Second Annual Risk Conference 2024
What it is about
Fernando Prieto, partner at Management Solutions, explained at the Second Annual Risk Conference (September 26, 2024) how to make artificial intelligence sustainable, both in its environmental impact and in its ethical use.
The 25 slides start from the context (AI's electricity consumption and the position of regulators and agencies), set out the areas of action for sustainable AI (hardware, infrastructure and software), present use cases (an energy-efficient AI system and a framework for building ethical and fair algorithms) and close with the risks of generative AI and six final reflections.
Objectives
- Size the environmental impact of artificial intelligence and the regulatory response.
- Present concrete practices for AI that is more efficient in energy and resource use.
- Explain how to build ethical algorithms and control the risks of generative AI.
Key takeaways
- In 2023, two large technology companies each consumed 24 TWh of electricity, more than the total consumption of the Dominican Republic.
- The European AI regulation (article 40) promotes energy-responsible development of models, and bodies such as the OECD stress the need to measure the environmental impact of AI.
- Sustainable AI combines efficient hardware and renewable energy in data centers, "green" algorithms (efficient architectures, model compression) and measurement of the CO2 footprint.
- One case presented decouples the structural and quantitative knowledge of a neural network to drastically reduce training and incremental retraining time while keeping predictive power.
- The risks of generative AI cited include privacy, interpretability, hallucinations, bias, regulatory non-compliance, intellectual property and excessive dependence.
Connect with our networks