Whether it likes or not, big language models are quickly embedded in our lives. And due to their intense energy and water needs, they can also cause us to spirals rapidly in climate chaos. Some LLMs, however, may release more planetary-warming pollution than others, get a new study.

According to a new study published in the query made for some models, 50 times more carbon emissions are generated than others. Limitations in communication, Unfortunately, and perhaps indefinitely, models that are more accurate have the biggest energy cost.

It is difficult to guess how bad LLMs are for the environment, but Some study It is suggested that training chatgpt uses 30 times more energy than average American uses in a year. What is not known whether some models have energy cost compared to their peers because they are answering questions.

Researchers at the University of Munchen University of Applied Sciences in Germany evaluated from 7 to 72 billion parameters to 14 LLMs- liver and dial that fix the understanding and language production of a model- about 1,000 benchmark questions about different topics.

LLMS converts each word or parts of words into a signal in a string of numbers called a token. Some LLMs, especially logical LLMs, also include special “thinking tokens” in the input sequence to allow additional internal calculations and arguments before generating outputs. This conversion and subsequent calculations that perform on LLM tokens, use energy and release co2.

Scientists compared the number of tokens generated by each model he tested. Reasoning model, on average, created 543.5 thinking tokens per question, while the brief model required only 37.7 tokens per question, found in the study. In the world of chat, for example, GPT-3.5 is a brief model, while GPT-4o is a logic model.

This logic process increases energy needs, the authors found. “The environmental influence of questioning the trained LLM is strongly determined by his logic approach,” the study author Maximilian Dunar, a researcher from the Hochchule Munchen University of Applied Sciences said in a statement. “We found that the logical-competent models produce 50 times more CO2 emissions than the concise response model.”

The more accurate models were, the more carbon emissions they produced, found in the study. The logic model Cogito, which has 70 billion parameters, reached 84.9% accuracy – but produced three times more CO2 emissions than the same size model that produces more brief answers.

“Currently, we see an apparent accuracy-sustainability trade-off in LLM technologies,” said Dunar. “None of the models with emissions equal to 500 grams of CO2 receives more than 80% accuracy when the correct answer to 1,000 questions.” Co2 equivalent is a unit used to measure the climate effect of various greenhouse gases.

Another factor was the subject matter. According to the study, for example, detailed or complex arguments are required, for example abstract algebra or philosophy, reached six times more emissions than more directly subjects.

Although there are some caves. The emission is very dependent on how the local energy grids are structured and the model you examine, so it is not clear how common these conclusions are. Nevertheless, the study authors said that they hope that the work will encourage people to be “selective and thoughtful” about LLM use.

Dunar said in a statement, “Users can reduce emissions by induction of AI to generate brief answers or to limit the use of high -capacity models, which actually requires that power,” Dunar said in a statement.