We like it or not, large language models have quickly embedded in our lives. And due to their intense energy and water needs, they could also be causing spiral even faster in climate chaos. However, some LLM could be releasing more planet -dating pollution than others, according to a new study.
Consultations made to some models generate up to 50 times more carbon emissions than others, according to a new study published in Communication borders. Unfortunately, and perhaps, as expected, the models that are more precise tend to have the highest energy costs.
It is difficult to estimate how bad are the LLMs for the environment, but some studies have suggested that Chatgpt’s training used up to 30 times more energy than US average uses in a year. What is not known is if some models have more pronounced energy costs than their peers, since they are answering questions.
Researchers from the University of Applied Sciences of Hochschule München in Germany evaluated 14 LLM that vary from 7 to 72 billion parameters, levers and dials that adjust the understanding and generation of languages of a model, in 1,000 reference questions about various subjects.
LLMS converts every word or part of the words into a request into a chain of numbers called Token. Some LLM, particularly reasoning LLM, also inserts «special thought tokens» in the entrance sequence to allow additional internal calculation and reasoning before generating the output. This conversion and the posterior calculations that the LLM performs in the tokens use energy and releases CO2.
The scientists compared the number of tokens generated by each of the models they tested. The reasoning models, on average, created 543.5 tokens of thought per question, while the concise models required only 37.7 tokens per question, according to the study. In the chatgpt world, for example, GPT-3.5 is a concise model, while GPT-4O is a reasoning model.
This reasoning process increases energy needs, the authors found. «The environmental impact of the LLM trained in question is strongly determined by its reasoning approach,» study author Maximilian Dauner, a researcher at the University of Applied Sciences Hochschule München, said in a statement. «We discovered that the models enabled for reasoning produced up to 50 times more CO2 emissions than concise response models.»
The more precise the models were, the more carbon emissions they produced, the study found. The Cogito Reasoning Model, which has 70 billion parameters, reached an accuracy of up to 84.9%, but also produced three times more CO2 emissions than similar size models that generate more concise responses.
«Currently, we see a clear compensation for the sustainability of precision inherent to LLM technologies,» Dauner said. «None of the models that maintained emissions below 500 grams of CO2 equivalent reached a precision greater than 80% to correctly answer the 1,000 questions.» The CO2 equivalent is the unit used to measure the climatic impact of several greenhouse gases.
Another factor was the theme. The questions that required detailed or complex reasoning, for example, abstract algebra or philosophy, led to emissions up to six times higher than simpler, according to the study.
However, there are some warnings. Emissions depend a lot on how local energy networks and the models it examine are structured, so it is not clear how generalizable these findings are. Even so, the authors of the study said they expect people to encourage people to be «selective and reflexive» about the use of LLM.
«Users can significantly reduce emissions by inciting AI to generate concise responses or limit the use of high capacity models to tasks that really require that power,» Dauner said in a statement.
