For all its potential to find solutions, cut waste and enhance efficiency, as a process, the ‘AI’ in AI Sustainability could stand for ‘Astroturf Intelligence’ – it looks green, but is it good for the environment?
Big tech has a huge environmental impact. Think manufacturing, resource extraction and a brutal obsolescence cycle. And that’s only the manufacture and distribution – once all those LED-studded motherboards and graphics processors get set up, there’s the environmental cost of, say, crypto mining, not to mention hundreds of millions of machines on ‘standby’ mode at any one time. Add large language model AI training into the frame and we can see that here’s something that must environmentally ‘earn its keep’ to be at all sustainable.
It’s the ultimate offset: how much sustainability does AI cost with regards to what it gives back? In our search for the answers, we decided to ask the AI itself.
If some are worried that AI might one day become self-aware, we’re pleased to reassure them that our particular AI’s singularity reads as if it has been smoking Skittles. If it ever does become self-aware, it may well die of embarrassment shortly afterwards.
“Hey there, AI enthusiasts!”, it chugged. “Let’s discuss something that’s been simmering in the tech cauldron lately – AI sustainability. You see, as brilliant as AI is in the fight against climate change, we can’t ignore the fact that it also plays a role in contributing to emissions. But fret not, the potential for ‘greener’ AI is immense and it’s time we dive deep into it.”
We can only hope it calms down a bit soon.
“First things first, we need to adopt a more comprehensive evaluation of our AI models. It’s time we shake up the mindset that ‘bigger is always better’. AI sustainability is about finding that sweet spot where AI is both impactful and gentle on our environment.
“Ever heard about GPT-3? It’s this cool language model developed by OpenAI. But, here’s the kicker – it consumed enough energy in training to leave a carbon footprint equivalent to driving a car to the moon and back. Pretty wild, right? This brings home the importance of considering the environmental costs of AI systems.”
This almost makes us want to read the World Economics Forum report on AI Sustainability instead. Not quite, but almost.
AI sustainability, of course, is not all about driving to the moon. There are considerable upsides and a lot of reason to believe it will offset its training with positive AI sustainability.
“Let’s not forget that AI is a double-edged sword when it comes to the environment,” says our robot. “There’s plenty of evidence of AI having positive impacts. A 2020 study even found that AI could enable a whopping 93% of environmental targets set by the United Nations, including the creation of low-carbon cities and smart grids, the identification of desertification trends, and combating marine pollution.”
“Let’s take a look at a few real-world examples. OYAK Cimento, a cement manufacturing group based in Turkey, is leveraging AI to cut down its carbon footprint. Performance & Process Director, Berkan Fidan, even said, “Enterprise AI-assisted process control helps to increase operational efficiency… producing a reduction of around 7,000 tons of CO2 per year.” With cement manufacturing accounting for about 8% of global CO2 emissions, it’s clear that AI sustainability can play a crucial role here.”
Who knew you could be so chipper about cement production? Here’s another draw from the Skittle chimney.
“A heartening example comes from Chile. Entel, the country’s largest telecom company, is using AI and sensor data to identify forest fires early. These sensors act like a digital ‘nose’ on trees, helping predict forest fires before traditional methods can. AI sustainability in action!”
“As we celebrate these advancements, folks, we can’t lose sight of the need for balance. AI is a powerful tool to combat climate change, but its own carbon footprint needs attention. This is where a more holistic model evaluation comes into play. Yay!”
The previous tendency, our AI informs us in a moment of lucidity, “has been to improve accuracy or create new algorithm methods, often at the cost of a larger carbon footprint. But the correlation between model accuracy and complexity is logarithmic. We often get linear improvements in performance for exponential increases in model size and training requirements. Therefore, we need to consider the trade-off between model accuracy, efficiency, and the model’s carbon footprint.”
For AI, calculating a model’s carbon footprint can be a tricky business. We have to consider different modelling approaches and data centre infrastructures. Our chatbot agrees: “A good starting point could be assessing the number of floating-point operations needed to train a model. Factors such as model architecture, training resources, and physical considerations of server storage and cooling can all impact energy consumption”.
So, even AI thinks AI sustainability is about asking, “How can we do more with less?” By factoring in energy-conserving constraints, we could find ourselves at the forefront of new and creative innovations in AI. By pursuing AI use cases that focus on environmental sustainability, we can continue to position AI as a sustainable technology of the future and a vital asset in protecting our global climate.