Energy to train AI tools, wasted?

Energy used to make and provide online services is an important consideration for many reasons. Production and delivery of online services require energy, and that energy has a direct impact on the environment. The energy used to create and provide online services often comes from burning fossil fuels, such as coal, natural gas, and oil. This burning releases carbon dioxide (CO2) and other pollutants into the atmosphere, contributing to global warming. Burning fossil fuels also releases other harmful pollutants, such as particulate matter, sulfur dioxide, and nitrogen oxides, contributing to air pollution and can cause serious health problems. Increased energy consumption also has a direct effect on our environment. As energy consumption increases, so does the demand for resources such as coal, natural gas, and oil. This can lead to the destruction of ecosystems and habitats, as well as the displacement of communities. Additionally, burning these resources to produce energy contributes to climate change, causing a shift in weather patterns, rising sea levels, and an increase in extreme weather events. The energy used to provide online services also has an impact on the cost of providing these services. The more energy used to power the servers and networks, the more expensive the services become. Additionally, higher energy costs can lead to higher consumer prices, as companies must pass on the extra costs to their customers. Finally, suppose energy used to provide online services is generated from non-renewable sources, such as coal and oil. In that case, it means that the energy used to power these online services will eventually run out, which could negatively impact the availability of these services in the future. Overall, it is essential to consider the energy used to make and provide online services. Burning fossil fuels to power these services contributes to air pollution and global warming while also increasing costs. Additionally, the use of non-renewable resources to generate energy could lead to a decrease in the availability of these services in the future. 

A prominent example of online services is AI chatbots that can provide the user with answers to almost any topic. Other than a search engine that only finds matches of the search text in the indexed documents, AI chatbots can compose new information by drawing connections between the vast amount of information they have been trained with. AI programs like ChatGPT are a highly relevant development because they significantly improve the user experience and enable people from all domains to access sophisticated AI technology. Open AI programs make AI more accessible, allowing developers to share and collaborate on AI models. It also helps reduce development costs and makes integrating AI into existing applications easier. By allowing developers to access and build upon existing models, they can create new and innovative applications that can benefit everyone. Developing AI models helps automate tedious tasks and reduce the time spent on manual labor. By using AI models, businesses can automate mundane tasks and improve their workflow. AI models can also help to improve customer support and increase customer satisfaction. AI models are also important for predicting future trends and predicting customer behavior.

But, despite the fact that users of AI often get free access or a generous free trial, developing and training an AI model does not come for free when we consider the energy budget. The Carbon footprint of training ChatGPT has been estimated to be 1287 MWh [1], in addition to running the services. Are 1287 MWh a number to be concerned with? Probably yes. Is it a number so high that we immediately need to banish AI training for the sake of the environment? I don't think so.

When relating 1287 MWh to a single person, it is a lot. It would mean driving an average European car on fossil fuels for 4,5 Mio km. That is enough to travel the whole road network of the United States or equivalent to the carbon footprint of a flight passenger going form London to New York 320 times.

Nevertheless, ChatGPT has more than one user. In fact, it is one of the fastest-growing online platforms in the world, with around 100 Million users at the time of writing. Dividing the development costs by the users, it amounts to 0.01287 kWh or roughly 1% of the energy required to print a book. 

In other words, if users can utilize the AI system to automate mundane tasks and improve their workflow, the energy spent on creating the AI is probably well-invested. Many usages are recreational, and sometimes the AI provides more fiction than facts, but so is the case with books.

However, we need to keep our eyes open on two issues:

• The operational cost of running the system: "Cost" would mean here energy cost as well as financial cost. If the system does not work here efficiently, we could end up in a much higher energy waste than 1287 MWh
• Further developments in training new AIs: competitors might train their own AIs, no matter the (energy) cost. And models are expected to grow in complexity and capabilities, probably also significantly raising the energy required for training a single model.

So let's keep an eye on further developments.

[1] Patterson, D., Gonzalez, J., Hölzle, U., Le, Q., Liang, C., Munguia, L.-M., … Dean, J. (4 2022). The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink. Computer, 55, 18–28. Retrieved from http://arxiv.org/abs/2204.05149