AI summary by TwinMind
AI More Energy Efficient than Humans, New Study Finds
Sabine Hossenfelder
6m 33s
Research Paper Analysis - AI vs Human Carbon Emissions
- Study compared CO2 emissions from AI vs humans producing text/images
- AI methodology: counted training phase + operational energy
- Training emissions divided by expected lifetime queries
- Operational energy (e.g., MidJourney unicorn generation)
- Human methodology: used typical carbon intensity of living in different countries
- Calculated emissions for time required to write/paint
Key Findings
- Typical US American needs 100-1,000x more CO2 per page written or image produced vs AI
- Exact numbers depend on AI model and human location
- GPT energetically wasteful compared to MidJourney for image generation
- Outsourcing to India reduces carbon footprint
- Hardware production/recycling contributions orders of magnitude smaller than training/operation
Energy Consumption Comparison
- Human brain: ~20 watts power consumption
- Supercomputer: ~20 megawatts (1 million times more)
- AI image generation: ~3 watts (equivalent to charging phone)
- Mark Twain productivity benchmark: 300 words/hour
- Speaker's actual productivity: 400 words in 10 minutes
Critical Analysis Points
- Humans produce CO2 emissions regardless of writing/painting
- AI emissions are additional, not replacement
- Human carbon intensity includes entire body maintenance overhead
- More relevant comparison: writing human vs non-writing human
- Difference likely unmeasurable
Broader Implications
- Energy demand tied to societal progress
- Problem: current energy sources not environmentally sustainable
- AI purpose: free humans for other tasks
- Future scenario: AI-written scripts, AI-generated footage, AI viewers
- Question: what's the point of fully automated content creation?
Babble Language Learning Sponsorship
- Language learning app promotion
- Different languages activate different brain connections
- 60% discount offer with 20-day money back guarantee
