I literally linked to a scientific article explaining why it cannot be done in a reliable fashion. I even pointed you to the part of the article where it says so.

I did not write the article, it was written by scientists a lot smarter than me and it was peer reviewed and accepted for publication in a highly respected international publication.

If you claim to be smarter and has more knowledge than them, then all the more power to you - but until now, you have only made unsubstantiated claims with no documentation or references.

LLMs are just software. if you run it on distributed system, then you have to measure accordingly. But in most case, especially for image generation, it runs on a single node, a single gpu. And of course you can measure that. And everything that can not be measure most definitively can be estimated. I really don't much care how much effort it is for them to do this, they dug this grave themselves.

What, you're saying they don't know how much energy usage their tokens translate to? What kind of bullshit business is this supposed to be where no one knows anything about anything they do? How would this ever be supposed to work?

If that really is the way these companies work it's no wonder they're all so deep in the red, i guess.

Of course you can't. To the best of my knowledge there is currently no standard for measuring energy usage for running LLM's in any way, shape or form, probably beacuse it all depends on where the data center is located, how much the hardware is loaded at the time of inference and how cooling is achieved.

Even training is not a fixed cost. Depending on who you ask, training can make up anywhere from 10% to 40% of the total energy consumption, and considering that the average energy consumption per query for training is getting lower with every query, it's impossible to calculate on a per-query basis.

See this report for a deeper explanation:

https://www.sciencedirect.com/science/article/pii/S1364032125008329

Scroll down to this part and what comes below it:

"2.4. Discrepancies in estimations and underlying reasons
Existing estimations on LLM energy consumption, including both training and inference stages, exhibits substantial discrepancies, even for the same model."

And i honestly don't care if people hate it less than i do. There are plenty of people who defend combustion engines, that doesn't make me want to protect their fragile egos either.

There is no way you can determine the energy usage like that. Hate AI all you want, but please don't make others have to defend why the don't hate it as much as you do. Seriously. Don't.

With that said, I would really like that there are some pretty clear consequences if someone publishes AI-generated content without labelling it as such, at least when it is used for commercial use.

AъBCEFC

@digitaleu