This landmark 200+ page report by over 100 researchers from Stanford's Human-Centered AI Institute coined the term "foundation models" and provided the most comprehensive mapping of their opportunities and risks:
Key Insights
- Homogenization risk: When many applications build on a few foundation models, any flaw in the base model propagates to all downstream applications
- Emergent capabilities: Foundation models exhibit behaviors not explicitly trained for — both beneficial (reasoning) and concerning (deception)
- Concentration of power: Only a few organizations can afford to train foundation models, creating unprecedented concentration in AI capabilities
- Societal impact: Effects on labor, education, healthcare, law, and the environment are interconnected and require holistic assessment
Source
Bommasani, R. et al. (2021). On the Opportunities and Risks of Foundation Models. Stanford HAI. arXiv:2108.07258.