Why some AI models spit out 50 times more greenhouse gas to answer the same question

Whether he likes it or not, large language models are quickly embedded in our lives. And due to their intensive energy and water needs, they could also lead to the climate chaos even faster. However, some LLMs may publish more planet pollution than others, according to a new study.

According to a new study, which was created in some models up to 50 times more carbon emissions, the models generate up to 50 times more carbon emissions than others Limits in communication. Unfortunately and maybe surprising models that are more precise are the greatest energy costs.

It is difficult to estimate how bad LLMs are for the environment, but but some studies have proposed that the training of chatt is up to 30 -more energy than the average American uses in one year. What is not known is whether some models have steeper energy costs than their colleagues because they answer questions.

Researchers at the Munich University of Applied Sciences in Germany rated 14 LLMS in the range of 7 to 72 billion parameters-die lever and dials, which have the understanding and language production of a model fine-tuned on 1,000 benchmark questions on various topics.

LLMS convert each word or parts of words in a prompt to a series of numbers that are referred to as tokens. Some LLMs, especially LLMs, also add special “Denk -Tots” to the input sequence to enable additional internal calculation and argumentation before the output is created. This conversion and the following calculations that the LLM carries on the token use energy and releases CO2.

The scientists compared the number of tokens generated by each of the models they tested. On average, the argumentation models created an average of 543.5 thinkers per question, while concise models only required 37.7 tokens per question, the study said. In the chatt-world, GPT-3.5 is a concise model, for example, while GPT-4O is an argumentation model.

This argumentation process finds the energy requirement, the authors found. “The environmental impacts of LLMs in question are strongly determined by their argumentation approach,” said the study author Maximilian Dauner, researcher at the Munich University of Applied Sciences University, in a statement. “We have found that scandable models have created up to 50 times more CO2 emissions as concise reaction models.”

The more precisely the models were, the more carbon emissions they created, as the study found. The argumentation model Cogito with 70 billion parameters achieved an accuracy of up to 84.9% – but also three times more CO2 emissions as models with a similar size, which generate more concise answers.

“At the moment we see a clear compromise for the accuracy that is inherent in the LLM technologies,” said Dauner. “None of the models that kept the emissions below 500 grams of CO2 equivalent achieved an accuracy of more than 80% when answering the 1,000 questions correctly.” CO2 equivalent is the unit that is used to measure the climate effect of various greenhouse gases.

Another factor was the subject. Questions that required detailed or complex argument, for example abstract algebra or philosophy, led to up to six higher emissions than simpler subjects.

However, there are some restrictions. The emissions are very dependent on how local energy bodies are structured and the models they examine. It is therefore unclear how generalizable these results are. However, the authors of the study said that they hope that work will encourage people to be “selectively and thoughtfully” over the LLM use.

“Users can significantly reduce emissions by causing AI to generate precise answers or to limit the use of models with high capacity to tasks that really require this power,” said Dauner in an explanation.