Summary
Machine Learning Street Talk hosts Daniel Kokotajlo (founder of the AI Futures Project and former OpenAI researcher) and Thomas Larsen (lead author of the AI 2040 Plan A report) for an extended conversation on long-range AI forecasting. The episode centers on “Plan A,” a new scenario in which the AI industry voluntarily slows development to pace the frontier, deferring superintelligence to roughly 2040 — a deliberate contrast to the near-term trajectory laid out in the pair’s earlier AI 2027 report, which Kokotajlo says has been tracking closer to reality than he expected at publication.
The discussion examines what human-level AI deployed at cloud scale would actually mean economically: autonomous humanoid robots constructing cities, strip mines run by AI-directed machinery, solar infrastructure expanding across oceans, and compute doubling on sub-annual timescales. The hosts challenge the guests on whether AI agents can genuinely compose and transfer skills across domains, probing the gap between theoretical weight-merging and the practical brittleness researchers see in multi-agent deployments today.
Kokotajlo, who left OpenAI citing concern about the gap between insider knowledge and public information, also addresses alignment questions — specifically, who AI systems are actually being aligned to when stated values diverge from observed model behavior. The episode is essential listening for anyone tracking AI governance, existential risk discourse, and the emerging public debate over voluntary industry pacing mechanisms.
📺 Source: Machine Learning Street Talk · Published September 08, 2026
🏷️ Format: Podcast
