I still don’t see how you can know that the majority of long covid is misattribution and the like. If (1) is large, and (4) + (5) are both negligibly small, then belief-in-covid will be a better predictor of long covid just because symptomatic covid is a better predictor of long covid than asymptomatic+symptomatic covid is.
Yet, one would still think that Having covid would be more predictive of Long covid than Believing you’ve had covid, since Believing and Long ought to be correlated only through their shared association with Having (common cause rather than mediation). The fact that this is not the case could indicate that people with chronic conditions come to think they Had covid (discussed at the end of the study) or that the measure of Having covid is not that good (see Siebe’s comment), or that it’s psychosomatic (loose usage of the term), or something(s) else.
Or that long covid is mainly caused by symptomatic covid, and Believing is a better predictor of symptomatic covid than antibody tests. Which seems pretty likely.
Why does the post imply that a majority of long covid symptoms are psychosomatic?
Let’s say covid is entirely non-psychosomatic, and that we have the following groups:
(0) People who never had covid, never thought they had covid.
(1) People with asymptomatic covid, who don’t believe they had covid.
(2) People with noticeable covid, no long covid.
(3) People with noticeable covid, including long covid.
(4) People who mistake something short-term (like a cold) for covid.
(5) People who have some serious long-term issues, that they mistake for long covid.
Now we have three variables:
(a) Antibody group = (1) + (2) + (3)
(b) Belief-in-covid = (2) + (3) + (4) + (5)
(c) Claimed long covid = (3) + (5)
If group (4) is relatively small and/or group (1) is relatively large and/or group (5) is relatively large, then it makes sense that (b) is a way better predictor for (c) than (a) is.
The french study found that (a) isn’t a good predictor for (c) if you control for (b). I don’t have a good enough intuition for regression with multiple variables to know whether this is unsurprising given the previous paragraph; but my guess is that this is unsurprising given the previous paragraph.
A French study found that long covid is barely associated with having had covid according to an antibody test, yet associated with believing one has had covid (which itself is unrelated to the antibody test results).
Sorry, are you claiming that “belief in covid” is uncorrelated with antibody test results? I think:
That’s a wild claim (people are seriously no better than random at determining if they’ve had covid?).
As far as I can tell, the study doesn’t claim that.
At a glance, the most important result of the study seems to be that antibodies are correlated with long covid as long as you don’t control for belief-in-covid; but antibodies are not associated with long covid once you do control for belief-in-covid. That makes a lot of sense if antibodies and belief-in-covid is very correlated with each other, but is harder to explain otherwise.
Relatedly, your story for why the french study doesn’t imply that covid is psychosomatic is way more extreme than it needs to be. Claiming that you have long covid is totally associated with having covid, as long as you don’t control for belief-in-covid. See e.g. this post.
Edit: See also the comment I wrote here.
I’m suprised about the claim that rapid tests are a good indicator of when you stop being infectious.
When a housemate of mine got covid, he was still testing positive 10 days after his first symptoms (using UK rapid tests), and I assumed this was normal and didn’t indicate infectiousness.
I also note that the US’s current rules for flying into the country is to either have a negative test or to have recently had a covid infection that you’ve recovered from. I assumed this latter rule was because some people keep testing positive past infectiousness. But the US does allow rapid tests when flying in.
that one study that showed babies born during the pandemic had lost 2 standard deviations of cognitive development compared to babies born earlier
Whoa, that effect size is huge. Too big for me to believe it without more evidence. Seems more likely to be a confounding factor. The discussion section of that paper is pretty good, listing a bunch of hypothesis of what the reason could be, but not finding any obviously good ones.
One thing that stands out to me:
One aspect also not investigated here is the impact of mask-wearing by the study staff during childvisits and assessments . The inability of infants to see full facial expressions may have eliminatednon-verbal cues, muffled instructions, or otherwise altered the understanding of the test questions and instructions.
This seems like it could be a big confounder. (Though it only makes sense if it has a differentially larger effect on younger children, since the cognitive loss supposedly applies to babies born during the pandemic rather than babies tested during the pandemic.)
When I count microcovids for airplanes, I assume people are silent. That makes a big difference.
My interpretation of Zvi’s point wasn’t that your model should account for past lack of nuclear war, but that it should be sensitive to future lack of nuclear war. I.e., if you try to figure out the probability that nuclear war happens at least once over (e.g.) the next century, then if it doesn’t happen in the next 50 years, you should assign lower probability to it happening in the 50 years after that. I wrote someone a slack message about this exact issue a couple of months ago; I’ll copy it here in case that’s helpful:
So here’s a tricky thing with your probability extrapolation: On a randomly chosen year, actors should give lower probabilities to p(nuclear war in Nyears) than the naive 1-[1-p(nuclear war next year)]^Nyears.
The reason for this is that the absence of nuclear war on any given year is positively correlated with absence of nuclear way on any other given year. This positive correlation yields an increased probability that nuclear war will never happen in the given time period.
One way to recognise this: Say that someone assigns a 50% chance to the annual risk being exactly 0.2, and 50% chance to the annual risk being exactly 0.01. Then their best-guess for the next year is going to be 0.105. If this was the actual annual risk, then the probability of nuclear war over a decade would be 1-(1-0.105)^10 ~= 0.67. But their actual best guess for nuclear war next decade is going to be 0.5*(1-[1-0.2]^10)+0.5*(1-[1-0.01])^10 ~= 0.45
I think one useful framing of this is that, each year that a person sees that nuclear war didn’t happen, they’ll update towards a lower annual risk. So towards the end of the period, this person will have mostly updated away from the chance that the annual risk was 0.2, and they’ll think that the 0.01 estimate is more likely.
This whole phenomena matter a lot more if the risks you’re dealing with are large, than if they’re small. Take the perspective in the most recent paragraph: If the risk is small each year, then each year without nuclear apocalypse won’t update you very much. Without updates, using constant annual probabilities is more reasonable.
To be concrete, if we lived in the year 1950, then I think it’d be reasonable to assign really high probability to nuclear war in the next few decades, but then assume that — if we survive the next few decades — that must be because the risk is low. So the risk over the 200 years isn’t that much higher than the risk over the next few decades.
In the year 2021, we’ve already seen a lot of years without nukes, so we already have good reason to believe that nukes are rare. So we won’t update a lot on seeing a few extra decades without nukes. So extrapolating annual risks over the next few decades seems fine. Extrapolating it all the way to 2100 is a little bit shakier, though. Maybe I’d guess there’d be like 2-10 percentage points difference, depending on how you did it.
I should also mention CLTR@EAF.
I want to note that the preferred acronym is CLR. This wouldn’t matter except that there’s now another EA-adjacent organisation (the center on long-term resilience) who do use CLTR as their acronym.
It’s very easy to construct probability distributions that have earlier timelines, that look more intuitively unconfident, and that have higher entropy than the bio-anchors forecast. You can just take some of the probability mass from the peak around 2050 and redistribute it among earlier years, especially years that are very close to the present, where bioanchors are reasonably confident that AGI is unlikely.
+1. I will also venture a guess that:
OpenPhil: Well, search by evolutionary biology is more costly than training by gradient descent, so in hindsight, it was an overestimate. Are you claiming this was predictable in foresight instead of hindsight?
is a strawman. I expect that the 2006 equivalent of OpenPhil would have recognised the evolutionary anchor as a soft upper bound. And I expect current OpenPhil to perfectly well understand the reasons for why this was predictable in foresight.
Sam Altman explicitly contradicted that in a later q&a, when someone asked him about that quote.
A factor 1.5-3 of that could be immune erosion, in which case the R0 would be more like 10-20. And more importantly, I don’t know anything that contradicts Zvi’s intuition that this little data shouldn’t push us far away from our priors.
I think this final graph is a bit confused here, unless ‘the original strain’ here means Delta. Delta had about a 120% advantage over ‘the original strain’ or 70% over Alpha. I’m going to take this to mean 500% as compared to that 120%, so 600% of original versus 220% of original, or about a 170% additional increase. Which is… better, but still quite a lot.
I’m pretty sure the graph is trying to say 500% over delta. If you run the numbers on a very similar graph, you get that the new disease is slightly above 5.5 times as infectious as delta, which you could round to a 500% increase. (Assuming identical generation length. Also, I haven’t looked into sources for the numbers.)
South Africa’s vaccination rate is sufficiently low, and this rate of spread so high, that it wouldn’t much matter if there was vaccine escape properties, although it would presumably matter if there was escape from natural immunity.
I don’t know why you expect a large difference there. I’d guess that roughly equal numbers of people in south africa has been vaccinated as has natural immunity.
Or, okay, I can see two reasons to expect natural immunity to be a bigger deal:
natural immunity is more common in the most exposed part of the population.
there’s more uncertainty about the amount of natural immunity in south africa, and if south africa has very sizeable natural immunity and there’s substantial immunity erosion, that would be a pretty good explainer for why omicron would be spreading so much faster than delta. So there’s some bayesian update towards south africa having sizeable natural immunity.
Don’t know where this nets out.
To be clear: Do you remember Sam Altman saying that “they’re simply training a GPT-3-variant for significantly longer”, or is that an inference from ~”it will use a lot more compute” and ~”it will not be much bigger”?
Because if you remember him saying that, then that contradicts my memory (and, uh, the notes that people took that I remember reading), and I’m confused.
While if it’s an inference: sure, that’s a non-crazy guess, and I take your point that smaller models are easier to deploy. I just want it to be flagged as a claimed deduction, not as a remembered statement.
(And I maintain my impression that something more is going on; especially since I remember Sam generally talking about how models might use more test-time compute in the future, and be able to think for longer on harder questions.)
Can you say more about what this picture depicts and why it’s relevant?
Quick googling, numbers for south africa:
Population 59 million.
28% of population at least one dose, 24% fully vaccinated.
89,771 deaths. At 0.5% fatality rate that would be 18 million people ~= 30% of the population.
So maybe 100% immune escape would be a factor 1.5-3? Leaving at least a factor 2-3 for generally increased infectiousness. (Assuming unchanged generation length.)
Oh, come on. That is straight-up not how simple continuous toy models of RSI work. Between a neutron multiplication factor of 0.999 and 1.001 there is a very huge gap in output behavior.
Nitpick: I think that particular analogy isn’t great.
For nuclear stuff, we have two state variables: amount of fissile material and current number of neutrons flying around. The amount of fissile material determines the “neutron multiplication factor”, but it is the number of neutrons that goes crazy, not fissile material. And the current number of neurons doesn’t matter for whether the pile will eventually go crazy or not.
But in the simplest toy models of RSI, we just have one variable: intelligence. We can’t change the “intelligence multiplication factor”, there’s just intelligence figuring out how to build more intelligence.
Maybe exothermic chemical reactions, like fire, is a better analogy. Either you have enough heat to create a self-sustaining reaction, or you don’t.
This is my take: if I had been very epistemically self-aware, and carefully distinguished my own impression/models and my all-things considered beliefs, before I started reading, then this would’ve updated my models towards Eliezer (because hey, I heard new not-entirely-uncompelling arguments) but my all-things considered beliefs away from Eliezer (because I would have expected it to be even more convincing).
I’m not that surprised by the survey results. Most people don’t obey conservation of expected evidence, because they don’t take into account arguments they haven’t heard / don’t think carefully enough about how deferring to others works. People will predictably update toward a thesis after reading a book that argues for it, not have a 50⁄50 chance of updating positively or negatively on it.
You’re right, that makes it ~5.5.