Summary

  • Sasha Luccioni, AI and climate lead at Hugging Face, has said that data scientists should be focusing on making AI smarter, not simply increasing computational power, according to a VentureBeat report.
  • Task-specific or distilled models can require as little as 20%-30% of the energy of a general-purpose model and are often more efficient for specific tasks.
  • Open-source models, which do not have to be trained from scratch, are more efficient; Hugging Face has launched an “AI Energy Score” to rate the efficiency of different models, as an “Energy Star” equivalence.
  • Default settings increase costs and compute requirements as models do more work than is required; adjusting batch sizes and memory usage can lead to significant increases in efficiency.
  • Changing the mindset that “more compute is better” can lead to better results through smarter architectures and better-curated data, rather than simply bigger GPU clusters.

By Taryn Plumb

Original Article