jsomae

The AIs can definitely get more advanced, sure, but with that should come some sort of efficiency.

This is what AI researchers/pundits believed until roughly 2020, when it was discovered you could brute force your way to have more advanced AIs (so-called "scaling laws") just by massively scaling up existing algorithms. That's essentially what tech companies have been doing ever since. Nobody knows what the limit on this is going to be, but as far as I know nobody has any good evidence to suggest that we're near the limit of what's going to be possible with scaling.

We’re also seemingly on the cusp of quantum computing, which I imagine would reduce power requirements.

Quantum computing is not faster than regular computers. Quantum computing has efficiency advantages for some particular algorithms, such as breaking certain types of encryption. As far as I'm aware, nobody is really looking to replace computers with quantum computers in general. Even if they did, I don't think anyone has thought of a way to accelerate AI using quantum computing. Even if there were a way to, it would presumably require quantum computers like, 15 orders of magnitude more powerful than the ones we have today.

We have very, very real and very, very large environmental concerns that need addressing.

Yeah. I don't think AI is really at the highest level of concern for environmental impact, especially since it is looking plausible it will lead to investing in nuclear power, which would be a net positive IMO. (Coolant could still be an issue though.)

Did you not notice I said "otherwise"?

Regardless, terrorist is a subjective term, and typically one excludes agencies with jurisdiction or organized militia; so the police, ICE, U.S. military, IDF, etc. don't qualify regardless of your disposition toward them. This is why newspapers are usually quite hesitant to use the word "terrorist." There's a good case for calling the CIA (and other foreign-intelligence TLAs) terrorists though, since they operate outside of places where they have the authority to do so. Obviously, it's semantics at the end of the day.

I agree with every part of this except for the word "most likely," since the incidence of ICE arrests are extremely high right now but terrorism is otherwise not.

Regarding (1), it still gets the network effect because of iOS users. That it's harder for android users to install does not mean the marginal utility of an android user installing the app is lower than that of an iOS user installing the app.

That's a good point. I rescind my argument that training is necessarily more expensive than sum-of-all-deployment.

I still think people overestimate the power draw of AI though, because they're not dividing it by the overall usage of AI. If people started playing high-end video games at the same rate AI is being used, the power usage might be comparable, but it wouldn't mean that an individual playing a video game is suddenly worse for the environment than it was before. However, it doesn't really matter, since ultimately the environmental impact depends only on the total amount of power (and coolant) used, and where that power comes from (could be coal, could be nuclear, could be hydro).

I think the more important takeaway from this article is not the political one, but that problems with a biological fix can be extremely easy to solve compared to problems that require societal change, even though we normally think of it as being the other way around.

If you are expecting AI to not have much impact and turn out to be a bubble, then I guess there isn't much reason to believe it it will have much environmental impact. If you expect AI to not be a fad, then yeah it could have big environmental consequences if we can't find renewable power and coolant. If AI is all it is hyped up to be, then it would dwarf the rest of humanity's power consumption down to a footnote. So it really depends on how bullish you are about AI, or at least how bullish you expect the market to be going forward.

Regarding proof-of-work crypto, well, bitcoin is currently at its all-time high in terms of value, exceeding USD$100k/BTC. So I'm not sure I exactly buy the idea that it's less popular, though perhaps people aren't reporting on it as much. If the power consumption of crypto has levelled off, which I don't know if it has, then it might be because it's expensive to build a mining rig and the yield goes down over time as more bitcoin is mined. (It's presumably true of other proof-of-work crypto, too, but as more BTC is mined, the marginal yield of mining more BTC decreases.)

What community would you suggest? I didn't realize /c/science was for pop-sci publications.

How is it any worse than crypto farms, or streaming services?

These two things are so different.

Streaming services are extremely efficient; they tend to be encode-once and decode-on-user's-device. Video was for a long time considered a tough thing to serve, so engineers put tons of effort into making it efficient.

Crypto currency is literally designed to be as wasteful as possible while still being feasible. "Proof-of-work" (how Bitcoin and many other currencies operate) literally means that crypto mining algorithms must waste as much computation as they can get away with doing pointless operations just to say they tried. It's an abomination.

never heard of the ben franklin effect before, neat.

Maybe you should stop smelling text and try reading it instead. :P

Running an LLM in deployment can be done locally on one's machine, on a single GPU, and in this case is like playing a video game for under a minute. OpenAI models are larger than by a factor of 10 or more, so it's maybe like playing a video game for 15 minutes (obviously varies based on the response to the query.)

It makes sense to measure deployment usage marginally based on its queries for the same reason it makes sense to measure the environmental impact of a car in terms of hours or miles driven. There's no natural way to do this for training though. You could divide training by the number of queries, to amortize it across its actual usage, which would make it seem significantly cheaper, but it comes with the unintuitive property that this amortization weight goes down as more queries are made, so it's unclear exactly how much of the cost of training should be assigned to a given query. It might make more sense to talk in terms of expected number of total queries during the lifetime deployment of a model.

Statutory rape in particular. We need to normalize the usage of the term statutory rape.