Training a single AI can emit as much carbon as five cars in their lifetimes

technologyreview.com

11 points·jasaloo··3 comments

"In particular, they found that a tuning process known as neural architecture search, which tries to optimize a model by incrementally tweaking a neural network’s design through exhaustive trial and error, had extraordinarily high associated costs for little performance benefit."

This is specifically the ONE instance that uses five cars worth of CO2. The next highest example used 435x less CO2. Click-bait titles like this incite unnecessary ire.

If you're spending $1,000,000+ on cloud compute costs, you're (a) hopefully very aware that this is burning a lot of CPU/GPU cycles, (b) very unlikely to try it again if the gains are minimal as the paper states.

I think the ire is necessary.

The article gives a fair range of the different baseline operations and their carbon footprint. Sure the title is provocative, but it’s accurate.

And even the operation that’s still “435x less” is just under the CO2 footprint of a passenger on a transatlantic flight. That’s not insignificant at all.

The fact is that more industries are trying to get their hands on this tech, and they won’t go about self-regulating the types of energy-intensive operations they use if they think there’s profit at the end of it.

“In a new paper, researchers at the University of Massachusetts, Amherst, performed a life cycle assessment for training several common large AI models. They found that the process can emit more than 626,000 pounds of carbon dioxide equivalent—nearly five times the lifetime emissions of the average American car (and that includes manufacture of the car itself).“

Reply on news.ycombinator.com