I'm not going to knock someone who actually did their homework and measured something and reported the results dispassionately. Even if it was on a smallish and not-too-complex problem. Too many people just won't do that, will make wild assumptions, and will just fiddle with things until they sort of work.

And the way you learn how to do large optimization problems is by practicing on smaller ones.

If you don't adequately specify the problem, you can't blame people for taking shortcuts within the inadequate or unspecified problem.

Also, it is kind of hard to design and write good tests without a specification.

Using a simple "find . -type f | wc -l" tells you a lot.

On a Raspberry Pi running some embedded stuff and a lot of python libraries:

272706

On a MacBook Pro with a home directory transported from machine to machine since about 2002:

760360 (in my home directory)
4611064 (all files)

You can easily manage that if you worked in the Asia-Pacific region and were flying around a lot over the course of a month. I did a three week press tour between Brisbane and Tokyo (with stops at Singapore, Saigon, Hong Kong, Taipei, Shanghai, and Seoul) and ran up over $45k in airfare and hotel charges.

I've seen far, far worse.

One company laid people off while they were engaged in business travel, and shut off their company Amex cards the very same day. So they had to get home at their own expense and were often stuck with hotel bills.

Another company for just refused to pay the Amex bills of laid-off employees. Since we had tech support people flying all over the country there were a few dozen laid-off folks stuck with several thousand (I recall one case that was over twenty thousand) dollars in charges. When you get that card you are effectively cosigning for it so they were all screwed.

Of course, if you include all of the preexisting conditions you end up with about half the population, which is a lot of people to write off.

Also, focusing on the fatality rate completely ignores what we've learned about the long-term health impacts of infection. And that seems to hit a lot more people including many younger people without preexisting conditions. Depending on how you define "long covid" some long-term effects linger in somewhere between 20% and 50% of infections.

Forget the planet. How would the plane (and pilot) not be ripped apart by tidal forces?

You need to look at Table 2, linked to in the study:

https://www.cdc.gov/mmwr/volum...

That "five times more likely" is for one scenario and one question.

If you got the Moderna vaccine an unvaccinated person is seven times more likely to be reinfected and hospitalized than you.

If you are considering when the Delta variant was dominant an unvaccinated, previously infected person is over seven times more likely to be reinfected and hospitalized than a fully vaccinated (with the mRNA vaccines).

Please follow the links.

My first thought was, who in their right mind in private industry is going to share datasets? Training datasets are the most valuable piece of intellectual property in this new world and most companies in this space spend huge sums of money curating those datasets. The exact contents of those datasets and (most importantly) the processes used to curate them are tightly held secrets.

I think a better approach would be to fund the development (probably via the existing NSF process) of relatively generic training datasets (e.g. like ImageNet and PASCAL) that are useful for comparing different models and techniques. And hopefully without some of their flaws and quirks.

On the other side of it, you don't need cloud services to run or train AIs. At this point in time a lot of the biggest leverage, in my opinion, is going to be from fairly lightweight AI implementations that you can put in low-power and low-cost edge computing devices. You can train those models on an inexpensive Linux box with a few RTX cards (I train 'em in about an hour with dual RTX cards and an older Ubuntu box). So there are a couple of things you could do there: one is to fund the open-source development of standardized training tools and evaluation tools; another is to subsidize the purchase of such hardware (already not very expensive) by researchers, even a very modest subsidy of around $1000 would likely go a long way.

The final thing you could do is break the damned NVIDIA monopoly.

Also, please have specific research objectives that you are going to prioritize before you start writing checks.

Note that none (except possibly fixing the NVIDIA monopoly) of these things would cost very much money and all of them would speed up AI development quite a bit.

An important point is that the population of people, either vaccinated or unvaccinated, that got sick enough to get hospitalized or even showed symptoms was tiny.

Out of around 32000 people who were infected, 238 were previously vaccinated and 19 were unvaccinated but had previously had COVID-19. And 8 of those hospitalized were vaccinated and 1 was unvaccinated but previously had COVID-19.

So your risk of getting symptomatic COVID-19 when you are vaccinated is around 1.5%, as opposed to around 0.06% if you previously had COVID-19.

These numbers are a pretty tiny piece of a larger sample and making big extrapolations from them does not seem wise. Even very small random effects or biases in the overall population could completely negate those numbers.

You might have some traction here if you could rent a place in Maui, Crested Butte, Park City, Banff, or Fiji. Otherwise I don't see the play here.

Of course it has.

But as you said, that was well before humans. The last ten thousand years has been a period of remarkable stability with respect to climate, and you'll recall that human civilization has flourished during that time and human populations have dramatically increased.

So it seems we are running an experiment where we see if human civilization can survive during a period of much less stable climate. With the lives of billions of people depending on the results of that experiment. I for one would rather that experiment be ran on some other planet and not the one I live on.

1. Alexnet learns a particularly inefficient representation of the training data. Do more modern models, in particular ones that are fully convolutional (e.g. no fully connected layers), have similar problems?

2. How does this steganography technique work in the face of weight quantization and pruning?

Got evacuated when a fire (started by someone welding irrigation pipe) blew up just south of my house. My house was saved but it was a close-run thing. I've paid $12k per year for homeowner's insurance for over twenty years here, and they will be happy to pay for both the private firefighting crews and cleaning up the mess.

So I don't care whether people think it is caused by climate change or not. Right now I am more focused on adapting and figuring out how to live with the new reality on the ground.

Not only was that an all-time Canadian temperature record, that is an all-time temperature record north of 45 degrees North latitude.