After Robert Downing Jr.’s announcement at Amazon’s re:MARS conference to start The Footprint Coalition – an organisation that will be committed to using AI, robotics and other such technologies to combat climate change – it’s inevitable that even more people – especially the Marvel Comic Universe fans– would join the AI-bandwagon. However, a recent article by MIT Technology Review on a study by the University of Massachusetts – Amherst provides evidence that our use of AI may in fact be a key player in contributing to climate change.
The UMass researchers conducted an assessment of the environmental impacts associated with the different stages in training several common large AI models – specifically NLP models. They did so by measuring the energy consumed in a single day to train the models, then used the estimated figures of the total expected training time of the models (as found in their original papers) and calculated the total energy consumed during the entire training process.
They found a shockingly alarming result: the training process can emit an equivalent of over 626,000 pounds of carbon dioxide.
It costs about the same amount CO2 emissions to train and optimise the NLP model as the lifetime emissions of 5 average American cars – including their manufacturing!
What is NLP anyways and why is it important?
NLP (or Natural Language Processing) is the sub-domain of Artificial Intelligence concerned with the interactions between computers and human languages:
“Alexa, what’s the weather today?”
“Okay Google, how do I say happy birthday in Japanese?”
NLP is the backbone structure on top of which a bulk of commonly used software and applications run. This includes things like personal assistants, grammar checkers (like Grammarly or MS Word), language translation (such as Google Translate), basically everything that deals with the words we use and/or what we say. It is no surprise that NLP is also the core component of search engines like Google, Bing, and the rest, as it is used to analyse the searched query in order to return relevant results. Therefore, with our increasing dependence on these systems (to set alarms, check our spelling, search the web and now even auto-compose email responses), the spreading scope of NLP-driven AI activities and the growing demand for increased optimisation, NLP training is going to get more intensive – and therefore have a greater carbon footprint.
So as Robert Downing Jr. works towards the launch of the Footprint Coalition in April 2020 and other organisations start developing AI-driven systems to help combat climate change, maybe we should take a moment to reflect on whether the very tools we pride might just be a double-edged sword.