At ANSAM, we have chosen to assume this responsibility through our 2024 sustainability report and our commitment to B Corp certification. But we also wanted to go further by giving our experts a say, through a series of testimonials in the form of 3 key questions to shed concrete light on the challenges of the digital transition.
Guillaume Van Lierde, Director of Novatix, shares his views on the environmental impact of artificial intelligence and the need to promote a more sober AI.
AI is particularly energy-intensive during the model training phase. How can we reconcile innovation and responsibility in such a competitive market?
1. What are the major environmental impacts of AI development today?
The development of artificial intelligence has a very real environmental impact. The training phase of models, particularly those used for language or vision processing, is particularly energy-intensive: it mobilizes hundreds, even thousands of hours of computation on powerful servers, consuming large quantities of electricity and material resources. Once the models have been trained, their day-to-day use during a query is much less energy-intensive. But on a global scale, the accumulation of billions of interactions ends up consuming a considerable amount of energy. The problem is that impact measurement tools are not yet standardized, which complicates awareness and regulation.
2. What is "sober AI" and why is it important?
Sober AI is first and foremost a responsible approach to the development of artificial intelligence. It involves designing lighter models, calibrated to real needs, rather than systematically aiming for maximum performance. It also means reducing training iterations, pooling resources and, more broadly, questioning the systematic use of energy-intensive models for tasks that could be accomplished otherwise. Promoting sober AI also means making users aware of the impact of their technological choices.
3. What are the challenges involved in implementing this sobriety in a highly competitive environment?
The main challenge is the constant pressure for rapid innovation. In a market as competitive as AI, it's tempting to go faster and harder, without always taking the time to optimize resources. There is also a lack of clear standards for measuring the environmental impact of AI models, making it difficult to compare and make informed decisions. Finally, there's a real job to be done in helping customers understand these issues: explaining to them why a more sober model may be a better choice in the long term, even if its gross efficiency seems slightly lower.