Post

They should provide more detail and evidence if they want to persuade people on this

That’s around 10 seconds of energy of a nominal human. Most queries would take me more than 10 seconds to answer :)

That's true, but it doesn't change the cost of a query. The more recent difference is reasoning models, with more compute at inference time. These models consume a lot more.

I read this with about the same skepticism I have when Bitcoin bros talk about Bitcoin's energy footprint. "No no, it's good, bro, because it makes me rich enough to afford a Tesla, which reduces Bitcoin's energy footprint". You think Altman's figures include training costs?

This article finds that a similar model uses about 1.4 Wh per prompt fwiw

For what it’s worth, I found a similar number of 0.3 watt-hours for an average GPT 4o query diving down the AI is power hungry rabbit hole the other day. Fun Fact: at 0.34 Wh per query, each query uses the same amount of power as streaming 16 seconds of HD content on netflix epoch.ai/gradient-upd...

Maybe he's "forgetting" some crucial detail like it being per token or "when it hits the cache"...

www.technologyreview.com/2025/05/20/1...

Nice to at least have someone in a company put a specific number on it

Prompt processing has never been the computationally heavy part. Also OpenAI, Anthropic, etc. have been caching prompts for a while now, amortizing the system prompt to nothing.

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

I wonder if it included training time

But can we really trust anything Sam Altman says?

Sam says he'll be gentle and we should all lay back, relax, and it'll be over soon. I feel so reassured.

Interesting article. Def agree all the equations around intelligence that shape our society are changing, and we have no idea what to do about that

I do, now. My phone still defaults to Google though.