It’s easy to forget that, for every online task that makes our lives easier, there’s a cost in terms of energy used and carbon released into the atmosphere. Dr Amanda Williamson discusses the climate impact of artificial intelligence.
Did you know?
According to a study by researchers at the University of Massachusetts, training a large language model with 1.75 billion parameters can emit up to 284 tonnes of carbon dioxide, which is equivalent to the emissions from five cars over their lifetimes.
Another study by researchers at the Alan Turing Institute suggests that training a model such as GPT4 generates as much carbon dioxide in 90–100 days as 60 average humans would use over a year.
In an analysis conducted by OpenAI, it was determined that the computing power required for training a large AI has been doubling every 3–4 months since 2012.
Questions for discussion
- Do you think the energy cost of building a model such as ChatGPT is justified by the applications people will find for the technology?
- What ways can you think of to keep up with or reduce the carbon cost of AI?
Transcript
Dr Amanda Williamson
What is the climate impact? And it’s huge. These language models, to build them, it was a huge, huge … it came at a huge cost in terms of climate costs. And the computing that we do does come at a cost as well just when we’re using it. But my biggest concern is, if we continue to keep building these models, that’s only going to continue to cause that issue.
But if we use what we’ve got, it’s not so expensive just to use it. It’s that training – that takes months of training with a huge server to build the language models that we’re using at present day.
Acknowledgements
Dr Amanda Williamson, Senior Lecturer in Innovation and Strategy at the University of Waikato and Manager in AI & Data Consultancy at Deloitte.
Video footage courtesy of the University of Waikato. The video is from a recording of the Kaupapa Kōrero panel discussion Bots vs Beings: How Will AI Impact Your Life and Work? held at the University in June 2023.
The full discussion with Professor Mike Duke, Dr Amanda Williamson, Professor Nick Agar and Associate Professor Te Taka Keegan and facilitated by Professor Bryony James can be viewed here.
