It's (mostly) not the networking layer where people pay to target us. It's the application layer that would most benefit from being forked.
Of course the problem is that what can be forked already has been. Federated social media. Distributed git hosting. However most "essential" uses are centralized and often also commercial in nature. If you fork Amazon you're ... still Amazon. That sort of thing.
We need to go full Oracle and charge an excise tax per logical CPU core. For GPUs we can count SIMT lanes.
More seriously they should be taxed per watt, likely in an asymptotic manner because most of the externalities don't scale linearly. Any additional infrastructure requirements should be directly rolled into their electric and water bills, which is to say that they should receive a very unfavorable rate.
> that they consume a lot of water based on a discredited book with elementary math errors.
How exactly do you think they dissipate the heat of a continuous 100 MW or 1 GW power draw? I have no idea what book you're referring to but you can do the math yourself it's quite straightforward.
Basically, the author (of the book) compares a data center outside Santiago to usage of water by humans, erroneously imputing that the average human uses only 200 cc of water per day.
Not true. Electric vehicles have been threatening to collapse residential grids for quite a few years now. The US hasn't been making the necessary infrastructure investments for a long time. See PG&E for example.
For something the size of the electrical grid, you can find regional variations, but the national trend is quite clear. One report from a quick search[0]
Consumption Growth Acceleration: After 14 years of near-stagnant growth (0.1% annually from 2008-2021), US electricity consumption surged 3.0% in 2024, driven by data centers, electric vehicles, and economic recovery, signaling a new era of demand growth.
> A modern AI data center uses 20-100MW+ of electricity.
I understand the high end builds to have exceeded 100 kW per rack at this point, with the largest sites exceeding 1 GW (ie 10x your upper bound). So the smallest datacenters use as much as the largest auto plants, and the largest datacenters use 100x that.
They get their own unique third category as unlike industrial sites there's no hazardous chemicals and even the noise pollution is substantially different in nature.
The old datacenters are analogous to office buildings that emit some unusual noise and consume large amounts of electricity.
The new ones (ie gigawatt class) consume enough electricity for ~1 million households and at minimum enough water for 100k households (but possibly many times that).
I believe evaporative cooling is the norm (thus my "possibly many times that" remark doesn't apply) however theoretically they could provide hot water as a utility or (as you say) just dump it into the sewer. If located next to a river or the ocean they could conceivably dissipate it that way but I'm not aware of any examples.
It's the sort of externality that could be solved with a well placed megaproject. A related question to my mind is why we're building such expensive strategic assets in the open rather than under a mountain.
That isn't the whole story. At least some of these new datacenters are gigawatt class. That's multiple sq km of solar.
Water usage is also an issue. A continuous 1 gigawatt is enough to boil off 1.3 million liters per hour which over 24 hours equates to very roughly 90k residential users. If it isn't boiled but is instead returned lukewarm it will require many times that amount due to how large the heat of vaporization is. Compare to the entire state of Florida at "only" 23.5 million people.
What? The water is not getting boiled off. Datacenters, for the most part, have closed liquid loop cooling systems. Electricity goes in, hot air and bits come out.
did you move the goal post, or erect a new one? either way- residential use is penny ante in terms of water usage. So much so that comparing data center use to residential use without including industrial, commercial, and irrigation can only be in bad faith.
Particularly since usage reports typically present all the numbers in the same chart or grid.
The concern is resource usage. Water had been left out, so including it isn't shifting the goalposts given the context.
The comparison was intended for illustrative purposes. Residential usage provides something relatable and is the general standard for these sorts of discussions.
Even comparing to industrial most operations don't use anywhere near as much electricity or water. The new gigawatt class datacenters are in the same ballpark as aluminum smelters, but rather than melting metal they sink all that energy into water.
Perhaps more importantly, will their chain of thought be "real"? So far the ones I've seen seem to be elaborate fakery. They look good unless you dig in at which point you often find that it merely looks plausible on the surface but that something else is going on under the hood.
I don't know what you mean by that. We know what's going on under the hood always: linear algebra, the attention mechanism etc.
To my first approximation all "Chain of thought" means is that instead of having to prompt the model to discuss everything in text and then decide at the end[1], now it sort of automatically does that so you don't need to prompt it.
[1] Which used to bring about very substantial improvements in performance on some tasks
I think it was clear from context that "under the hood" wasn't referring to the math but rather to the contents of the trace. What's written (often?) isn't what's actually being "thought" about. The trace is a trained output similar to the final output, which is to say that it's fake. There are research papers on the topic, particularly that models can be trained to print other arbitrary stuff during the "thinking" phase instead.
You can easily see this for yourself by carefully walking through a given trace with a critical eye. Here's an example from myself a few days ago. https://news.ycombinator.com/item?id=47623324
Yeah now I get what you're saying. Yes the trace isn't what's actually happening. What's actually happening is just the attention mechanism etc. The model doesn't "think" in human language, it thinks in linear algebra. The thing is that before chain of thought it used to be necessary to get the model to output some language because that's the only thing it had to attach processing to (so if you wanted more processing you needed to get it to generate more text). Whereas now we get the model to generate some text that is a simulcrum on the thought that it might hypothetically be doing but in actual practise chain of thought is just something they get the model to do by training it in a certain way.
> Drip feeding arms, arms meant to intimidate through the prospect of overwhelming force no less, into air defenses below replacement rates is just dumb.
That probably depends on the cost of the arms, the cost of the interceptors, and any number of other externalities or indirect goals. If you can reliably induce high end interceptors to fire against cheap rockets (granted, that's a big if) you are definitely winning the immediate economic exchange.
> If you can reliably induce high end interceptors to fire against cheap rockets (granted, that's a big if) you are definitely winning the immediate economic exchange
Tactically sensible. Strategically foolish.
The deterrent value of Hezbollah’s arsenal was in overwhelming Israeli defenses and causing loss of life. That is what democracies, first and foremost, respond to. (Second being cost of living.) Spending a potent deterrent to play economic attrition with Israel, a rich country with a richer friend, was stupid.
Of course the problem is that what can be forked already has been. Federated social media. Distributed git hosting. However most "essential" uses are centralized and often also commercial in nature. If you fork Amazon you're ... still Amazon. That sort of thing.
reply