Sofia Puccini asks Nenad Tomasev how scientific institutions can keep pace with agentic AI. Nenad Tomasev argues that more generated hypotheses will not automatically produce more scientific progress: physical laboratories, domain experts and the surrounding research ecosystem also need investment to validate or disprove those ideas.
Nenad Tomasev distinguishes near-term use of AI from the longer-term problem of training the next generation of experts. The discussion considers cognitive friction, automation bias and the danger of either accepting model judgments too readily or overriding useful predictions. Nenad Tomasev links faster progress in mathematics and coding to their comparatively clear verification signals.
Sofia Puccini and Nenad Tomasev examine value lock-in, personalization and the feedback between users and models. Nenad Tomasev presents the alignment research as conceptual modeling rather than a complete deployment solution, emphasizing changing social norms, sycophancy and the need for environments and benchmarks that measure desirable behavior.
Nenad Tomasev closes by contrasting unstructured agent swarms with governed teams. Human institutions offer useful organizational ideas, but agent management is not identical to managing people; safety, security and robustness constraints must be designed for the technical systems themselves.
Watch on YouTube
