And here I thought we were getting the code for the oldest denial of service attack known. I'm not sure anyone really cares about old versions of DOS.

Even unlimited sodas is a better cost model than this.

Ed Zitron interviewed a guy who said that there are supposedly thousand-fold reductions in inference costs coming down the pipeline. Ed is an AI skeptic, and I don't remember who the guy was, but he seemed knowledgable. Anyway, if that is true then "running" a model could be cheap, whereas "training" a model is still horrendously expensive. Somewhere between the two would be true cost.

I personally doubt it will ever be a viable business beyond what you're describing.

Also, there is one customer who doesn't care what it costs -- the military. So that is always an option.

Sweet, so almost everyone will be charged zero for the thing they aren't using.

Agreed. We're in the final stages of this AI "race to IPO", which is the final step in the tech grift cycle, for the VCs get "made whole", while "retail investors" take all the losses. Elon is trying to bootstrap this by tying everything to the SpaceX IPO.

It's truly just a cash-out for the original investors. Twitter is bought for $54 billion, loses money hand over fist, and has no path to profitability. The answer: form your AI spite company, vacuum up the "Elon" money, then buy Twitter to pay out those original investors (some of which just rolled into xAI). Oops, xAI is burning money and isn't going to be able to IPO quickly, let's bootstrap that into SpaceX. So now we have this rocket/AI/Social media company (which is really just a satellite ISP), and we need to pump the valuation before we dump it on the idiots. So space AI data centers is the anwers!@@#

I think Antarctica would have a problem with electricity during the dark months. But there are more accessible places that are nearly as cold. The Canadian Shield is cold, extremely stable geologically, and more accessible, even if you had to lay undersea fiber along the coast instead of over land. I'm sure there are many more places.

The issue would be that few of them are "politically" convenient, for all the jerbs.

The saddest part is that our society worships one-hit wonders. Get one thing right once and you're suddenly a genius at everything. It's literally just probability at work, a winning lottery ticket. How many times do you see the headline "PERSON WHO PREDICTED X DOES Y!!!!!!!!!!!!!". I used to work with a guy who would claim the sky was falling every day, was wrong 99.999% of the time, but you never heard the end of it the vanishingly few times he was right. Would have been more accurate flipping a coin.

There are about 60 GW of AI compute data centers in the world. I saw a Youtube video where some space guy claimed to do the analysis that a starlink-sized satellite could possibly handle the thermal load for 20kW of compute. If we triple that to 60kW, that means Elon needs to build and launch 1,000,000 satellites into space. Elon had 100 launch sites sending up one rocket per hour, you would be launching continuously for over a year. Note, there are only 20-something launch sites in use worldwide. The logistics of supporting such a cadence would be inconcievable.

Then, if you actually did this, how would you ever orchestrate such a constellation of satellites, each needing fantastic amounts of bandwidth between each other as well as ground stations. It's just absolutely ridiculous. Even if you said you were putting them on the Moon, we've only ever sent a few hundred metric tons of equipment there ever.

This whole "data centers in space" is just the next space elevator / dyson sphere bullshit concept.

That's because evolution is solving for a different problem. If organic "wheels" solved the problem, we would have them.

Being limited to a bipedal model is also pretty limiting. Animals with additional pairs of legs can navigate with much greater precision, ease, and efficiency.

I bet a one-wheel, battery-powered skateboard can beat it.

Pretty sure you can fly over every jungle that has ever existed.

But what is the use? This isn't a bending unit from Futurama. This is finally accomplishing the first of many rudimentary, basic tasks. This kind of generic robot suffers from an enormous "80-20" "long tail" problem, and they aren't even close to the 80% mark yet.

When one of these things can scale Mt. Everest unassisted, I'll say they have something to celebrate.

Literally not how the math works. The probability is the same, every year. Just like rolling a dice.

Also, this guy is maybe a great physicist but he's absurdly stupid about this. If his radioactive decay model of humams was correct, we would all be dead because the Soviets got their first usable H-bomb in 1955, far past the "50 year limit'.

That would have required uncoupling the heisenberg compensators.