Subscribe to Our Bi-Weekly AI Newsletter

AI's Impact on Climate Change

AI impacts climate change and carbon emissions in a few different ways.

On the one hand, AI models can be used to reduce carbon emissions by helping people and businesses operate more efficiently. On the other, AI models produce carbon when they are trained.

Those are the first-order effects. To get to the optimization, you have to pass through training. That is, AI must produce carbon emissions in order to reduce them. But the potential reductions are enormous.

How AI and Machine Learning Can Help Fight Climate Change

The core value of AI is that is helps us make smarter decisions, which is the nature of intelligence itself.

Let’s take weather predictions and the power grid. Solar power is clean, but it needs to be replaced when clouds block sunlight. Gas generators are kept idling in order to replace solar power, but keep those generators on idle, instead of turning them off, produces significant carbon emissions. Better weather forecasts would enable the grid to turn the generators off entirely during periods of consistent sunshine. AI is capable of producing better forecasts.

Making he physical operations of large businesses more efficient is another way that AI can help fight climate change. One good example is DeepMind’s success reducing Google data center cooling bills by 40%. Deep reinforcement learning has achieved similar breakthroughs optimizing other industrial operations and business processes.

[WIP]

Training Large AI Models Produces Carbon

Training large AI models uses energy and produces carbon emissions.

For example, OpenAI’s natural language model GPT-3 ranges in size from 125 billion to 175 billion parameters. The largest model needed 3.14E23 FLOPS of computing to be trained. Analysts have estimated that amount of compute would require between 355 GPU-years and 665 GPU-years, depending on the chip.

The same arguments that are being made against the training of large models could be made against any compute intensive process, ranging from mining crypto-currency to collecting, moving and processing large amounts of data, such as logs from IoT. So in a sense, arguments against AI training are arguments against digitization, the public cloud and the semiconductor industry, all of which are predicated on massively increasing the amount of compute we throw at humanity’s problems.

Chris Nicholson

Chris Nicholson is the CEO of Pathmind. He previously led communications and recruiting at the Sequoia-backed robo-advisor, FutureAdvisor, which was acquired by BlackRock. In a prior life, Chris spent a decade reporting on tech and finance for The New York Times, Businessweek and Bloomberg, among others.


Newsletter

A bi-weekly digest of AI use cases in the news.