As an example, Ren and his colleagues calculated the emissions from training a large language model, or LLM, at the scale of Meta’s Llama-3.1, an advanced open-weight LLM released by the owner of Facebook in July to compete with leading proprietary models like OpenAI's GPT-4. The study found that producing the electricity to train this model produced an air pollution equivalent of more than 10,000 round trips by car between Los Angeles and New York City. (https://news.ucr.edu/articles/2024/12/09/ais-deadly-air-poll...)

see also:

https://www.techrepublic.com/article/news-ai-data-centers-dr...

https://www.scientificamerican.com/article/a-computer-scient...

> The study found that producing the electricity to train this model produced an air pollution equivalent of more than 10,000 round trips by car between Los Angeles and New York City.

I am totally on board with making sure data center energy usage is rational and aligned with climate policy, but "10k trips between LA and NY" doesn't seem like something that is just on its face outrageous to me.

Isn't the goal that these LLMs provide so much utility they're worth the cost? I think it's pretty plausible that efficiency gains from LLMs could add up to 10k cross USA trips worth of air pollution.

Of course this excludes the cost of actually running the model, which I suspect could be far higher

> 10,000 round trips by car between Los Angeles and New York City.

That seems like very low impact, especially considering training only happens once. I have to imagine that the ongoing cost of inference is the real energy sink.

It doesn't happen only once. It happened once, for one version of one model, but every model (and there are others much larger) has its own cost and that cost is repeated with each version as models are continuously being retrained