Andy Masley finds ChatGPT's water usage is 50 to 250 times lower than the viral 'bottle of water' claim
Simon Willison urges OpenAI to release retired GPT-4 specs.
#193Simon Willison@SIMONWGiven how much of the original "bottle of water per generated email" water estimate came from guesses at the architecture of GPT-4, it would be very much in @OpenAI's interest to publish the architecture of that now-retired, three year old model
Andy Masley@AndyMasleyhttps://blog.andymasley.com/p/i-might-have-found-the-specific-way
Andy Masley@AndyMasleyA lot of people have started saying that the "AI uses all the water" meme started with Empire of AI. I think this is definitely wrong, and what actually started it was the Washington Post article from 8 months earlier claiming that each ChatGPT prompt used a whole bottle of water. This kicked the meme into the stratosphere. There are so many infographics made about AI mentioning that it uses a bottle per prompt, so many thousands of popular videos and images where someone holds a bottle begging people not to use AI. It got so much news coverage. My claim in my new post (link below) is that I've found extremely strong evidence that this claim was based entirely on napkin math that ignored lots of simple things we knew about the hardware and software running GPT-4, and if you just account for those the cost drops by 50 to 200 times. This shouldn't have ever been allowed to influence the discourse at all, never mind get an infographic made in a major newspaper. I think the Washington Post should either retract or correct the original article and make it clear that it was based on napkin math, or publish the methodology. The authors have never made it clear how they got the number, and if I'm right it shouldn't have been allowed to ever be the basis for a claim this strong.