Conversation

Notices

1. Embed this notice
   Security Writer :donor: (securitywriter@infosec.exchange)'s status on Tuesday, 29-Jul-2025 23:48:23 JST Security Writer :donor:

   Maybe it’s time to start the anti-AI propaganda by telling people their utility bills will rise due to AI usage.

   Because they will.

   Taxes and utility bills seem to be the only language some people understand.

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:50:00 JST bemmesr

     in reply to

     Given the sheer number of flights world-wide, it's not unfair to suggest that the energy cost of training all LLMs thus far is eclipsed in a relatively short amount of time of average air-traffic (Perhaps a few weeks or months?).

     But how about the energy cost of **using** AI? According to blog.bestai.com (linked above) and https://dev.to/nilanth/how-much-energy-does-chatgpt-use-per-prompt-a-look-at-its-hidden-environmental-costs-2j3a the energy usage for one prompt is somewhere between 5Wh and 10Wh, whereas making a cup of tea with a kettle costs something like 50Wh.

     (continued)

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:50:01 JST bemmesr

     in reply to

     @SecurityWriter I struggle to grasp this somewhat common concern.

     After reading some articles sharing the same concern, I've been unable to fully justify those worries.

     https://blog.bestai.com/demystifying-ai-energy-consumption-everyday-comparisons-that-make-it-real/ claims that training something like GPT-3 consumes as much energy as '515 transatlantic flights,' but according to https://travelasker.com/how-many-transatlantic-flights-are-there-per-day/ there are about 1,400 transatlantic flights each day, and according to https://en.macromicro.me/series/7251/flight-tracking-commercial there are more than 140,000 flights total every day.

     (continued)

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:51:42 JST bemmesr

     in reply to

     They then state that 1,287 megawatt hours is enough to power 120 average US homes for a year. According to https://housegrail.com/how-many-houses-are-in-the-us/ there were more than 140 million houses in the US in 2020. Thus, the energy cost of training one LLM represents roughly 3/35000 of a percent of the energy used by all homes in the US in a year.

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:51:43 JST bemmesr

     in reply to

     I don't know about you, but I have a couple cups of tea every day, and honestly, I usually overestimate the amount of water needed by a substantial margin, so I think it's safe to say that I easily use more energy making tea than I do by using AI. Although, this might not be true for everyone, as I drink **a lot** of tea and only **rarely** use an LLM.

     (continued)

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:51:43 JST bemmesr

     in reply to

     Still, tea isn't the only energy cost. There's hot showers, gaming PCs, TV screens left on, driving your car instead of using public transport, central heating, air conditioning, and endlessly more.

     All in all, if the energy costs of AI are really such a big worry, then perhaps we ought to be cutting back on unnecessary travel and at-home energy use, too.

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:51:43 JST bemmesr

     in reply to

     In response to @SecurityWriter (despite the fact he blocked me):

     I'd like to thank you for recognizing my talent for producing 'propaganda, misinformation, and logical fallacies,' and also fully appreciate that this is likely a subject that causes you a lot of stress, which at least slightly justifies the hostile tone of your reply.

     Either way, I'm glad you took the time to respond and even read some of the material I linked to.

     Now, I want to address your response piece-by-piece:

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:51:43 JST bemmesr

     in reply to

     1. 'ChatGPT doesn't consume jet fuel.' I'm aware of this fact, but I think we can both agree that jet fuel is a store of energy, and can in fact be used to power computers too. I also think it's worth noting that jet fuel is much harder to use in an environmentally friendly manner than electricity.

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:51:43 JST bemmesr

     in reply to

     2. '[Generative AI] serves little to no purpose.' Clearly, this is a statement reflecting your opinion, as there is no consensus suggesting generative AI to be utterly useless. I would argue there are strong indications that it is in fact useful, for example, the large number of people and businesses willingly paying to use it. Now, I'm willing to accept that businesses are overestimating the utility of generative AI and overspending on it, but I'm not willing to accept that it...

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:51:43 JST bemmesr

     in reply to

     that it 'serves little to no purpose' unless you'd be willing to back that up.

     3. You're right, it does consume water. I've not done enough research myself to have a position in relation to whether the amount of water it consumes is too much, but suffice to say I wasn't even referring to water, my points mostly had to do with its energy consumption. Perhaps I should have written a thesis statement expressing that.

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:51:43 JST bemmesr

     in reply to

     4. It makes sense that '[Generative AI] does consume locally produced energy' (I mean, all energy is local to *somewhere* right?) but I'm not sure how it could significantly increase the price of energy when most countries have a national grid (meaning prices are likely balanced across the grid, and not dependent on local circumstances. Although, correct me if I'm wrong about that), and considering my previous points...

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:51:43 JST bemmesr

     in reply to

     considering my previous points indicating that the energy consumption of generative AI is in some ways less than that of a single household on a per-user basis (whether that holds for the training phase is not clear given that you've raised concerns about the validity of comparing AI to airlines), I would argue that we should not see significant price increases.

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:51:43 JST bemmesr

     in reply to

     5. 'But also, the source you use to evidence energy usage is bestai dot whatever.' True, the name is a little unfortunate and it's something I should have taken more seriously. Rest assured there are other sources citing similar numbers. I did provide a second source for power consumption on a per-prompt basis, because I was worried about that name.

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:51:43 JST bemmesr

     in reply to

     I think it's potentially fallacious to assume a source is unreliable simply because it has a vested interest in the subject matter, but I do concede that the selection was a little irresponsible of me. https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117 states 'the training process alone consumed 1,287 megawatt hours' in reference to GPT-3.

     Ethan Black repeated this.
   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:53:53 JST bemmesr

     in reply to

     7. 'a GPT prompt uses 6-40x the energy of a standard search.' I would point out that the given range of 10-15x lies within your range, so you don't seem to be in disagreement.

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:53:53 JST bemmesr

     in reply to

     8. ' "we need to stop using standard conveniences to make up for AI Bros blatant disregard for humanity so they can steal more IP and cut down more rainforests".' Well, I'm afraid you're just missing my point now. I would never suggest the oppressed should yield to their oppressor, or that we should make way for billionaires simply because they're richer than us.

     Ethan Black repeated this.
   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:53:53 JST bemmesr

     in reply to

     I'm trying to argue that the energy consumption of AI is comparable and perhaps even smaller than basic household amenities. That's to say, we should be willing to accept such an amount of energy consumption if we are already accepting it elsewhere. I think your mock interpretation of my viewpoint reveals some strong ideological oppositions you have to AI, besides the energy consumption alone.

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:53:54 JST bemmesr

     in reply to

     I think it's also clear to see that there are orders of magnitude of wiggle room in these numbers and the point would still stand. I hope this new source is enough to back up my original point.

   • Embed this notice
     bemmesr (bemmesr@mastodon.social)'s status on Wednesday, 30-Jul-2025 14:53:54 JST bemmesr

     in reply to

     6. 'the information is so far off the mark ... A standard Google search uses 15x the energy of a prompt' I think you've misread the article here, as it indicates the reverse https://dev.to/nilanth/how-much-energy-does-chatgpt-use-per-prompt-a-look-at-its-hidden-environmental-costs-2j3a 'To put this into perspective, [a prompt's energy usage] is roughly 10 to 15 times the energy consumed by a standard Google search.'