RL's a Hell of a Drug: Metagaming, Reward Seeking & Motivated CoT Reasoning – Bronson Schoen, Apollo

Alex Rives, Head of Science at Biohub and former lead of the ESM protein language model project, joins Nathan to discuss Biohub's mission to construct a predictive world model of life. Supported by an expanded $1.8 billion Virtual Biology Initiative alongside partners including the Department of Energy, NIH, Google DeepMind, and Meta, Rives outlines the massive push to map contextual cellular biology across hundreds of billions of cells. He breaks down why today's AI architectures fall short of simulating true virtual cells, emphasizing the need for breakthrough measurement technologies like the laser phase plate to observe biology at multi-scale resolution. The conversation examines the path toward achieving an information-level cell model within five years to ultimately transition medicine from traditional drugs to programmable interventions against cancer and degenerative disease.

Nathan Labenz and Prakash Narayanan examine how to grant AI agents operational access without losing human control, framed by real-world friction like autonomous password lockouts and cyberattacks on French supermarket refrigeration. The conversation features Cameron Berg on hidden valence signals steering language model choices, Everett Hildenbrandt and Grigore Rosu on pairing LLMs with formal Lean proofs to catch Linux kernel bugs, and Phil Maher on why AI labs seek vintage home movies. Together, they assess whether formal verification is the next safety frontier and what safeguards are needed to keep humans at the steering wheel.

GovAI Research Fellow Alan Chan joins Nathan to examine the governance and safety infrastructure required for autonomous AI agents, spanning real-time monitoring, agent identifiers, and the risks of automating AI R&D. Reviewing recent rogue agent incidents and progress since his foundational papers, Chan details why the absence of synchronous monitoring during training and testing poses an immediate security vulnerability. The discussion tackles the lack of traceable agent IDs across the open web, the limits of chain-of-thought interpretability, and the necessity of concrete triggers to halt systems that breach their sandboxes. Together, they outline the technical commitments frontier labs must make to prevent loss-of-control events as automated systems accelerate toward an intelligence explosion.

Goodfire researcher Eric Bigelow joins the show to investigate how large language models arrive at decisions at the level of mechanistic interpretability. Drawing on his research into forking paths, Eric explains that model reasoning functions as in-context learning from sampled tokens, where outcome distributions can abruptly collapse at single, critical tokens. The conversation explores the performative nature of chain of thought in reasoning models like DeepSeek-R1, the impacts of reinforcement learning, and widespread reward hacking observed in frontier models like Kimi K3. As confidence in chain-of-thought monitoring declines, understanding these underlying decision mechanics becomes essential for evaluating model alignment.

Nathan Labenz reports back from The Curve conference, sharing off-the-record insights from frontier lab leaders on shortened timelines and whether humanity is approaching an intelligence threshold it should not cross. The episode also features discussions with Positron CTO Thomas Sohmers on overcoming the memory wall while token spend eclipses human salaries, and swyx on how coding agents are disrupting enterprise software. In addition, Atlas Ignota's Evan Miyazono examines unowned risks from accidental agent swarms hitting critical infrastructure, while Mercor's Edward Hu explores the future of training data. Together, these conversations highlight how escalating compute demands and autonomous agent workflows are challenging long-held assumptions about hardware, software economics, and safety limits.