The growing energy footprint of artificial intelligence ("AI") [pdf] | Hacker News Reader

The growing energy footprint of artificial intelligence ("AI") [pdf]

cell.com

The next Bitcoin. Energy usage will be argued about by all sides.

Model training can at least follow the sun or clean energy forecasts. Latency from distance isn’t an issue as it would be for serving users requests. You can stage data for training runs. If you train during periods where the price of power goes negative, you get paid to take the power.

These workloads are going to run, so best to run them where the impact is lowest for the effort involved.

https://blog.google/inside-google/infrastructure/data-center...

https://www.electricitymaps.com/blog/announcing-our-partners...

https://www.electricitymaps.com/blog/how-companies-reduce-th...

https://arxiv.org/abs/1906.02243

Bitcoin and other PoW crypto burn energy by design and found few to no enduring use cases.

Talking about the energy usage of AI is like talking about the growing energy usage of lighbulbs at the turn of the 20th century. It's a byproduct of adoption and reduced-energy solutions will readily be taken up if/when found.

It seems reasonable to suspect that Silicon Valley will turn AI into a bubble; just like it did with the Internet, just like it did with Bitcoin. Mystical futuristic technology always has greed hot on its heels.

Bitcoin was always an ideological gambit hoping “step 3: here a miracle happens” would make it viable - there was never a path to mainstream adoption without changes which would make it unrecognizable to improve security, privacy, and, above all, making it as fast and convenient as the competition.

In contrast, AI has plenty of hype but, similar to the dotcom-era internet, you can easily find a ton of people working on real things which actually work. Translation and copy-editing, for example, used to be expensive services, OCR and image search are now practical for problems which they previously weren’t, and I know people who have found LLMs helpful enough for things like learning or debugging code that they’re paying monthly subscriptions. That doesn’t mean it’ll print money or take over every job the way the big investors are claiming but there’s far more substance than there ever was in cryptocurrency.

Always look for what people spend their own money on, not what the VCs and analysts at big consulting companies are pitching as the second coming.

I mean it seems like the conversation these days is never about how Bitcoin was a libertarian anti-corruption response to the 2008 banking crisis. The miracle then is hoping enough people stop trusting banks. I feel like institutional money getting into crypto was the thing that killed it. A cheap investment to maintain control of the value of a dollar.

The ideology I was referring to was the builtin deflationary model: that’s a very strong ideological position in libertarian circles but since it immensely privileges wealthy early holders over everyone else, it’s a huge barrier for adoption.

Reply on news.ycombinator.com