Interview, Fireside Chat
Kevin Scott, CTO @ Microsoft: An Evaluation of Deepseek and How We Underestimate the Chinese
- Product managers are expected to transition into domain experts to establish feedback loops for agents, while the speaker predicts 95% of net new code will be AI-generated within five years, though strategic authorship will remain human-driven.
- Models will evolve to focus on reasoning rather than factual recall, operating as a diverse ecosystem of specialized agents rather than a single universal model, with user interactions shifting from session-oriented transactions to complex, long-term tasks.
- Agents are anticipated to gain significant memory capabilities over the next year and will increasingly handle asynchronous work while users are inactive, facilitating the delegation of complicated tasks similar to human coworker dynamics.
- Future value will reside primarily in products that connect models to user needs rather than infrastructure alone, driven by the trend that usage follows utility as skepticism regarding agent feasibility diminishes.
- Tool accessibility will remain high in the immediate future despite scaling laws eventually hitting an asymptote with diminishing marginal returns, as high-quality synthetic data and expert feedback become more critical than low-quality web data.
- Quantitative measurement of incremental data value remains a current gap, while inference optimization continues to improve through hardware and software advancements beyond specific launches like DeepSeek R1.
- Developers will demand more choices regarding payment and interaction models, while the future landscape will sustain both open and closed source models similar to the search industry.
- The immediate future includes more asynchronous agent interactions, but scaling laws will eventually hit an asymptote with diminishing marginal returns, making synthetic data and expert feedback more critical than web data in post-training phases.
- Software abstraction will rise to allow prompt engineers to function as programmers, yet computer scientists will remain necessary for solving complex computational problems beyond current human capability.
- AI tools are expected to eliminate technical debt and enable small teams to achieve significant results, while Microsoft and others plan heavy education investments to ensure accessibility for every child.
- AI models are already capable as health diagnosticians surpassing general practitioners, and the speaker anticipates a shift toward using this capability for public good, noting that Microsoft cannot identify every opportunity alone and requires a vibrant startup ecosystem.