The potential of artificial intelligence is immense – but its equally large energy consumption must be hampered, by asking shorter questions in a way of carrying out, said a UNESCO study unveiled on Tuesday.
A combination of shorter queries and the use of more specific models could reduce IA energy consumption up to 90% without sacrificing performance, UNESCO said in a report published to mark AI for the good world summit in Geneva.
Openai CEO Sam Altman recently revealed that each request sent to its popular generative application AI Chatgpt consumes on average 0.34 Wh electricity, which is between 10 and 70 times a Google search.
With Chatgpt receiving approximately one billion requests per day, this amounts to 310 GWh per year, equivalent to the annual electricity consumption of three million people in Ethiopia, for example.
In addition, UNESCO has calculated that AI’s energy demand is doubled every 100 days as generative AI tools become integrated into daily life.
“The exponential growth of the computing power necessary to manage these models exerts increasing pressure on global energy systems, water resources and critical minerals, which raises concerns concerning environmental sustainability, equitable access and competition from limited resources,” warned the UNESCO report.
However, he was able to obtain a reduction of almost 90% of electricity consumption by reducing the duration of his request, or invites, as well as using a smaller AI, without lower performance.
Many models of AI such as chatgpt are models for general use designed to respond to a wide variety of subjects, which means that they have to scrutinize an immense volume of information to formulate and assess the responses.
The use of smaller specialized AI models offers major electricity reductions necessary to produce an answer.
The same goes for the cutting of cutting prompts from 300 to 150 words.
Being already aware of the energy problem, the technology giants all offer miniature versions with fewer parameters of their respective large language models.
For example, Google sells Gemma, Microsoft A Phi-3 and Openai A GPT-4O Mini. French IA companies have done the same; For example, Mistral IA introduced his model, minister.
