Practical AI Green AI 🌲
Mar 2, 2021
Roy Schwartz, a senior lecturer at the Hebrew University of Jerusalem, and Jesse Dodge, a postdoctoral researcher at the Allen Institute for AI, discuss the urgent need for Green AI. They reveal the surprising environmental costs of AI research and advocate for more efficient, inclusive practices. The conversation highlights the disparities in research contributions and emphasizes the importance of transparency and sustainability. Innovative solutions like smaller, efficient models are explored, showcasing their potential to transform AI workflows and reduce carbon footprints.
AI Snips
Chapters
Transcript
Episode notes
Shifting Computational Demands
- In 2013, Jesse Dodge could run experiments on a used laptop.
- By 2019, large models required cloud instances and extensive resources, a trend that continues.
The Problem with Scaling
- Scaling AI models improves performance but increases the carbon footprint and research inequality.
- Training a single model can emit as much carbon as five cars, highlighting the environmental impact.
Reporting for Efficiency
- Report model evaluations at different training stages to enable comparisons with lower budgets.
- This promotes competition and drives down costs by focusing on performance-efficiency trade-offs.