The environmental cost of genAI
In the last few years, many of us have started to see the benefits of using genAI in day-to-day tasks. But we’ve also been asked to reckon with the enormous environmental cost.
Reporting has highlighted that these popular AI technologies have a significant environmental impact through high energy consumption, carbon emissions, and water use.
AI-focused data centres are increasing in size to accommodate larger and larger models and growing demand for AI services. According to a recent report from the International Energy Agency, a hyperscale, AI-focused data centre can consume as much electricity annually as 100,000 households.
Large AI models require many energy-intensive calculations, necessitating liquid cooling systems. Large data centers can consume up to 5 million gallons per day, equivalent to the water use of a town populated by 10,000 to 50,000 people.
A greener alternative to large language models
So, what’s the alternative?
Popular genAI technologies all operate from ‘large language models’; advanced AI systems built on deep neural networks designed to process, understand and generate human-like outputs.
‘Small language models’ offer similar outputs but with a comparably smaller scope.
While large language models need massive data sets and hundreds of billions (or trillions!) of parameters (learnings) to do their work, a small language model will typically require just one billion.
In this way, small language models require less memory, processing and storage, which are essential for reducing the overall resource footprint of AI systems. In some cases, small language models can even require 90 per cent less energy to achieve similar results to their larger peers.
Think of it like using a jumbo jet to travel to your local shops as opposed to your car. For most of the tasks we want to use genAI to complete, there is no need to use a large language model. It’s overkill; a small language model would suffice.
Wider benefits of small language models
Not only do small language models have smaller energy requirements; they also cost far less to build and maintain. An important consideration for businesses that want to develop their own genAI technologies.
Small language models can run on devices that users or companies already own, making them easier to deploy locally, improving accessibility and reducing the need for extensive infrastructure support.
This means small language models also offer better security.
For example, using genAI in a defence context, you don’t want your national security to depend on a third-party product or be vulnerable to cyberattack.
Instead, small language models can be deployed on a local device, which allows your data to stay with you, and there’s no need for internet access.
Barriers to using small language models
So, what’s stopping us from using these greener, cheaper, safer generative AI alternatives already?
Small language models do currently face some limitations. Because they are smaller, their performance is different. They can’t help you with complex tasks and are better used for more specific topics or simple tasks.
Many of the things we look to large language models for assistance with day-to-day can be accomplished with a properly trained small language model, like asking for a recipe, booking a hotel room or writing a report.
But so far, much more money has poured into the development of large language models, and big tech companies are now heavily invested in ensuring that investment pays off. Arguably, it’s in their business interests to maintain the performance gap their large, paid, models currently offer over smaller, open access, alternatives.
This has left small language models somewhat overlooked. As a result, it’s mostly researchers and start-ups who are developing this space right now, which means investment is lower and progress is therefore slower.
Deakin researchers are improving AI technology
More research is required before small language models can be widely adopted into genAI technology.
That’s where our work at Deakin’s Applied Artificial Intelligence Initiative comes in.
One of our insights is around using multiple small AI models that collaborate rather than relying on one large model. Our early work suggests groups of smaller specialised models can share reasoning and solve complex tasks together.
A drone, for instance, could run a small vision model to spot hazards alongside a separate reasoning model to decide its next move, both operating locally on a chip with no internet required.
Another part of our work looks at training AI for tasks without a single correct answer. Current approaches need clear right-or-wrong answers to learn. We’re developing approaches that train models on open-ended problems: writing a caption, building a story from images, or finding a creative solution where no single answer is correct.
We’re also solving AI’s forgetting problem. Current models lose old skills when they learn new ones. We’re building systems that accumulate knowledge, like a prosthetic limb that learns your walk, your run, and adapts to injury over years, never losing a movement pattern it has ever known.
We hope this work can help build a safer and more sustainable approach to generative AI.
*This article was originally published in Deakin Research News
About the author
Dr Hung Le is an Australian Research Council DECRA Fellow and a Senior Lecturer at Deakin University. He is a senior member of Deakin Applied Artificial Intelligence Initiative (A2I2), where he currently supervises several PhD students in research areas focused on machine learning (ML) and reinforcement learning (RL). He specialises in deep learning and is dedicated to pioneering new agents equipped with artificial neural memory.
Stories worth sharing
The text of this article is licensed under the Creative Commons Attribution (CC BY) 4.0 International license. We’d love for you to share it, so feel free. Images, videos, graphics and logos are not covered by the CC BY license and may not be used without permission from Deakin or the respective copyright holder. For more information on how to share or reuse this content, please contact researchcomms@deakin.edu.au.
Share