This article has been republished from The Conversation. Please read the original article.
Generative AI is the hot new technology behind chatbots and image generators. But how hot is the earth?
As an AI researcher, I often worry about the energy costs of building artificial intelligence models. The more powerful the AI, the more energy it requires. What does the emergence of increasingly powerful generative AI models mean for society’s future carbon footprint?
“Generative” refers to the ability of AI algorithms to generate complex data. An alternative is a “discriminative” AI that chooses from a fixed number of choices and he only generates one number. An example of discriminatory output is choosing whether to approve a loan application.
Generative AI can create more complex output such as sentences, paragraphs, images, and even short videos. It has long been used in applications such as smart speakers to generate voice responses or to suggest search queries with autocomplete. However, it only recently acquired human-like language and the ability to generate realistic photographs.
Using more power than ever before
Estimating the exact energy cost of a single AI model is difficult and includes the energy used to manufacture computing equipment, build the model, and use the model in production. In 2019, researchers found that creating a generative AI model of hers, called BERT, with 110 million parameters would consume as much energy as one person making a transcontinental flight back and forth. The number of parameters refers to the size of the model, and generally the larger the model, the higher the proficiency. The researchers estimated that creating the much larger GPT-3, with 175 billion parameters, consumed 1,287 megawatt hours of power and produced 552 tons of carbon dioxide. This is equivalent to driving 123 gasoline-powered passenger cars in one year. This is just to prepare the model for launch before consumers start using it.
Size is not the only predictor of carbon footprint. His open-access BLOOM model, developed by the French BigScience project, is similar in size to GPT-3 but has a much lower carbon footprint, consuming 433 MWh of electricity at 30 tonnes of CO2 equivalent. . Google research shows that for the same size, more efficient model architectures and processors and greener data centers can reduce carbon footprint by a factor of 100-1,000.
Larger models consume more energy during deployment. Data on the carbon footprint of a single generated AI query is limited, but some industry statistics estimate it to be four to five times higher than the carbon footprint of a search engine query. As chatbots and image generators grow in popularity, and as Google and Microsoft incorporate his AI language model into their search engines, the number of queries they receive every day could grow exponentially.
Search AI bot
A few years ago, not many people were using models like BERT and GPT outside of the lab. Things changed on November 30, 2022 when OpenAI released his ChatGPT. According to the latest available data, ChatGPT visits will exceed 1.5 billion in March 2023. Microsoft has incorporated ChatGPT into its search engine, Bing, making it available to everyone on May 4, 2023. If chatbots become as popular as search engines, that energy could actually increase the cost of deploying AI. But AI assistants have many uses beyond just searching, such as writing documents, solving math problems, and creating marketing campaigns.
Another issue is the need to continuously update the AI model. For example, ChatGPT was only trained on his data up to 2021, so we don’t know about what happened after that. The carbon footprint of ChatGPT’s creation has not been made public, but he is probably much higher than GPT-3’s carbon footprint. Energy costs will increase further if knowledge needs to be recreated periodically to update.
One of the advantages is that asking a question to a chatbot gives you more direct information than using a search engine. Instead of getting a page full of links, you’ll get a direct answer, just like you would from a human, if accuracy issues were mitigated. Quick access to information may offset increased energy usage compared to search engines.
Our goal
The future is hard to predict, but large-scale generative AI models are here to stay, and people will probably rely on them more and more for information. For example, if a student needs help solving a math problem, ask a tutor or friend, or consult a textbook. The future will rely on chatbots. The same is true for other expertise, such as legal advice or medical expertise.
No single large AI model will destroy the environment, but if a thousand companies develop slightly different AI bots for different purposes, each used by millions of customers, Energy usage can be an issue. Further research is needed to make generative AI more efficient. The good news is that AI can run on renewable energy. Placing computations where green energy is more abundant, or scheduling computations during times when renewable energy is more , emissions can be reduced by a factor of 30-40.
Finally, social pressure may help push AI models to publish their carbon footprint, as some companies and research institutes are already doing. In the future, consumers may be able to use this information to choose “greener” chatbots.
