Posts
Comments
I hate that I actually liked Answer to Job
Thus, I must currently hold Sam Altman guilty
*innocent
The largest models should be expected to compress less than smaller ones though, right?
The main problem with crawlers is that their usage patterns don't match those of regular users, and most optimization effort is focused on the usage patterns of real users, so bots sometimes wind up using the site in ways that consume orders of magnitude more compute per request than a regular user would.
And Twitter has recently destroyed his API, I think? Which perhaps has the effect of de-optimizing the usage patterns of bots.
Hinton says he partly regrets his life’s work.
This may be another Cade Metz moment. (38:20)
Right. From what I've seen, the people that support censoring misinformation are almost never doing it out of worries that themselves will get misinformed.
I'm assuming dsj's hypothetical scenario is not one where GPT-6 was prompted to simulate an actor playing a villain.
It's a nice analogy, but it all rests on whether infinite evidence is a thing or not, and there aren't arguments one way or the other here. (Sure, infinite evidence would mean "whatever log odds you come up with, this is even stronger", but that doesn't rule out it is a thing).
Like, how much evidence for the hypothesis "I'll perceive the die to come up a 4" does the event "Ok, die was thrown and I am perceiving it to be a 3" provide? Or how much evidence do I have of being conscious right now when I am feeling like something? I think any answer different from infinity is just playing a word game.
Aiming for convergence on truth. I guess it's true this might lead to a failure mode where one seeks for convergence more than anything else. But taken literally, this should not discourage exploring new wild hypotheses. If you are both equally wrong, by growing your uncertainty you get nearer to converging on truth.
True. Still, using 1960's prices with current production assumes a 1960 flat demand curve, right? It's like using off-season avocado prices when no one buys them to compute real GDP during avocado season.
Maybe the UK's case curve has flattened after the end of the spike due to the asymptomatic people that are getting tested for whatever reasons and turn positive for the reason you state? It doesn't feel likely (perhaps it's just the other omicron subvariant giving it a push? or just the "control system" of people relaxing?). The hospital admissions continued to go down as one would expect if this was the case, though the data at ourworldindata is a few days behind.
I know it's unlikely, but if it was indeed omicron, its faster generation time also would make its numbers drop faster if they managed to move R under 1
I presume 12 feet is a quarter of the risk of 6 feet [...] there is no magic number
My intuitive oversimplified model of this has been analogous to the direct sound vs reverberant sound in acoustics (in slow motion).
I'd expect the risk from direct viruses to follow the inverse square law (at least to the extent that the risk is linear to the expected number of viruses around you, which can't be true for high risks). And maybe be even be reduced by cloth masks which stop big droplets (?).
But the reverberant viruses are supposed to be the main drivers of the pandemic, right? And those don't care about distance for small enough rooms where virosols (heh) have more than enough time to travel everywhere before falling down. This is where N95s and ventilation become crucial, but distancing not so much.
In this model, there is a special distance, a "critical distance" (which depends on the context, masking, etc), after which the direct viruses are as important as the virosols and extra distancing starts not mattering.
Is my intuitive model nonsense?
Can immune escape by itself explain the transmission advantage or do we also need it to be spreading better?
Makes sense to me...
On the other hand, if it takes longer to show symptoms but it's still equally transmissible since early but for longer, you get higher Rs without surprising new mechanisms of transmission. Also, it may also be escaping our current precautions instead of the immunity.
Why does it follow that a longer time to develop symptoms suggests immune escape?
Also, if the timeline is longer, then the estimates of how much more transmissible Omicron is, based on the time it's taken for it to displace Delta, should be even greater, right?
On the other hand, those lockdowns may only last until the cases start going down again, but you can't get unvaccinated.
If not mandating vaccination for indoor dining, then what?
Even that minimally coercive approach you describe is pretty coercive; I don´t expect the benefits to outweigh the ugly side of making many tens of millions of people be injected with something they don´t like or trust or want. Some people are still getting convinced to get vaccinated just with time alone, and many other things could be done better to convince more people without more restrictions. I don´t know what to expand on without making this too long.
Thanks!!
the WHO who still refuse to admit Covid is airborne
Sort of. For some months now, the WHO states that it can spread "in poorly ventilated and/or crowded indoor settings [...] because aerosols remain suspended in the air"
EDIT: (used to ask why the link wasn't formatting properly)
And if the choice is between ‘no indoor dining (or other X) for anyone’ and ‘no indoor dining (or other X) for the unvaccinated’ I know which one I’m choosing, and which one leaves me more free.
I agree that "no indoor dining for anyone" is worse than mandating vaccination for indoor dining. But I also don´t think the situation merits either. Protecting the immunocompromised and people that want but can´t get vaccinated doesn´t make up for the concerns.
we’d not only not make them mandatory, they’d be forbidden.
The space between those two is very small, maybe even negative.
Regarding the intelligence tests after COVID: Fourth, I can imagine some people that had COVID and go test themselves might actually want/expect to see some effect and end up not doing their best, to be a victim or have an excuse or something to blame for whatever.
It's unlikely that these corporations would make the assumption that all future IP would also be "confiscated"
Do you have a good explanation to Moderna's market price drop?
Even if they made that assumption, what are they supposed to do? Stop investing in future developments, and slowly go out of business?
Borrow less, invest less, or, as you say in your last line, focus on other ways of making money that don't require innovation and IP?
Right! My untrained intuition still resists a bit; I should play with the numbers.
Niice, it makes sense! Thanks!
So to recap, I was right in that riskier assets can have higher avg returns, but I was missing the usually bigger and opposing effect where as the assets gets riskier, the same avg returns rely more and more on lucky very big gains while doing worse more often (at least if they are sort of lognormal).
My second point I still think was correct, right? -- i.e., that if Scott believed ETH had some chance of total collapse (a mixture distribution), then this skews it to the other side and pushes the median below the mean, and gives some reason to think ETH is more likely to outperform BTC. Does this make sense?
If ETH is less risky than BTC then the median performance of ETH will outperform BTC and his probability could be consistent with EMH
Wait. Does this mean that EMH expects less risky investments to have higher performance on average? That sounds shocking enough that I must be confusing something here. Or is this some sort of median vs mean distinction that I'm not seeing
?
About 17 and the EMH. Can't Scott be just thinking that ETH is sufficiently more risky than BTC so it may have higher expected returns even with the EMH (the EMH allows this, right?). Or even that he might think ETH has some chance of total collapse (like an outlier at 0) so even with equal expected returns it's much more probable that ETH outperforms BTC than the other way around (?)
What's this supposed to be estimating or predicting with Bayes here? The thing you'll end up doing? Something like this?:
Each of the 3 processes has a general prior about how often they "win" (that add up to 100%, or maybe the basal ganglia normalizes them). And a bayes factor, given the specific "sensory" inputs related to their specific process, while remaining agnostic about the options of the other process. For example, the reinforcer would be thinking: "I get my way 30% of the time. Also, this level of desire to play the game is 2 times more frequent when I end up getting my way than when I don't (regardless of which of the other 2 won, let's assume, or I don't know how to keep this modular). Similarly, the first process would be looking at the level of laziness, and the last one at the strength of the arguments or sth.
Then, the basal ganglia does bayes to update the priors given the 3 pieces of evidence, and gets to a posterior probability distribution among the 3 options.
And finally you'll end up doing what was estimated because, well, the brain does what minimizes the prediction error. Is this the weird sense in which the info is mixed with bayes and this is all bayesian stuff?
I must be missing something. If this interpretation was correct, e.g., what would increasing the dopamine e.g. in the frontal cortex be doing? Increasing the "unnormalized" prior for such process? (like, it falsely thinks it wins more often than it does, regardless of the evidence). Falsely bias the bayes factor? (like, it thinks it almost never happens that it feels this convinced of what should happen in the cases when it doesn't end up winning.)
Whatever prevents the most infection, hospitalization and death is the right answer either way
I first read this sentence as suggesting that killing people is the best way to prevent infection.
Yeah, if R0 is held constant and also COVID-UK is going up in absolute numbers.
Israel's deaths are dropping more slowly than I would have intuitively expected given the vaccinations; I now wonder if it's because of longer duration of the new strains which means we may have to wait a little longer until most of the previous infections resolve. Anyone that's been looking at detailed data (like strain prevalence, the ages of the people still dying, etc) has an opinion? (I just looked at the daily death and vaccination rate)
I haven't read the papers so, please correct me if I guess wrong (most likely), anybody.
I'm guessing the UK strain was estimated from relative growth between strains when the UK cases were skyrocketing, and that gave around ~40% higher R0 than COVID-classic.
Now, say they were underestimating the duration of the UK strain. That would mean it is actually more transmissible than estimated -- but it was masked by the long timescales (transmissibleness means R, right?). And that would mean that it's that much harder to contain than we thought (yet it was contained in the UK, which is great and suggests I'm talking BS). And it also means that it comes to dominate COVID-classic that much faster when COVID is going down.
> This means that we should expect the English strain to arrive in numbers somewhat slower than its level of infectiousness would otherwise indicate.
I'd instead guess that we should expect it to arrive faster since it's would be more infectious than previously expected and the US seems to be mitigating much more decently than the UK at that time? Does this make any sense?
I think you get more points for earlier predictions.
So one should interpret the points as a measure of how useful you've been to the overall predictions in the platform, and not how good you should be expected to be on a specific question, right?
Yeah, I wasn't trying to be tautological.
I am under the impression that you are thinking something like: "Bezos has ~100 billion to spend. If he spends 1 million in X, then he has 1 million less to spend on the rest. But he won't even get to spend it in his lifetime, so that extra million in X doesn't change how much he would spend in Y. Therefore, it's wrong to say that Y will become more available because Bezos spent in X.".
I don't think that's the right way to think about all this. (Warning: oversimplification coming):
Bezos earns some income, say, in a year. Almost all of it will be spent. Most will be invested and not consumed, so it will still increase his net worth, but that demand for stuff is still there, affecting the economy. Bezos is already probably spending about as much as he can, and what he is not spending he is saving which probably means transferring it to someone else who will spend it. So, if he spends USD 10 in X, it's reasonable imho to "expect" the economy to get USD 10 less spending in non-X stuff (on avg)
I think I disagree a bit with both (but what do I know).
For someone like Jeff Bezos, an increase in spending on Item A probably just results in slightly less money spent by his great-grandchildren in 100 years.
This doens't seem to me to be the right way to think about it. Short term, the more he spends on Item A will result in lower spending on Item B, or lower investment in his companies, a lower transfer of money from him to someone else (like through lower savings). Or more money being spent overall if he just uses up cash he had hidden in his pillow; which increases prices for everyone (but this will be made up for in some future).
If we want to make sure that the starving guy gets some of the food, can't we just allocate the food to him directly, rather than having to give him enough money to win a bidding war with Jeff Bezos?
Who produces the food and can set the prices? If it's private companies, then they wouldn't sell it to the state for cheaper than to Bezos, so it would be as expensive to the state as giving that same money to the poor and let them outbid Bezos. If the state owns the stuff, then [insert standard anti-socialism arguments]. If the prices are fixed by the state, then its inefficient and there may not enough production for all. If the prices fixed by the state but depend on the person -- or on how many of X you have bought this month or stuff like that -- then that introduces whole new types of messes.
I doubt that kind of hidden information can affect PredictIt betting odds as it limits the amount each person can bet.
There is bias or Zvi is reaching wrong conclusions with the same info.
It's the placebo effect, obviously; you can't get sick if you zinc it works.
Do they really get higher expected returns from that?
I know they do when the market isn't efficient (relative to the specific investor), but that doesn't help me.
Why is it that riskier investments should give higher expected returns?
I ask not because I don't get that the avg person would rather invest on something safe than something unsafe, all else being equal. I get that. I ask because I imagine that investors could bring their total risk down through diversification without harming the expected returns, so big money would prefer the higher expected returns even if they are risky, and in doing that, they'd bring down the extra returns from the riskier investments.
Is it because investments options are so correlated that diversification isn't enough to bring the risk of a portfolio down to acceptable levels? Or some other reason?
I have the opposite impressions. Science should embrace causality more and do it better. And as a layman term it should be refined so that we stop talking about the causes of any event as a cake where each slice has a name and only one name.
I find it hard to summarize why, at least right now, but my view is sorta similar to Pearl's (though I don't totally like how he puts it). Hopefully later I'll re-read this more attentively and comment something more productive (if no one has done a strictly better job already).
I meant hiding just the CWish posts. There're enough non-CWish posts to attract people that value the way of thinking in general.
Also, it doesn't sound that bad to attract users through 1 to 1 recommendations only. Or allow unlogged people to read all, but only little by little release the power for new users to interact with the content. Maybe release it all at once if a high karma user vouches for you (they lose it that person gets banned or something). Maybe instead of karma, there could be another value that better reflects how much you are likely to value proper manners and thinking (e.g., it could be obtained by summing karma from different topics i
in a way that overvalues breadth of interest ).
I'm just thinking out loud in real time. My main point one can go a long way just by limiting the rate at which new users can invade and screw with the content.
I was 80% kidding. I do believe that the type of people that could attack this community are hugely people that can't tolerate trying to read and understand the kind of content in here; let alone Scott's 999999 word analytical yet clear essays. They didn't sign up for real thinking and nuance when they went into activism.
And unlike others, I don't think mobs are organized. They look like it, but its some sort of emergent behaviour that can be managed by making it boring for the average mob member to attack.
Fair enough.
I still don't get it. agc asked how is the retaliation NOT at attempt to stifle criticism. TurnTrout answered that it is not: it's retaliation for a doxing attack, not for criticism. Then wolflow said something that's "literally" wrong, and metaphorically I didn't get it; probably TurnTrout didn't get it too so he answered the literal interpretation. Etc etc.
But the upvotes-downvotes show I'm not seeing something here.
Of course, that was a given. I just assumed that most of us don't need days of exclusive focus to write an email.
Simply requiring log-in to read some posts, and limiting the rate of new users (maybe even make it invite only most of the times, like a private torrent tracker), should go a long way to prevent mob attacks.
Make a captcha with GPT-X rationalist content against real rationalist content. If you can't tell the difference, you are out :P
Also, train GPT-X on content that triggers mobs, and then use it to keep them busy elsewhere :P
Given the news cycle speed, it makes sense to get ready for the likely scenarios.