I am Andrew Hyer, currently living in New Jersey and working in New York (in the finance industry).
aphyer
Karma: 3,739
I can’t help but read this post as something like this:
Current government mandates around children are very harmful to children.
Enforcement of current cultural norms around children is very harmful to children.
???
We need to add on and enforce these three new government mandates around children and these two new cultural norms around children.
There is one section arguing that schools are prisons that children hate and are miserable in. And then there is another section advocating for the schools to crack down harshly on children using their phones in school. I find this somewhat depressing.
Haven’t found anything particularly good, but I’ve probably gone as far as I’ll go. I’ve done some analysis trying to predict how much variance we expect from each turtle so that I know how much to overestimate, and for the non-special turtles I’m predicting:
Abigail: 23.0lb
Bertrand: 19.0lb
Chartreuse: 26.2lb
DonatelloDontanien: 21.1lbEspera: 17.3lb
(Flint is already estimated as a gray turtle as 7.3lb)
Gunther: 30.0lb
(Harold is already estimated as a six-segmented clone as 20.4lb)
Irene: 23.7lb
Jacqueline: 20.0lb
I’m rounding these to 0.1lb even though I’m allowed to go more granular, because if the Tyrant weighs to the same precision we do he will also be rounding to 0.1lb, which means we don’t gain anything from more precision (estimating 7.25lb gives a payoff exactly halfway between estimating 7.3 and 7.2).I’ll put these estimates in the parent comment for ease of GM extraction.
The one interesting thing I’ve turned up is that Abnormalities appear to convey a very large amount of variance: each abnormality adds ~1lb of average weight, but actually slightly over 1lb of stdev-weight. I suspect that abnormalities are adding weight in a highly-random way: my weight estimates for Espera, Irene and Jacqueline (0-abnormality turtles) are relatively low as a result because my confidence was higher, while my estimate for Gunther (6 abnormalities?) has a lot more safety margin built in.
A simple linear regression analysis on the remaining turtles (everything that isn’t a Fanged Gray Turtle or a Six-Segmented Harold Clone) gives the following formula:
10.56lb base weight if green...
+2.02lb if grayish-green,
+5.47lb if greenish-gray,
+0.359lb/Wrinkle
+0.142lb/Scar
+0.598lb/Segment
+1.000lb/Abnormality
This does a reasonable job of prediction, but has a residual with a fairly-large ~2lb standard deviation. Our standard-deviation math suggests that this means we should give the Tyrant answers overestimating each turtle by 2.4-2.5lb, and should expect to lose on average ~35gp/turtle to error.
That seems like we might be able to improve on it, but I’m not sure how. I haven’t been able to find any useful interactions yet. There does seem to be an obvious explanation of all the traits except Abnormalities being driven by some hidden Age variable: old turtles start getting grayish, are wrinklier, have grown more shell segments and accumulated more scars, and are larger. However, I’m not sure how actionable this is for us.
The one thing it does look like I can do is adjust the amount of overestimation I do: it does seem that our estimate is less accurate as turtles get older and larger, and so rather than overestimating by 2.44lb for every turtle I should overestimate the larger ones by more and the smaller by less. That’s not going to be a very large improvement, though. I feel like there ought to be something else to do, but haven’t found anything yet.
What is QC?
The Fanged Gray Turtle seems relatively simple, so we look at that first.
The weight of a Fanged Gray Turtle seems well-approximated by (0.425 + 0.4568*#segments) lb.
This leaves behind a residual that looks roughly like a normal distribution with stdev ~0.357lb. I’m not able to find any interaction of this residual with any other properties of the turtles—scars, mutations, etc. all seem unpredictive for the Fanged Gray Turtle.
Some quick math reveals that the Tyrant’s asymmetric payoff distribution encourages us to overestimate a turtle’s weight by ~1.22 standard deviations. Therefore, we’re going to bump up all our weight estimates by 0.435lb in order to flatter His Tyranny.
(We could bump them up a bit further if we thought that reducing the odds of him having an unflattering portrait of us was worth trading off money for. However, I actually think we can plausibly use that to extract more money: whatever itinerant artist he kidnaps to do that portrait, we can demand that they give us part of their commission in exchange for us being helpful and sitting for the portrait! Kaching!)
There’s only one Fanged Gray Turtle among the Tyrant’s pets: Flint, with 14 Shell Segments. Our best guess of Flint’s true weight is 6.8lb, but we’re going to overestimate this to 7.3lb in order to optimize our payoff.
And two(low-priority) questions for the GM:
Are we unusually careful and competent at weighing turtles in a way that the Tyrant is not likely to be? If he is careless about weighing his turtles, and introduces additional error, that increased variance makes us want to slightly increase how far we overestimate by.
What level of granularity are we able to give the Tyrant in our weight estimates? I think that an estimate of 7.25lb for Flint is slightly higher-payoff than 7.3lb in expectation, but don’t know if that’s something I’m allowed to give.
When we look at the distributions of variables individually, there’s a startling number (5-6k out of 30k) of green turtles with 6 shell segments (the lowest number, never seen otherwise), zero wrinkles, and zero abnormalities, that weigh exactly 20.4lb.
They do have varying numbers of scars, though, which makes me incline more towards ‘some very particular turtle subspecies’ and less towards ‘one very friendly turtle that figured out that it can get extra attention by wiping off the mark you put on it and coming by again’.
Harold from the King’s pets matches this pattern (and thus presumably is one of these strange clone turtles).Removing those and looking at the rest of the universe:
The remaining green turtles now resemble the grayish-green and greenish-gray turtles, making me draw the following three species:
Fanged Gray Turtles.
Six-Segmented Harold Clones.
All Other Turtles.
Most variables are now reasonably smoothly-distributed:
Scars and Wrinkles look Poisson-like.
Abnormalities peak at 0 and fall off: that might also be a poisson distribution, just from a lower mean, or might be something else.
Weights are bimodal (with one peak around 5-6lb for the Fanged Gray Turtles, and one wider peak around 15-25lb for All Other Turtles).
EDITED TO ADD FINAL ANSWER:
Abigail: 23.0lb
Bertrand: 19.0lb
Chartreuse: 26.2lb
DonatelloDontanien: 21.1lbEspera: 17.3lb
Flint: 7.3lb
Gunther: 30.0lb
Harold: 20.4lb
Irene: 23.7lb
Jacqueline: 20.0lb
Getting started with my favorite first step of calculating a bunch of correlations:
It turns out that all fanged turtles are gray, and all gray turtles are fanged.
This suggests some kind of speciation by color.
When we break down by color:
Grayish-green and greenish-gray turtles show near-identical patterns—I assume those are the same species and you’ve just categorized them a couple different way:
Weight is positively correlated with wrinkles, scars, shell segments and abnormalities.
The first three of these are positively correlated with one another, and probably reflect some hidden ‘age’ variable. Number of abnormalities is not correlated with the others, and seems to do its own thing. (Turtles grown larger with age, and also weigh more per extra mutant tentacle they have grown?)
Nostril size has no effect.
Gray turtles work differently:
They show the same pattern of wrinkles, scars and shell segments being positively correlated.
However, weight in this case seems to be almost entirely determined by # of shell segments.
Perhaps these turtles grow at a more predictable rate, with one shell segment per year that adds a regular amount of weight?
And green turtles also work differently:
The correlations between wrinkles, scars and shell segments have broken down.
Additionally, those variables have only small correlations with weight.
The most predictive variable towards weight is the # of abnormalities.
Perhaps these are strange mutant
ninjaturtles of some kind thatare perpetually teenagedon’t have a regular growth lifecycle?
It looks like these three species of turtle behave differently enough that I’m probably going to end up modelling the three of them all separately (except maybe the what-I’m-assuming-is-age effect that shows up on both gray and mixed-color turtles).
My planned next step is to try three independent simple regressions and see how predictive they are for each of those three types of turtle.
The argument Zvi is making.
They open with the Battle of the Board. Altman starts with how he felt rather than any details, and drops this nugget: “And there were definitely times I thought it was going to be one of the worst things to ever happen for AI safety.” If he truly believed that, why did he not go down a different road? If Altman had come out strongly for a transition to Mutari and searching for a new outside CEO, that presumably would have been fine for AI safety. So this then is a confession that he was willing to put that into play to keep power.
I don’t have a great verbalization of why, but want to register that I find this sort of attempted argument kind of horrifying.
I’m surprised to hear you say that. I would consider it perfectly reasonable to use a line graph without a zero-based y-axis to plot gravity against altitude: the underlying reality is in fact a line (well, a curve I guess)! Gravitational force goes down with altitude in a known way! But the effects of altitude on gravity are very small for altitudes we can easily measure, and extending the graph all the way down to zero will make it impossible to see them.
If I measure gravitational force against altitude, and end up with points like the following:
0 ft above sea level, force is 9.8000 m/s2
1000 ft above sea level, force is 9.7992 m/s2
2000 ft above sea level, force is 9.7986 m/s2
3000 ft above sea level, force is 9.7980 m/s2
would it be egregious for me to plot this graph without a zero-based y-axis? Do I need to plot it with a y-axis going down to zero?
Certainly there are cases where it’s misleading to not extend a graph like this down to zero. But there are also cases where it’s entirely reasonable to not extend it down to zero.
I did not understand #8 at all. I am confident that this is not because I don’t understand the general point. Does anyone have an explanation of #8?
<trolling>
The S&P500 has returned an average of ~8%/year for the past 30 years. As you say, we have on many occasions observed people lying, cheating, and scamming. But we have only rarely observed lucrative good ideas! Why, even banks, which claim much more safety and offer much lower returns than the stock market, have frequently gone bust!It follows inevitably, therefore, that there is a very high chance that the S&P 500, and the stock market in general, is a scam, and will steal all your money.
It follows further that the only safe investment approach is to put all your money into something that you retain personal custody of. Like gold bars buried in your backyard! Or Bitcoin!
</trolling>
‘Bike’ is sometimes used as shorthand for ‘motorcycle’, in which case the ‘absurdly dangerous’ claim stands. I agree that riding a pedal-powered cycle without a helmet is somewhat dangerous, and unnecessarily so, but not ‘absurdly dangerous’.
On the object-level of your particular case, I don’t see how you’ve ended up rate-limited. The post of yours that I think you’re talking about is currently at +214 karma, which makes it quite strange that your related comments are being rate-limited—I don’t understand how that algorithm works but I think that seems very odd. Is it counting downvotes but not upvotes, so that +300 and −100 works out to rate-limiting? That would be bizarre.
In the general case, however, I’m very much on board with rate-limiting people who are heavily net downvoted, and I think that referring to this as ‘censorship’ is misleading. When I block a spam caller, or decide not to invite someone who constantly starts loud angry political arguments to a dinner party, it seems very strange to say that I am ‘censoring’ them. I agree that this can lead to feedback loops that punish unpopular opinions, but that seems like a smaller cost than communities having to listen to every spammer/jerk who wants to rant at them.
The ‘full repayment’ part is only sort of true, in a similar way to what happened with MTGOX, due to bankruptcy claims being USD-denominated.
Suppose that:
You owe customers 100 Bitcoin and $1M.
You have only half of that, 50 Bitcoins and $500k.
The current price of Bitcoin is $20k.
You are clearly insolvent. You will enter bankruptcy, and the bankruptcy estate will say ‘you have $3M in liabilities’, since you owe $1M in cash and $2M in bitcoin.
Suppose that the price of bitcoin then recovers to $50k. You now have $3M in assets, since you have $500k in cash and $2.5M in bitcoin! You can ‘fully repay’ everyone! Hooray!
Of course, anyone who held a Bitcoin with you is getting back much less than a bitcoin in value, but since the bankruptcy court is evaluating your claims as USD liabilities you don’t need to care about that.
This ‘full repayment’ is plausibly still important from a legal or a PR perspective, but e.g. this part:
there is typically a legal fight over whether a company was insolvent at the time of the investment or that the investment led to insolvency. If every FTX creditor stands to get 100 cents on the dollar, the clawback cases that don’t involve fraud wouldn’t serve much of a financial purpose and may be more difficult to argue, some lawyers say
is better thought of as ‘our legal system may get confused by exchange rates and pretend FTX was always solvent’ rather than as ‘FTX was actually always solvent’.
Competition should improve meth and reading outcomes here.
Is this a typo, or a snarky comment on reducing student drug use?
We don’t care about how many FLOPs something has. We care about how fast it can actually solve things.
As far as I know, in every case where we’ve successfully gotten AI to do a task at all, AI has done that task far far faster than humans. When we had computers that could do arithmetic but nothing else, they were still much faster at arithmetic than humans. Whatever your view on the quality of recent AI-generated text or art, it’s clear that AI is producing it much much faster than human writers or artists can produce text/art.
This change would not get rid of the need for researchers to have a non-research skillset to secure funding. It would just switch the required non-research skillset from ‘wrangling money out of grant committees’ to ‘wrangling positions out of administrators’. Your mileage may vary as to which of those two is less dysfunctional.
Will that extra credit be worth...uh...at least 1.98 gp?