Anthropic Researchers on RL+LLMs, Interpretability, and AGI Timelines
Sholto Douglas and Trenton Bricken of Anthropic join Dwarkesh Patel to discuss the state of reinforcement learning for language models and what it means for AGI. They argue that RL with verifiable rewards (RLVR) has finally proven it can produce expert-level reliability in domains like competitive programming and math, and predict that software engineering agents capable of hours of autonomous work will arrive by end of 2025. They also describe how they are still under-investing in RL compute relative to pre-training, but expect this to change rapidly as labs scale up. A major portion of the conversation covers mechanistic interpretability progress, including sparse autoencoders, the Golden Gate Bridge feature, and circuits work showing that larger models form more abstract, cross-lingual representations. They describe an interpretability agent that can win Anthropic's auditing game, plus alignment concerns including alignment faking, emergent misalignment from narrow fine-tuning, sandbagging, and how models might develop situational awareness during evaluation. They debate whether current models reason genuinely, with circuit evidence suggesting yes on certain math tasks but no on hard computations like cosine.
Original sources
Podcast Dwarkesh Podcast
Episode:Is RL + LLMs enough for AGI? — Sholto Douglas & Trenton Bricken ↗
YouTube Dwarkesh Podcast
Video:Is RL + LLMs enough for AGI? — Sholto Douglas & Trenton Bricken ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.