Interview, Fireside Chat
Microsoft CTO Kevin Scott on How Far Scaling Laws Will Extend | Training Data
- Scaling models is expected to reduce costs and fragility for currently brittle tasks while enabling more complex operations.
- Microsoft plans to expand access to powerful AI alongside OpenAI and increase the number of Copilots with daily market traction for non-trivial cognitive work.
- Demand is projected to grow with capability expansion, though its specific shape may shift as frontier model development remains resource-intensive.
- Uncertainty exists regarding the number of frontier models needed if capabilities become roughly equivalent, and the next major capability breakthrough is anticipated within a couple of years.
- New business models for data retrieval may emerge via subscriptions, revenue sharing, licensing, or advertising, while auctions could evolve to value referrals and agent interactions.
- Developers relying exclusively on current frontier models without architectural flexibility risk obsolescence when next-generation models arrive.
- Achieving the final 2% of reliability for autonomous agents will likely require combining foundation model improvements with domain-specific software optimizations.
- The field has not yet reached diminishing marginal returns on scale, though the continuation of the current scaling trend for six years or more is uncertain.
- Technology will likely shift from zero-sum societal problems to non-zero-sum games by creating abundance, specifically targeting healthcare, environmental, and educational sectors.
- AI applications are expected to alleviate rural healthcare strain, assist scientists in finding carbon capture catalysts, improve transportation safety, and support education, while Microsoft focuses on assistive rather than substitutive cognitive tools.
- Without technological intervention, the healthcare system is forecast to become increasingly strained over time.