They didn’t account for training. From the paper:

> LLM training & data storage: This study specifically considers the inference and serving energy consumption of an Al prompt. We leave the measurement of Al model training to future work.

This is disappointing, and no analysis is complete without attempting to account for training, including training runs that were never deployed. I’m worried these numbers would be significantly worse and that’s why we don’t have them.

▲

sbierwagen 3 days ago | parent [-]

If I download a copy of llama and run a single query, what was the cost of that query?

	▲	progval 3 days ago \| parent [-]
		No, because you don't incentivize the training of the next version of LLama, and the current version was not trained because you wanted to run that query. This is not true of Gemini.