The FAIK Files Power Struggles
19 snips
Aug 22, 2025 Discover Google's surprising new insights on the environmental impact of AI models. Dive into the promising world of Hierarchical Learning Models, touted as the future of AI architecture. Hear about a baffling hack of a smart home via a Google Calendar invite. Plus, explore the art of deception with a captivating look into emotional manipulation and the 'long con'. This episode is packed with tech intrigue and practical tips that keep you engaged!
AI Snips
Chapters
Books
Transcript
Episode notes
Real-World AI Energy And Water Use
- Google published production-scale measurements showing a Gemini text prompt uses about 0.24 Wh and ~0.26 mL of water on median.
- Their study includes chip utilization, idle machines, CPU/RAM, and data-center overhead, making estimates much lower than earlier theoretical figures.
Energy Use Is Multi-Component
- Google's breakdown shows TPUs account for 58% of electricity while CPU/memory, backup equipment, and overhead make up the rest.
- The company also highlighted large renewable power purchases, lowering emissions per electricity unit versus average grids.
Avoid Unnecessary AI Calls
- Consider the trade-offs before adding unnecessary AI features like auto-rendering images for every app action.
- Avoid gratuitous AI calls that waste compute and energy when the feature adds no user value.
