You have only instructed a generative AI tool to design a logo to your friend who is opening a new coffee shop. The result pops up in seconds. It feels like magic, right? However, behind that one prompt comes a thirsty guzzling engine. Artificial intelligence is becoming an environmentally huge dilemma, but very quietly so. What is the solution? How to reconcile this astounding technological advancement and the necessity of sustainability?
The Magic of the Convenient Numbers
Let’s talk scale. The electricity used to train a single large model, such as that used by GPT-3, was estimated to serve more than 1000 average U.S. homes in a year. Moreover, it had an immense footprint in water. According to a recent study, even the mere act of talking to an AI tool such as ChatGPT uses the amount of water equivalent of a 500ml bottle per 10-50 prompts. This is because data centers require enormous cooling requirements. These numbers aren’t static. They are exploding.
- Every several months, computational demand is doubling. This is a trend that goes crazy compared to the Law of Moore.
- A training run of one model can result in an emission of more than 500 tons of CO2. That is a 300 round trip flights between New York and San Francisco.
The ugly side of the AI boom is the environmental cost.
The Reason Your Artificial Intelligence Question is Leaving the Tap Running
Why is it so inefficient in the process? At that, the strategy of brute force prevails at the moment. We pour additional information and more computational resources on the problem. Billions of parameters are there in modern models. All of them need to be calculated continuously. Imagine it in the following way: when a model is asked to write a haiku, it is like using a gigantic engine that is set to compress multi-layered paperwork works. It’s overkill.
Additionally, it is not only the initial training that is draining the energy. It’s the inference. It implies that each and every interaction with an AI tool, whether by you or a billion other users, provides power to it. This puts an irreversible high energy liability on our grids.
Green Artificial Intelligence Awakening: A Strategy Change
Thank God, the industry is coming to itself. The high price of compute is an excellent incentive. Therefore, now a competition over efficiency is in effect. The idea is to do a lot by doing less, and develop AI devices that do not require the planet to empty its pockets.
Key strategies are emerging. To begin with, we have smarter model architectures, such as Mixture-of-Experts. These models activate a limited set of their network in order to carry out a task. Second, there exists a massive trend towards specialized hardware. Such companies as Google and Amazon are developing AI-specific chips. They provide higher calculations per watt of power.
A Dilemma of the Developers: My Code vs. The Climate
This dilemma was something that I had to encounter as a machine learning engineer. My team was developing a new recommendation model. We were crazed on the issue of accuracy improvement by 0.5%. Weeks and weeks went by running thousands of experiments. One day, a colleague gave out carbon footprint calculator as a training model. The result was a gut punch. Our quest of a marginal gain had already emitted a number of tons of CO2. We had been totally ignorant of the ecological cost of our labour. It is a typical tale of the tech laboratories nowadays.
- We are optimizing accuracy and speed, and hardly ever joules/prediction. – AI Researcher in a large laboratory.
- The point of motivation is to release a larger model, not a more efficient one. That needs to change.” – A Lead AI Scientist.
Case Study: As Artificial Intelligence Comes to Life
Take an example of a real world situation. One of the financial institutions installed a strong AI model to identify fraudulent transactions 24/7. It was successful and millions of dollars in fraud were caught. Its continued use in a data center in a hot environment, however, necessitated immense cooling. The local electricity supply that still used natural gas was unable to sustain the demand. The artificial intelligence that helps keep financial assets unharmed was giving a contribution unwillingly to the very climate instability that creates economic shocks. This is the paradox in action.
Competent in Corporate Greenwashing: What Is Critical
A number of technology companies are offering to be carbon-neutral by 2030. They buy renewable energy certificates. But is this enough? It is the issue of transparency. As it stands, companies do not have a standard defining whether to disclose the footprint of their particular AI tools. AI requires some kind of a nutritional label. It would indicate energy used, water used, and carbon used per 1000 inferences. This would enable businesses and users to make sustainable decisions.
What can you do? Ask questions. When selecting an AI vendor, ask them how efficient they are. Fund more narrow and special models. The future of AI does not have to be a decision of smart or sustainable. It must be both.
But the brightest system is that that can be self-sufficient and does not destroy its surrounding in the process. Let’s build that one.
