There's always going to be new things to power, this is good. That's why we have to replace our non-electric infrastructure with more efficient alternatives and scale up green energy.
There are two ways to reverse course on anthropogenic climate change: reducing energy use and displacing dirty power sources with clean ones (including fission). I've heard takes like this one that its a net good that companies like Microsoft are investing in green infrastructure. I'd like to challenge that idea. I don't believe the juice is worth the squeeze here. At best, we're treading water here on energy usage. If we really consider the costs of running these systems, which is substantially more than more "traditional" systems like search engines, and we really consider the value of an unreliable ai text generator, is what we're doing here really an unambiguous good for the human race?
Neither of those ways reverses course, they only reduce emissions. But we have to go net negative on emissions to reverse global warming.
Electric devices are not a major issue, since they can be powered by green energy easily. There's no reasonable path forward that involves picking and choosing what software has acceptable "value" to budget energy for (besides just market forces). We just have to build the power infrastructure and try to make computing energy efficient.
Pedantic opening line aside, I think software engineers in general underestimate how much of our electrical infrastructure still runs on coal and natural gas, and also underestimate the amount of time required to stand up new capacity on "green" sources, especially nuclear.
I consider it reasonable to have a conversation about AI with respect to how worth it it is vis a vis climate goals. The verge just ran article that stated that Microsoft's emissions had increased substantially in 2023 even with investments in green infrastructure, with one of their executives stating explicitly that AI had pushed back sustainability goals. I also think it's obvious that the free market has failed to have a meaningful positive impact on this issue considering that this entire problem is caused the negative externalities of emissions, despite solar becoming the cheapest source of electricity.
I also disagree that it's not worth discussing computing efficiency with respect to this problem. Crypto is rightly criticized for using a ton of power to do essentially zero useful work. That power could have gone towards displacing some kind of fossil fuel usage but it didn't and that means we now have more work to do with respect to climate goals. According to the article we're commenting on here, Virginia could have powered 1.5 million homes with the power they used on the AI data center boom. And what value did we really get out of it? I believe strongly that it's fair to critique this.
Increasing efficiency is good, yes. Crypto sucks because it doesn't do anything useful and intentionally burns compute. If AI is not useful, then we will not need to spend the power on it because no one will want or need it. If it is useful, we will need to spend the power on it.
> Crypto sucks because it doesn't do anything useful and intentionally burns compute
Thats not true. Cryptocurrencies, such as Monero, enable you to completely anonymously send money fast and cheap over the internet. This has value for many people.
I disagree that that's always true. AI consumes some value and produces some other value. I'm skeptical that the output value will be worth more than the amount of money and electricity invested in this stuff.
This is kind of headline bait. I work in energy forecasting so I look at these numbers all day. Facebook's 350k H100s might draw around 1kW each, let's say 2kW with the rest of the rack and cooling. That's 700MW for what is probably the largest AI cluster in the world.
You know what's as big as that? Riot Blockchain's Rockdale mining facility (700 MW capacity).
Which one do you think provides more value, training Llama 3 or mining bitcoin?
The reason I'm not worried is because solar is now dirt-cheap, and compute is quite interruptible. If it costs you $50/MWh to train your LLM during the night but $5/MWh to do it during the day, maybe you're ok taking 2.5x as long to save $45mm on wholesale electricity costs.
I'm not saying it's trivial or even easy, but industrial energy usage is huge already. Freeport LNG is also ~700MW. EV-ification and crypto are already huge stresses on the grid, I don't think AI data centers are the only story.
And can you tell about open ai consumption ? I'm curious.
It's just a personal opinion but don't think because solar energy is cheap that it must be consumed.
Also if I had to choose, I very much prefer to have heat in winter and freshness in summer than have AI and crypto.
