The average is more useful if you wish to answer the question of how much energy the entire Google is using, but there are other ways to answer that question (such as Google's annual report, which tells you to 8 significant digits how much energy they used).
The median is more useful to answer the question of how much energy a typical Gemini interaction uses.
Agree, but given there is such a big distribution in prompt energy usage, wouldn’t it then make sense to break it down further into some meaningful categories?
Also, they say median and mean diverge a lot. So, in my view, google should disclose both and discuss it. Understanding avg/total consumption of AI use is relevant here.
Do they include AI search summaries here? It would be a big no no in my view.
Google has rolled out AI summaries extensively over the time of the study and they likely use more efficient inference than chatbot prompts to their larger models.
They discuss the median in their paper. But I couldn’t find any breakdown about how the prompts they analyze are distributed across their models.
In this thread alone there are many comments multiplying the median to get some sort of totalt, but that's just not how medians work.
If I multiplied my median food spent per day with the number of days per month, I'd get a vastly lower number than what my banking app says.