Red and Green AI

Illustration generated using DALL·E 2.

In episode 6: Red vs. Green AI, we discussed the environmental impact of AI. A Red AI is one that requires massive amounts of computational resources. While such systems can often produce highly accurate results, they also have a very large carbon footprint. Demands on increasing accuracy and performance result in AI systems and AI research becoming ever more resource-demanding over time, there-by also excluding researchers, companies, organisations and countries who do not have access or can afford such large amounts of resources.

A Green AI, meanwhile, is one where you strive to make gains without increasing computational resource demands, or possibly even decreasing resource demands – though potentially at the expense of accuracy. Measuring and reporting the resource requirements is a key characteristic of a Green AI, and the environmental impact is a central consideration from the very start.

How much is too much?

Can anything that is as energy intensive as AI ever truly be green? And the answer is, probably not entirely. A more appropriate question might be how red or how green is the AI in question? Now the answer to that depends very much on what the AI is doing, and how much energy it needs to do it. The most energy intensive part of AI is normally in the beginning stages, when training the underlying machine learning models. When it comes to training, the pertinent question that should be asked is – how accurate does the model actually need to be?’ A degree of error of two percent for example might be acceptable, and so therefore it is no longer necessary to continue development once this point has been reached.

Setting the limits

The danger is, researchers continue to develop their models simply for the sake of it, consuming greater and greater amounts of energy in the process. So if there is ever to be something that could ever be considered ‘green’ AI, it is, at least in part, down to the researchers being willing to set and keep to certain boundaries. Of course, it can be argued that the energy source used can play a part in helping make the AI somewhat redder or greener. Surely solar, wind or hydro power, all themselves considered to be green, must have an influence? But then again, there is the question of just how green solar and wind really are, but that is a debate for another time.

No ‘U’ turn

The technology road we are travelling is certainly not one that will lead us to a greener future as long as we continue to exploit current energy resources in ever vaster quantities. But a 180 degree U-turn is also not something that is going to happen either, and one has to suppose shouldn’t happen. The answer to the question of how to green AI and technology in general currently has no answer. Until a completely green energy solution becomes available, if ever, we can only try to restrain ourselves in the uses, requirements and energy consumption of the AI we wish to exploit in order to prolong the life expectancy of our planet.

Written by Ian Bowie