Unsupervised Learning with Jacob Effron Ep 84: OpenAI’s Chief Scientist on Continual Learning Hype, RL Beyond Code, & Future Alignment Directions
159 snips
Apr 9, 2026 
Jakub Pachocki, OpenAI Chief Scientist focused on model capabilities, RL, and alignment. He discusses the rise of coding agents and autonomous research tools. He talks about math and physics as benchmarks, extending reinforcement learning to long-horizon tasks, chain-of-thought monitoring for alignment, and the societal risks of highly automated AI research organizations.
AI Snips
Chapters
Transcript
Episode notes
Research Interns Arrive Before Fully Autonomous Researchers
- Research intern vs automated researcher hinges on task specificity and autonomous runtime rather than raw capability.
- Jakub expects near-term systems that autonomously execute specific technical ideas (e.g., run an eval or prototype an approach) but not open-ended self-directed research yet.
Math Benchmarks Serve As A Clear North Star
- Math and physics act as measurable north stars because success is verifiable and arbitrarily hard to scale.
- Improvements in mathematical reasoning transfer to AI research skills and signal general model intelligence gains useful for applied science.
Prioritize Evals And Context Over Building RL Pipelines
- Collect domain evals and examples and prioritize feeding them into context before replicating full RL pipelines.
- Jakub suggests in-context learning will improve, so using contextual examples may be more effective than building a bespoke RL system now.