Sustainability in the Age of AI

I'm the father of two awesome kiddos, ages 12 and 14. They're both smart as a whip and are growing up to be kind, empathetic, and wise. Both children are curious by nature, and they are growing up in a very interesting era marked by an incredibly disruptive technology that will change the way they learn. 

The kids both have taken an interest in my work as they know that it involves AI, and they consistently have shared their concerns about the environmental impact that AI is having on our planet. With massive amounts of power, water, and other resources being used to train and run LLMs, they're right to be concerned. 

So, where's my head at with this complex issue?

Technology Trade-offs

First and foremost, it's important to acknowledge the facts – LLMs require a lot of compute to create and use. Between training, inference, and data requirements, AI systems require massive amounts of electricity, water, and other natural resources. Sustainability encompasses more than just environmental impact; it also touches on social justice and other, more human impacts.

That said, it's important to remember that every major technology shift has been environmentally impactful. Cloud computing, for example, has resulted in an explosion of data center capacity, creating similar demands on natural and human resources.

Now, the question we have to ask ourselves is "Is this all worth it?" 

In other words, is the amount of energy required to train and use LLMs creating enough benefit to the world to justify the cost? That's a complicated question, but as a technologist, I think the answer is "yes, eventually." AI is already creating so many practical benefits after an extremely short time, and I expect that will accelerate over the next decade.

Accountability and Action

It is our responsibility as technologists to understand the impacts of our innovations and to do what we can to ensure that they're built on a foundation of accountability.

The good news is that many of the biggest names in cloud and AI are making efforts to become increasingly sustainable. AWS has made significant investments in its own data centers, with a "Power Use Effectiveness" (PUE) of 1.15. PUE is a measure of efficient energy usage in data centers, and while it's not perfect, it's one of several useful metrics to guide more sustainable data centers. The "perfect" PUE is 1.0, where the ratios that balance energy with efficiency are in complete harmony.

AWS also provides its customers with tools and best practices to help them reduce the carbon footprint of their workloads. This is a noble pursuit, but it also has positive business impacts, as more efficient workloads running on IaaS are generally less expensive.

Sustainable AI

If you're interested in understanding a bit more about the path to sustainable AI, I highly recommend taking a look at The Gradient, which is a non-profit organization with a mission to facilitate healthy discussion about AI. They've been around since the very early days of generative AI, and back in 2021, they published The Imperative for Sustainable AI Systems, an op-ed by Abhishek Gupta, the founder of the Montreal AI Ethics Institute. While it is an opinion piece, it's exceptionally well-informed and worth a read.

In the piece, Gupta advocates for several key approaches to achieving "Sustainable AI." Notably, he advocates for:

• Elevating smaller models – smaller LLMs can be trained more quickly, operated more efficiently, and still be incredibly effective. Too often, AI workloads use models that are much larger than the use case demands. By taking care to select the smallest model effective for a task, we can be more efficient and reduce the impact of our workloads.
• Alternate deployment strategies – innovation in hardware, such as purpose-built silicon like AWS' Trainium processors, can drive massive efficiencies both in training and in execution. In addition, continued work on federated learning can decentralize training, pushing the compute to the most carbon-friendly regions.
• Carbon efficiency and awareness – being intentional about optimizing workloads for the realities of the grid can make a large impact by understanding the carbon footprint required to train and use LLMs in different regions, or even different times of day.

Wrapping Up

When I speak to my children about the impact that AI can have on the world, I first discuss the incredible utility of the technology. But, I also share that there are many ways that we can hold businesses accountable for their impact on the environment, and that there is a global community of researchers, scientists, and advocates working on creating more efficient silicon, driving the adoption of clean energy, and pushing for a future of sustainable AI. The future can be bright provided we're willing to light the way.