Your quoted cost for training the model is for training such a model **once**. This is not how the researchers do it, they train the models many times with different hyperparameters. I have no idea, however how hyperparameter tuning is done at such scales, but I guarantee that the compute cost is higher than just the cost for training it once.
MichaelLowe
Karma: 242
Any remark on the recent reports of reinfections? In previous pieces you were quite critical of such reports, IIRC you estimated a minimum of 4 months of immunity. However, the US reinfection case seems to have been reinfected after less than two months, and with a much more serious reinfection to boot.
We need more discussion of Slovakia’s approach: Mass testing everybody with cheap antigen tests, testing is semi voluntary, if you don’t have a negative test result you have to quarantine.
I maintain that the Slovak government has shown extraordinary competence here, when compared against an admittedly low international baseline. Even more astonishingly, the Lancet article makes it seem like they went against the advise of their scientific advisors who pushed for a 45 days lockdown.
Before doing nationwide testing, they implemented a successful local test run.
They understand what an incubation period is, and are therefore doing two runs, last weekend and this weekend.
With antigen tests they settled for a good enough solution, understanding that they will catch the vast majority of infectious people who PCR tests would have caught as well.
They managed to mobilize this whole initiative seemingly in a span of 2-3 weeks, with soldiers making sure distances are kept in waiting lines.
What I find most surprising is how discerning the government was in picking out the right advice and ignoring the bad scientific advice (at best governments follow one set of advisors completely, mostly they implement their recommendations in a worse than random way). Why bad? Because most epidemiologists only consider medical harm but not economic, and let the perfect be the enemy of the good:
The bad advice included: Antigen tests are less reliable than PCR tests, that there would be high infection risks at the testing sites, that not enough doctors would be available, that it would be immoral to forcibly quarantine people who do not want to be tested (apparently much better to effectively quarantine everybody in a lockdown), that a lockdown that would crush the economy was the better approach( what comes after the lockdown when the numbers will most likely rise again?)
This is not to say that each individual point was wrong or unlikely to happen, just that the epistemic standard when trying something new should not be “this will definitely work without any problems” but “this is worth a shot”, particularly when the other alternatives have huge downsides as well. Neither am I saying that everything was done perfectly (I am unsure whether it is best to exclude 65+ citizens, citizens could maybe also do the swab testing supervised but themselves)
The jury is still out on whether this will flatten the numbers, but we should all hope that it works out for them so we have a viable alternative.
Regarding the cough identifying AI: Due to the online collection method I suspect that most of the positive samples were already quite advanced in their disease progression. Since Covid-19 deposits in the lungs mainly in the latter part of the disease it is easier to identify them at that point, but also not that useful anymore because most of the transmission happens during the earlier part of the infection (both for symptomatic and asymptomatic people).
These researchers had a much better sample procedure, cough samples were mostly acquired at testing sites, where participants did not know yet whether they have Covid (much less risk of subconscious bias) and were presumably at an earlier stage of their disease. They also had much worse results, which I suspect are more realistic for a real world setting.
What actually needs to be done is to do a longitudinal analysis, i.e. you record your baseline cough when you are healthy. Then if you want to check if you are infected, you cough again and compare that “potentially sick” cough against your baseline “non-covid cough”. The potential of this approach is much higher since baseline characteristics of the cough can be accounted for (smoker, asthmatic, crappy mic in phone).
I have been thinking that it should be possible to gather training data for this quickly by identifying a subset of people that are somewhat likely to get sick in the near future like e.g. people participating in big parties, and acquire coughs from them prior and subsequent to infection. If somebody has ideas how to acquire such data, feel free to share. As an aside, I am somewhat surprised that we as a community interested in AI and out-of-the box thinking have not focused/discussed AI for Covid detection much earlier.
Reposting my comment under Zvi’s post:
Due to the online collection method I suspect that most of the positive samples were already quite advanced in their disease progression. Since Covid-19 deposits in the lungs mainly in the latter part of the disease it is easier to identify them at that point, but also not that useful anymore because most of the transmission happens during the earlier part of the infection (both for symptomatic and asymptomatic people).
These researchers had a much better sample procedure, cough samples were mostly acquired at testing sites, where participants did not know yet whether they have Covid (much less risk of subconscious bias) and were presumably at an earlier stage of their disease. They also had much worse results, which I suspect are more realistic for a real world setting.
What actually needs to be done is to do a longitudinal analysis, i.e. you record your baseline cough when you are healthy. Then if you want to check if you are infected, you cough again and compare that “potentially sick” cough against your baseline “non-covid cough”. The potential of this approach is much higher since baseline characteristics of the cough can be accounted for (smoker, asthmatic, crappy mic in phone).
I have been thinking that it should be possible to gather training data for this quickly by identifying a subset of people that are somewhat likely to get sick in the near future like e.g. people participating in big parties, and acquire coughs from them prior and subsequent to infection. If somebody has ideas how to collect such data quickly, feel free to share.
The Cuomo video does not have the quote “stop the distribution of the vaccine”, the clip says that Cuomo wants to shape or stop Trump’s *plan*. This could mean that an alternative plan would be implemented that would fulfill the Cuomo’s requirements.
I feel that getting the quote right is necessary if one is literally calling for pitchforks.
Does anybody have recommended resources that explain the timeline of clinical trials of interventions? Specifically why they take so long and whether that is because of practical necessity or regulatory burden. Bonus points if Covid-19 is included as a context.
This is a very important topic and question, but I fear that you generalize too much and your assessment of Western politicians’ understanding lacks subtlety. In particular, my opinion is that the obviously good strategies were just not politically feasible. In the beginning of the pandemic, I used to treat arguments of the form “The successful strategy of country A is just not possible in country B” as defeatism and status-quo bias, but I now believe that the South Korean model is indeed not possible in Western democratic countries. This can be seen by creative and smart initiatives of some Western countries that nevertheless failed.
You mention that the government holds the following misconception:
It’s fine to hover just below the point where hospitals get overwhelmed—it’s not important to bring down the number of active cases as low as possible
However, the German government is perfectly aware of the meaning of exponential spread, here is chancellor Angela Merkel explaining what R means and why a value of 1.1 would be too high.
While hand-washing was an important recommendation in the beginning here as well, our public health messaging has been focused for some time now on droplet and aerosol transmission. School and university classrooms are often required to be ventilated at regular intervals (which for most schools is not doable, but that’s a different topic). Hand sanitizer is also much easier to implement than any ventilation measure in Winter.
You also invoke the risk society thesis, but this would apply to Asian countries as well, which were able to contain the virus.
In addition, I think “the summer success in Western countries was not due to measures but due too weather effects” is far too strong a claim. European countries had a decent contact tracing system and cancelled mass events, while the US did not have the first part and had far worse numbers in summer.
Why the South Korean model would not work in the West:
South Korea did contact tracing very well, with huge invasions of privacy like checking CCTV data, publishing the whereabouts of infected individuals and using credit card transaction history. In the US and the UK contact tracers are happy if contacted individuals pick up their phone at all. It’s paradoxical, but it seems to me that Western populations would rather accept a wrecked economy, restriction of movement AND hundred thousands of deaths than a temporary surveillance program.
Examples of Non-Asian countries with smart but failed initiatives:
As far as I can tell, there has only been one Western country to try to eradicate the virus, namely Israel which implemented very tight border control policies and a mobile phone surveillance initiative very early. However, my impression is that cooperation of the populace is just not high enough, which is why a second lockdown had to be imposed.
A to me pretty saddening case is the initiative of the Slovak government to test its entire working age population through cheap antigen tests. Testing was semi voluntary, with the other option being mandatory quarantine. New infection numbers fell very rapidly, but because the testing was done in parallel with a partial lockdown it’s not exactly easy to determine causality. However, since many other European countries with similar lockdowns have at best a flattened curve it seems very likely that mass testing was a great idea which is why it’s copied now in parts of England, Austria and Italy. Despite the large success and subsequent reopening, another round of mass testing has in Slovakia been postponed indefinitely, mostly because the mandatory quarantine got many voters angry and popularity of the government has been waning rapidly.
So in conclusion, many smart policies are much harder to implement in Western countries and may actually reflect the preference of the population, and that our current situation is not because of governments “[...] making some silly errors, not updating their information, and not thinking through the long-term effects. ”
However, there was/is room for fairly cheap wins through scientific and regulatory adaptation. This post is already too long, but briefly put the failure seems to be in those two areas. Despite strong theoretical justifications, no country (AFAIK) has so far approved at home, cheap antigen testing.
I have also been experimenting with mind palaces for books this year, and have a slightly different experience. This is not meant to contradict your experience, just pointing out that it can vary for different people:
1)”Constructing a memory palace makes textbook reading more engaging and focused.”
While I am indeed more focused while memorizing boring content, I would not necessarily say that it is more engaging, in part because I have to mentally switch between reading/listening and visualizing. Furthermore, despite long experience, visualizing does not feel effortless for me, so unlike you I cannot study a textbook this way longer than when I am just studying it.
3) A memory palace helps you relate and understand concepts.
True, and a very important point. My biggest gripe with Anki/Flashcards is how split up the knowledge becomes after atomizing the pieces of information.
4) Visualization practice and memory palaces make you more able to see and manipulate 3D structures.
My loci in the memory palace are fairly generic, in that they would work just as well as 2D pictures, which is maybe why I do not feel like I got better at manipulating 3D structures. Certainly, I have never used a complex 3D structure in my mind palaces, it would just take too long in the first place.
I also think that visualization/mind palaces would be tremendously worthwhile in school or/and work. They are great for deeper mastery of concepts, because they allow for longer recall AND building links between topics, but also for stupid grinding of facts before an exam.
Very well written post, although I disagree with the main conclusions. But first of all, I agree that both in the original and in the updated version of the essay Kelly seems to imply that if an average artist goes for his strategy, they can expect (>= 50%) to be able to live off their work:
I don’t know the actual true number, but I think a dedicated artist could cultivate 1,000 True Fans, and by their direct support using new technology, make an honest living.
Having said that, I do not think your conclusions about the EMH hold here, or that they are somewhat inconsistent with your other statements. First of all you agree with the author regarding the superiority of the “1000 true fans” strategy over the broad appeal strategy.
However, then you assert that according to the EMH it is impossible that one job is much easier than the other. But that would also mean that the “1000 true fans” path cannot be superior to the normal path, so both of your assertions are (somewhat) in conflict with each other.
But the EMH brings it all together in a much bigger way. It says:
Is this author suggesting that it’s much easier to do one job than another and make the same money?
THEN THIS AUTHOR IS WRONG!!!!
Personally, I both believe that the 1000 true fans strategy is indeed more superior for many/most creators and in that sense a 20 dollar bill that is lying on the ground for a long time. If one believes in the EMH here, one would also need to explain why it took >10 years for OnlyFans and Substack to emerge.
Thanks for the explanation. I do not understand the formula however. As I read your explanation, if both strains had the exact same doubling time of 6.5 days, one strain would still be ln(2) *6.5/6.5 = 0.69 more infectious than the other one, so I must be misunderstanding.
[Question] Base rate of RCT from developing countries finding unexpected effects
While I agree that the review is written engagingly, and I personally would like to see more book reviews/summaries, I find this decision surprising. Surely, the most important question in evaluating a history book is whether it is accurate. The author addresses this point, but dismisses criticism with “the historians are upset they were bypassed”, and “it does not seem important to me”. This is neither true, kind or necessary: the authors of the review criticize that despite Heinrich’s claims, the Catholic church did not have the power to bring about the changes in social kinship norms that Heinrich attributes to them. This is, like, the central thesis of the book, so it is certainly not unimportant.
To his credit, the author of this review engages openly with the criticism of Charles Freeman in the comments.
Not very confident, but given that several what appears to be independent mutations have been identified, but with mutations in similar regions and at a similar time, it could make sense that a new or increasingly used therapy form could be behind this.
The rescuers were just random people hindered by bystander effect like everybody else.
You do not seem to go into the possibility that many people who were asked to help refused to do so, in contrast to the rescuers. Since it was a literal matter of life and death, I believe that many or even most Jewish people did try to ask for help but were declined.
But otherwise a great post, and I was happy to see it included in the Curiosity Book.
Thank you for this great post. I would like to comment on a particular part:
“Besides, even in the technical classes, forgetting is the near-universal outcome, and the long-term benefits are mostly conceptual — for if you don’t use these skills continuously for the rest of your life, you’re almost certainly going to lose them. Maybe more than once.”
This seems strongly like throwing the baby out with the bathwater. At least for math, it is, as you say, very clear that skills build on top of each other, and that weak students struggle greatly because they lack the foundational skills. I used to tutor friends and used to be regularly shocked by how much they forgot: how to divide by fractions, summation with exponents, just basic stuff. So the following is pretty likely:
1)If you don’t understand old concepts in math, you will not understand many new concepts
2) many students forget old concepts
3)Spaced Repetition (not necessarily software!) ensures remembering
We should be making sure they remember this material not in 20 years, but in 2 years if and when we make them struggle through calculus (unless one has a realistic plan for how to let the majority of underachievers in math just drop out of advanced math classes). This gives mediocre students a fair chance of actually building deeper understanding of what is going on on a conceptual level, and underachieving students a chance to get acceptable grades so that they do not become dejected and apathetic.
For math, students get assigned homework anyway. I see no reason why that homework should not include an automatically generated exercise aimed to repeat an old concept. Of course, multiple-choice flashcards would be a failure mode.
This is the key point. All other comments are about the base rates of general forms of thrombosis, but the concern is about a rare, special type. It is actually not inconceivable that the risk of death for young people from this vaccine might be higher than the risk from Covid-19.
This is right. For people who do not know, you cannot actually use AMD GPUs for deep learning (at least not productively, AMD is trying to get there though), so AMD’s rise has little to do with AI.
This is an interesting hypothesis, but I find it implausible that there is large temporal variation in vitamin D levels. Seasonal variation which might be even the biggest factor affects everybody the same, and it just does not seem to match my experience that the majority of the population changes their diet in such random ways that they could become Vitamin d deficient by chance. Same with indoor/outdoor activities, most people’s life is not that variable that they are spending each day outside one month, but not the next. Besides, Vitamin D deficiency is correlated very strongly with various commodities, which definitely do not randomly fluctuate.
I would also bet that the secondary household attack rate is similar across different age groups (except children) while it is known that Vitamin D deficiency is much likelier in older people.
Given that Steven Pinker retweeted Scott’s deletion post and this news article , this issue will probably keep getting publicity for better or worse. Given this, some people will start looking for Scott’s real name, and thus it would be a great idea to increase the entropy here by promoting a value for Scott’s real name that is not ahem entirely accurate. Thoughts?