Post

"[...] the average [ChatGPT] query uses about 0.34 watt-hours, about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes. It also uses about 0.000085 gallons of water" - Sam Altman in blog.samaltman.com/the-gentle-s...

blog.samaltman.comThe Gentle SingularityWe are past the event horizon; the takeoff has started. Humanity is close to building digital superintelligence, and at least so far it’s much less weird than it seems like it should be. Robots...

built this over the weekend — these estimates are very very different than what folks have assumed… think we can trust them?

dame@dame.is

New web app ✨ It’s hard to understand the impact of personal AI usage on water & energy consumption, so I made an interactive + visual conversion tool! 🧮 See how many ChatGPT queries you could make for 1 hamburger. 🤖 🟰 🍔 My goal was to help people make informed decisions. Try it: context.supply

would also be interested if you had any feedback on this tool in general, i’m not nearly as well educated on this topic as you are

What figures are you using for that? I have found it infuriatingly difficult to get numbers that feel credible to me

0 reposts4 likes

combination of things, there's an info section at the bottom of the site with some links... but it has an optimistic, neutral, and pessimistic calculation modes that gives a very wide spread to help folks think about it the neutral mode just assumes the 3 watt stat that most people reference/claim

GPU has become more efficient ( at least for Nvidia b200 and Google's TPU). And model size decreased thanks to distillation, all public models are distilled versions of bigger, slower, and more expensive models.

I've been assuming that efficiency gains are at least somewhat reflected by the prices the different vendors charge for their models via their APIs, and those prices have gone down 100x over the last few years

I think for the average user it probably has gotten signficantly better, but because of so many other more powerful models being used at higher rates by "pro" users, things could easily just be balanced or negated? We just don't know at the moment.

Yes, I'm aware. I think it is much much much smaller on average, but I chose to default to that because the tool's intent is to just paint a clear/simple picture that even if that number is true... it still isn't a signficant or meaningful impact

If I defaulted to something much smaller, a large number of people (critics/skeptics) would probably just dismiss the tool outright rather than playing with it at least, that was my thinking

it was a weekend project so i need to polish it a bit more, i want to make it clearer which numbers came from which data source