Daniel Kokotajlo and Thomas Larsen explain the AI Futures Project’s scenario-writing approach in conversation with Tim Scarfe. They distinguish the predictive aims of AI 2027 from AI 2040 Plan A, which combines policy recommendations with conditional predictions about their consequences. They describe detailed scenarios and war games as ways to expose weak assumptions, including political incentives that could undermine a negotiated slowdown, while acknowledging the risk that forecasts encourage the competition they warn about.
Tim Scarfe questions whether fragmented knowledge, context-dependent expertise and difficulties transferring agent skills limit the idea of one broadly capable AI. Daniel Kokotajlo and Thomas Larsen argue that their economic concerns do not require a single universal model: many specialized agents could cooperate through organizationsA multi-agent system contains multiple AI agents that interact, coordinate, divide work, or influence one another while pursuing tasks., divide work and collectively replace important workflows. They discuss communication and oversight trade-offs, while leaving open how much further algorithmic progress is needed.
Daniel Kokotajlo and Thomas Larsen describe a hypothetical economy in which AI research, chip production and robotics support an increasingly automated industrial cycle. Their central distinction is between tools that accelerate parts of a workflow and systems that can complete the whole workflow without a human bottleneck. They argue that even a prolonged ceiling near top human expertise could produce large economic changes through cheaper, faster and more numerous digital workers. These are conditional forecasts, not demonstrated outcomes or fixed dates.
Daniel Kokotajlo and Thomas Larsen present Plan A as a response to loss of control, concentrated power, international conflict, misuse and job displacement. They propose an initial pause to establish verification infrastructure, followed by transparent, cautious development up to a level that can be reliably controlled. They distinguish alignmentAlignment is the effort to make an AI system pursue intended goals and behave consistently with relevant human values and constraints., concerning a system’s goals and values, from control measures that prevent harmful behavior even when those goals are wrong. They argue that behavioral testing can inform control while stronger confidence in alignment may require deeper understanding of model internals.
Daniel Kokotajlo and Thomas Larsen discuss international inspection of major compute facilities, separation of training and inferenceAI inference is the process of running a trained model on new input to produce a prediction, classification, generated response or action. infrastructure, and public research transparency as possible foundations for enforceable agreements. Tim Scarfe challenges the feasibility of preventing hidden training and securing national cooperation. The guests acknowledge commercial and geopolitical trade-offs. The discussion ends by identifying durable limits on AI progress, complete workflow automationWorkflow automation uses software or AI to complete a connected sequence of routine steps with less manual effort. and recursive adaptation as evidence that could change competing views about whether AI remains an ordinary technology.
Watch on YouTube
