SpaceX made $75 billion actual real dollars. It's in the bank.

Sure, the individual VCs aren't allowed to take their actual cash out of the company until August 6. Want to bet the datacenter hype keeps going until at least then?

Elon Musk, or whoever manages him, already learned not to post speculative tweets about his companies followed up shortly later by "just jokes lol".

It's not great, but I don't think that's the least practical part of it. Reasonable people have done the math and you can almost make it work just by making the radiators the same size as, and putting them on the back of, the solar panels. Starlink satellites already generate and dissipate a kilowatt plus.

The impractical part is that the whole thing is going to deorbit and burn up after five years. Sure, maybe you don't want the five year old GPUs, but replacing the panels and radiators every five years is going to be more expensive than building twice as much on the ground.

Conspiracy theories are silly, and unnecessary.

Remember when Elon ruminated about hyperloop and everyone went nuts? Well, this time he ruminated about space datacentres. Oh, and SpaceX investors cashed out for ~$85 billion.

It is LEO.

Size is not free. Besides having to get the thing up there, which might come down to merely very expensive, there's drag in low Earth orbit, and the bigger the surface area of your satellite the more propulsion you need to keep it up there. The life of Starlink satellites is primarily limited by their propellant.

Even if you ignore launch costs entirely, is it cheaper to put your datacentre in space and replace it and your power plant every few years, or put it in a nice desert or on a floating island somewhere instead? Oh, and you have to engineer it to be completely maintenance free for the first option too.

What's the downside? SpaceX stock got pumped for their IPO. The money is made. As long as the hype keeps going they can raise more any time they want, or Elon could sell some of his shares. If it turns out to be unworkable, SpaceX (and subsidiaries) are back where they started.

There aren't really any unsolved engineering problems. SpaceX can absolutely put a rack of nvidia GPUs into low orbit. We could have done that in the 70s. The argument is whether it's economical or not.

Which is cheaper, putting a thousand square metres of solar panels, a rack of GPUs, a vacuum cooling system and propulsion in low orbit and incinerating and replacing it all every few years, or the panels, GPUs and a convective radiator that is ~50x more efficient on the ground and runs for twenty plus?

If I put some bricks or concrete blocks in a vacant lot and they assembled themselves into a house you'd think that was pretty cool.

I didn't say they didn't. Those missiles were very effective. Much more effective than ten times as many drones would have been.

Both missiles and drones are of limited use when you're not willing to send infantry to take and hold territory. Cheap light drones much more so than Tomahawks and GBU-57s.

The US didn't fail in the Iran war due to a lack of offensive air power, and its offensive air power definitely wouldn't have been improved by packing the Bush, Lincoln and the Burkes with cheap drones. It failed because it was a half assed effort organized by clowns who apparently don't know any military history and decdied to disregard all their advisors who do.

You can theoretically win a war with air strikes alone but it depends entirely on your enemy. If they fail to surrender then you lose, and there's a strong incentive to dig in when some foreign invader starts dropping bombs on you. That's why no real war has ever been won that way.

Using projected military power in general has always been hazardous. It's expensive, unpopular, and depends on either a quick victory or local allies, preferably both, or it will fail. "Local allies" are part of the soft power that Hegseth is so disdainful of, and the US has lately been lighting on fire just to watch the pretty flames.

Datacentres are coveted by lots of local governments. Lots of construction jobs, high paying jobs in operation and maintenance, no on-site pollution. It's like tourism without the tourists.

The locality in this story, along with the state of Virginia, offer a lot of incentives for datacentres to site there, which is why they have so many. Meanwhile, the electricity price today is just slightly higher than it was in 2000, corrected for inflation, and lower than it was in 2010.

Also, turning off the lights in empty rooms is something we learned to do as soon as we were tall enough to reach a light switch. The fact that adults have to be told to do this suggests electricity is still too cheap in Henrico County.

This isn't a new thing. Soviet doctrine for dealing with carriers was to fire as many missiles as they could at them to overwhelm their defences.

The US focus on expensive, highly capable weapons isn't baseless. The US military is designed to project power, and projecting power is expensive. If you're going to ship equipment halfway around the world and support it there, it might as well be the best you can make. That is very different from Ukraine or South Korea where they are, or expect to, fight in their literal back yards.

Expensive invading armies have always been vulnerable to the defending swarm, from hoplites and knights to empires getting their asses kicked in Afghanistan, Vietnam or Iran.

The people building the datacentres had a plan. The people approving them, and offering big tax incentives, also had a plan. Part of that plan was electricity rates going up.

No, the grandmaster is doing many, many more calculations, just in parallel instead of serial. The human is also not blindly searching a tree of possible moves but spending a lot of computation on figuring out which are promising branches to prioritize. Modern chess programs are so good because they do the same thing.

A human can learn with 20 hours of driving school

No, they can't. We don't generally let humans even attempt driving for something like 16 years. They're also pretty shit at it until they have a decade or so of pretty frequent practice.

Ah, right, that paper. I don't think they'd use the word "continuous" the way they did if they thought about it for a bit either. They use it as a vague throwaway in the abstract and then never again. Also "fairly discontinous" is silly. It's either is or it isn't, nothing in between. That's kind of the defining property of a discontinuity.

What they actually mean is this:

Our main result is that for deep neural networks, the smoothness assumption that underlies many kernel methods does not hold. Specifically, we show that by using a simple optimization procedure, we are able to find adversarial examples, which are obtained by imperceptibly small perturbations to a correctly classified input image, so that it is no longer classified correctly.

Deep neural networks can be "highly nonlinear" (which they note). The output of nonlinear systems can vary a lot in response to small changes in the input. In fact, one way of defining smoothness is as an upper bound on local nonlinearity, so this result shouldn't be at all surprising. However, discontinuities are only a subset of things that are nonlinear and not smooth. In fact, the optimization they use to genrate their adversarial examples depends on the neural network being continuous (down to the inherent discretization of the datatype).

Unfortunately a lot of other people only read the abstract of this paper and completely missed this (the very next paragraph after the one I quoted above):

In some sense, what we describe is a way to traverse the manifold represented by the network in an efficient way (by optimization) and finding adversarial examples in the input space. The adversarial examples represent low-probability (high-dimensional) “pockets” in the manifold, which are hard to efficiently find by simply randomly sampling the input around a given example. Already, a variety of recent state of the art computer vision models employ input deformations during training for increasing the robustness and convergence speed of the models [9, 13]. These deformations are, however, statistically inefficient, for a given example: they are highly correlated and are drawn from the same distribution throughout the entire training of the model. We propose a scheme to make this process adaptive in a way that exploits the model and its deficiencies in modeling the local space around the training data.

The concept of adversarial examples isn't about "lol, neural networks are dumb and easy to fool." They're about efficiently generating supplemental training data that makes the model more robust.

Szegedy et al also only study one specific type of model, although with some internal variations, and there is a discontinuity in that model. It has nothing to do with the neural network though. It's the very last step where you decide that, for example, a vector of probabilities like [0.3, 0.29999, 0.30001, 0.1] == [0,0,1,0]. It's called thresholding, and we sometimes do it because it's necessary to make a decision, but we also do it a lot because humans don't deal well with uncertainty.

I think you can find a similar adversarial-plus-thresholding example for the human brain in optical illusions like Rubin's Vase. You see a face or a vase, switching back and forth as you imperceptibly change your focus on different parts of the completely static image. You also don't see both at the same time, it's pretty strongly either or. "Fairly discontinuous" if you will.

That must have been a while ago. Computers are effectively unbeatable at chess. One of the best programs was written by a Norwegian nerd in his spare time and then forked by an impatient Italian. It will run on your phone and will almost certainly kick your ass, although if you want to be sure of beating every human who's ever lived you might want to give it a desktop computer.