Unexpected: AI training indirectly releases CO2 which is harmful to the environment

The process of AI training seems to be harmless but is indirectly affecting the environment and climate due to the operation of CO ₂ emissions from the process of electricity consumption.

AI or artificial intelligence is gradually becoming the driving force of the industry when it helps people detect cancer, drive and predict disasters, etc. But we all don't know that, the process of digging at AI also causes consequences for the environment.

The process of digging at AI also causes consequences for the environment.

According to a recent study by US scientists, the popular AI training model may emit more than 283.9 tons of CO2. This number is more than five times the average amount of emissions that a car in the United States emits into the environment during its life cycle, including car production.

Almost all AI researchers were surprised by the potential effects of AI on the environment.

Helping machines understand language is behavior that is harmful to the environment?

According to Interesting Engineering , the study explores natural language processing or thinking language programming (NLP). The NLP research community has contributed quite a lot to technology in the prime time, including machine translation and sentence completion.

However, in exchange for improvements in the past on many applications, AI training understands human language requiring huge amounts of data to be collected from the Internet. Especially the conversion of these data sets into a training program requires a huge amount of computer power.

Where does that energy come from?

Of course electricity is taken from power plants and to generate electricity, people need resources such as fossil fuels, nuclear, . and that's just one of the few causes of gas change. post and environmental pollution.

The four models used by researchers to determine the impact of AI training on the environment include: Transformer, ELMo, BERT and GPT-2 . To find out the level of CO ₂ emissions of these models, the research team conducted each model, based on a single GPU for up to 1 day. This activity is intended to measure the power consumption of AI.

They then used the number of hours of training the models to calculate the total electricity that the entire training process could consume. This number will then be converted into CO ₂ emissions. The results show that the cost of training AI and its impact on the environment is directly proportional to the size of the AI ​​training model.

The cost to develop AI is very big, but it also affects the environment.

The final and most expensive step is to improve the accuracy of the AI. However, this is also the step that causes the greatest impact on the environment.

With the final step removal model, CO ₂ emissions are estimated to be only about 635kg CO ₂ . This is equivalent to the CO ₂ of a flight around the United States that applies to one person.

What's worse is that when the researchers claim, these are only modest numbers because researchers only train AI models with minimal power consumption. Most practical AI training programs will need a greater amount of power than that.

The impact from AI training is huge. That's why the team proposed moving AI research out of the academy and applying it only internally. It can be seen that the cost of developing AI is very large because it not only consumes resources but also indirectly pollutes the environment.

Although the AI ​​is promoting the development of many industries such as defense, health, education. But it is clear that monitoring and monitoring of AI development will become more and more difficult if resources are gradually exhausted.