AI Tradeoffs ⚖️.

In October I gave a quick (5 minute) lightning talk at Mozilla about the environmental impact of using AI vs. the potential environmental benefits. In the talk I make a Deep Fake of myself and estimate the energy usage in creating it.

You can watch the video at the link above, but here’s a summary of things I learned in the process:

AI Training Energy Usage

  • Machine Learning Inference (using Machine Learning models to generate content vs training them), has a fairly minimal energy impact. Using a software called Code Carbon, I calculated that you could use Stable Diffusion for 3 hours and it would have about the same impact as a full charge of a smart phone.

  • However, energy impact is most strongly influenced by how clean the energy grid it is drawing from. I used a Machine Learning Carbon Calculator and determined that training a model for 100 hours in Google Cloud’s Asia South region would consume forty six times more carbon than a GPU in the europe west region

  • ClimateChange AI keeps a good list of academic papers on the positive application of AI in environmental response including traffic and energy forecasting and optimization, reducing system waste, and environmental monitoring and modeling.