ChatGPT and other generative AI technologies are revolutionizing the way we interact with machines. However, there are implications to the amount of energy that goes into creating and running these powerful AI models. A single query for a generative AI model such as BERT and GPT-3 can consume the energy of a round-trip transcontinental flight and generate large amounts of carbon dioxide equivalent for every query.
The BigScience project in France developed the open-access BLOOM model, which is comparable in size to GPT-3 but does not generate the same amount of energy consumption and carbon emissions. Google found that by using more efficient model architectures, processor, and data centers, their specific AI models can reduce energy consumption by a factor of 100 to 1,000.
To make matters worse, AI models need to be continually updated to keep up with advances in AI technology. ChatGPT, for example, had 1.5 billion visits in March 2023, and its carbon footprint may increase as more people use it.
The good news is that AI models can be run on renewable energy. By bringing the computation to where green energy is more abundant, or scheduling computation for times of day when renewable energy is more available, emissions can be greatly reduced. Additionally, transparency in publishing carbon footprints could help incentivize companies and research labs to reduce their energy use.
OpenAI is the company behind ChatGPT, as well as other generative AI modules. Founded in 2015, OpenAI has been at the forefront of groundbreaking research in AI, working with leading experts in the field. Co-founder Greg Brockman was the Chief Technology Officer of OpenAI until he stepped down in late 2020.
As AI technology continues to advance and become more widely used, it is important to consider the long-term environmental implications. By taking steps to reduce energy consumption where possible, and holding companies and research labs accountable for their carbon footprints, we can ensure that AI technology is doing more good than harm.
