The MAD Podcast with Matt Turck Sonnet 4.5 & the AI Plateau Myth — Sholto Douglas (Anthropic)
504 snips
Oct 2, 2025 
Sholto Douglas, AI researcher at Anthropic and former Google engineer, delves into the innovations of Claude Sonnet 4.5, claiming we're mere years away from AI matching human capabilities. He explains how reinforcement learning has suddenly made a breakthrough and how agents can maintain coherence during long coding sessions. Sholto also discusses the cultural differences across major AI labs, Anthropic's focused approach to coding, and the profound implications of AI's upcoming exponential progress, especially in economics and robotics.
AI Snips
Chapters
Transcript
Episode notes
Meter Evals Track Time-Based Progress
- Meter evals correlate model performance with time horizons and show rapid doubling of task time capability.
- As time horizon improves, models can be left running overnight and return useful, complex outputs.
Prioritize Taste And Persistent Memory
- Focus on fixing 'taste' and global context issues to reduce frequent human interventions.
- Improve memory so agents stop repeatedly rediscovering codebase-specific facts.
Many Small Wins, Then Big Progress
- Recent jumps arise from many incremental improvements across the stack plus added compute, not a single miracle.
- Cumulative engineering and plentiful compute produce smooth, law-like gains in capabilities.