Panel 5

Green AI Summit 2025: Panel 5 - Efficient and Responsible AI: Technology Solutions for a Greener Future

Session Date: April 26, 2025 Location: Boston University Duan Family Center for Computing & Data Sciences

Overview:

Opening the panel discussions on Day 2, this session focused on the technological solutions required to build a future where AI is both efficient and responsible. Moderated by Yasaman Khazaeni from Rue Gilt Groupe, the panel brought together researchers and industry leaders from MIT, Oracle, Meta, Harvard, and the University of Nottingham to highlight advancements in machine learning systems that prioritize energy efficiency, reduced resource requirements, and overall sustainability.

Key Discussion Points:

The conversation explored practical approaches and broader considerations for developing greener AI:

• Defining Green AI: Panelists offered nuanced definitions, emphasizing not just minimizing direct environmental costs (energy, water, carbon) but also ensuring AI is applied towards positive environmental outcomes and avoiding applications that counteract climate goals (like fossil fuel exploration). The concept extended to considering the entire lifecycle and systemic impacts. Paulo Carvao provocatively suggested Green AI might be the "only possible AI" given the unsustainable trajectory of current energy demands.

• Efficient AI Techniques: Speakers detailed technical approaches to reduce AI's footprint. Shreyank Gowda discussed parameter-efficient learning (like adapters and quantization), data-efficient learning (training effectively with less data), and edge deployment (reducing reliance on centralized data centers) as key strategies. These methods aim to lower compute, memory, and energy demands during both training and inference.

• End-to-End Optimization & Open Source: Jay Jackson highlighted the challenge within companies of achieving true end-to-end optimization across the entire AI stack (hardware, data, algorithms, compilers) due to siloed expertise. Kate Saenko stressed the critical role of open source – sharing not just model weights but data recipes, training code, and energy/hardware details – to enable better research, reproducibility, comparison, and the adoption of efficient techniques. The current lack of transparency and incentives for companies to share this information was noted as a major barrier.

• Incentives and Policy: The discussion touched upon motivating companies to adopt greener practices. While inherent economic incentives exist due to rising energy costs and the unsustainability of current trends, policy and regulation were seen as crucial. Examples included setting environmental constraints to drive engineering innovation (similar to past environmental regulations) and the need for independent auditing rather than self-regulation.

• AI for Climate Action (Applications): Priya Donti provided examples of AI actively accelerating climate action, such as monitoring emissions (Climate Trace), optimizing power grids (forecasting renewables), improving building efficiency, accelerating scientific discovery (e.g., battery materials), and potentially enhancing transportation systems. She emphasized the need for close collaboration between AI experts and domain-specific problem owners.

Panelists:

• Yasaman Khazaeni (Moderator): VP, Data Science & Machine Learning, Rue Gilt Groupe

• Paulo Carvao: Senior Fellow, Mossavar-Rahmani Center for Business and Government, Harvard Kennedy School

• Jay Jackson: Vice President, Artificial Intelligence & Machine Learning, Oracle

• Priya Donti: Assistant Professor, MIT EECS & LIDS; Co-founder and Chair, Climate Change AI

• Kate Saenko: Professor, Department of Computer Science at Boston University; Director, Computer Vision and Learning Group and member of the IVC Group

• Stephan Klinger: Senior Expert Counsel, LGP Law, Vienna

• Shreyank N Gowda (Online): Assistant Professor, the University of Nottingham

This panel showcased the technical ingenuity being applied to make AI itself more sustainable, while also highlighting the systemic, policy, and transparency challenges that must be addressed to ensure these solutions are widely adopted and contribute effectively to a greener future.