slopping in a couple places because I'm out of practice). I mean to use it only as a pedagogical tool here. When I suggest we can fix self-esteem, I mean we can work to adjust A(e) and E(e) so that they better match, and work to alter A such that the actual esteem you observe yourself to have also matches the set point s. What that looks like in the case of wanting more esteem than you currently have and repeatedly expecting to have more esteem than you observe yourself having (the case where E(e)<A(e)) and it being true that A(e)<s. The fix in this case is to take actions that cause E(e) to rise to match A(e) and take actions that cause A(e) to rise to match s, and even better if this can be done in concert by making A conditional on |E(e)−A(e)| such that A(e) increases as A(e)−E(e) decreases towards 0.