What’s the problem?
Everybody loved ChatGPT, but not so many people are aware of the tremendous damage from the point of view of the environment. Indeed, to train the model extensive use of computer was required which means electricity, which means CO2 to generate this electricity (502 Tons C02). We are talking of about the same footprint of 600 flights between New York and San Francisco, according to the AI Index report from Stanford University.
And this is only taking into account ChatGPT, other models or AWS services are not better. AI is largely dependent on the infrastructure that supports it, including data centers, cloud networks, and edge devices, all of which use up a lot of energy and resources. Given the exponential growth of data and the requirement for ever-more-powerful computer resources with the creation of ever-larger neural networks or large language models (LLM), the energy consumption of AI is a serious concern. AI’s energy needs are substantial in terms of carbon emissions as well as the resources needed to create and maintain the essential infrastructure.
Numerous studies have examined the carbon footprint of AI, emphasizing the severe negative environmental effects of this technology. “Energy and Policy Considerations for Deep Learning in NLP” Strubell et al. was a study that was presented at ACL 2019. It claims that training a single deep-learning model can produce up to 284,000 kg of CO2, which is comparable to the lifetime energy use of five cars. Other studies have similarly emphasized how AI affects the environment, especially in terms of energy use and carbon emissions. Unfortunately, apart the alarming analysis. The only conclusion was to provide easy to-use APIs implementing more efficient alternatives to brute-force grid search for hyperparameter tuning, as Bayesian hyperparameter techniques, which are never or rarely used for NLP training model.
I don’t want to sound alarming, but recent studies like the aforementioned from University of Massachusetts Amherst stated that “training a single AI model can emit as much carbon as five cars in their lifetimes.” LLM training requires a huge amount of energy, and it appears that a single ChatGPT prompt uses 100 times as much energy as a single Google search. Sam Altman estimates that each prompt uses “probably single-digits cents” of energy, which translates to 0.3 kWh each request as opposed to 0.0003 kWh per Google search.
Those numbers sound small, but we are talking about the ratio 1 to 1000 for 1 ChatGPT search and 1000 Google search in terms of energy costs and environmental impacts. Imagine this multiplied for all the users around the world.
But carbon footprint related to electricity is not the only problem. To train LLM we need more and more powerful cards [thank you NVIDIA], which have to be updated continuously. What does it happen to the old ones?
Some are sold second-hand but the majority becomes e-waste. E-waste having dangerous pollutant is incenerate.
I have seen many buzzword headlines like “AI will solve climate change”, “AI for good will save the planet”. I am sorry to open your eyes. Most of these initiative are just hype, and actually the increased usage of AI is contributing producing CO2 and harming the planet.
What can we do?
The AI sector MUST undergo a revolution while decreasing its carbon footprint thanks to a number of innovative projects, energy-efficient hardware designs, and renewable energy options.
Let’s examine some of the most pertinent strategies for addressing the problems associated with AI’s carbon footprint.
AI companies should enter Net-zero economy. Using as clean as possible energy to maintain their servers on and to train models. It seems Amazon and other big techs are already pledging to reduce to zero their emissions.
Energy efficient hardware should be adopted. This includes developing software that is energy-efficient and creating chips specifically for AI. The energy needed to run AI applications can be greatly decreased with this method, lowering the environmental effect. For example programmable resistors, and neuromorphic hardware.
Use strategies for training: “sparse” neural networks, which reduce the number of connections between neurons and minimize the computational load. Transfer learning and federated learning can reduce the carbon foot print as well, as we are just training part of the model and reuse previous ones. There are already initiatives to implement solutions as there are challenges to propose models in medical imaging related to energy efficiency.
Better models should be used. Using data sets that are 1/10,000 the size of the huge language models (like a portion of language a child can have compared to an adult), several researchers are attempting to develop language models. The BabyLM Challenge aims to teach a language model, based on a dataset of the words young children are exposed to, how to pick up linguistic nuances from scratch the way a human does. Young children hear between 2,000 and 7,000 words annually; for the BabyLM Challenge, the dataset’s maximum word count is 100,000, or about the quantity of words a 13-year-old will have heard. A smaller model uses less energy since it is easier to train and requires less time and resources.
More energy efficient cooling system should be used. Conventional cooling techniques like air conditioning sometimes fail to adequately cool data centers. While Microsoft is experimenting underwater cooling system, a company by the name of Lonestar plans to construct a few tiny data centers on the moon. Lonestar data centers would benefit from the plentiful solar energy and be less vulnerable to sabotage and natural calamities.
Sustainable AI is more than simply a trendy term; it refers to a long-term strategy for creating AI systems that can satisfy current requirements without jeopardizing the ability of future generations to satisfy their own. It is crucial to adopt a long-term view that considers the social, economic, and environmental effects of our activities as we continue to rely on AI to solve complex problems and spur innovation.
“Sustainable” means long-term. To have long-term strategies we need a fundamental change in how we view innovation, one that prioritizes community-driven problem-solving over specialized technological advancements. Giorgio Parisi, after winning the Nobel Prize, commented that to stop the climate crisis we have to stop economy. Unfortunately, this is an extreme approach but we need to find a way.
We need to collaborate with experts from a variety of disciplines, including environmental science, social science, engineering, and computer science, to develop AI systems that are truly sustainable from an environmental, social, and technological point of view for our future generations. This partnership may make it possible to design sustainable AI solutions that have a thorough understanding of the intricate social and environmental systems they are meant to solve.
Clearly right now the craze is to increase AI, which is literally in the opposite direction of creating sustainable society. People need to at least be aware of this.