Kevin Scott, CTO @ Microsoft: An Evaluation of Deepseek and How We Underestimate the Chinese

Microsoft CTO Kevin Scott argues frontier AI models already outperform average GPs, and DeepSeek surprised the public but not serious practitioners.

• Microsoft had models more capable than DeepSeek R1 that they chose not to launch; the public reaction surprised Scott.
• Scott believes frontier models already outperform average GPs as health diagnosticians, and the world needs to deploy this immediately.
• 95% of net new code will be AI-generated within five years; human authorship shifts to intent and architecture, not line-by-line writing.
• Agents’ core missing feature is durable memory — without it they remain transactional and session-bound, limiting delegation of complex tasks.
• Microsoft Research launched a dedicated lab ~1 year ago with a single mission: eliminate tech debt at scale using AI tools.
• High-quality data plus expert human feedback can be amplified into training signal; undifferentiated web tokens are far less valuable than commonly assumed.
• Scott rejects the one-agent-for-everything model; expects many domain-expert agents, with product managers becoming deep domain specialists who tune agent feedback loops.
• The gap between what frontier models can do and what they are actually being used for is wider now than it was two years ago.

2025-03-31 · Watch on YouTube