Abstract: Generative AI is driving a rapid surge in energy demand, with data-centre electricity consumption projected to grow by more than 160% by 2030. In response, AI companies are increasingly turning to nuclear power, seeking to secure tens of gigawatts of new capacity on timelines fundamentally misaligned with the technical, regulatory, and safety realities of nuclear energy. This talk examines how this mismatch is producing a wave of nuclear “fast-tracking” initiatives that threaten to undermine long-standing nuclear safety, security, and governance norms. We analyse three converging trends that are apparent in some nations: (1) policy efforts to weaken nuclear regulatory standards and reduce regulator independence in the name of urgency and national security; (2) proposals to use generative AI systems to accelerate nuclear licensing, safety analysis, and commissioning; and (3) the promotion of advanced and novel reactor technologies—such as SMRs, AMRs, and fusion—on ambitious timelines that outpace current technical and regulatory readiness. Although these initiatives are not universal, they risk eroding established safety principles, increasing public exposure to radiation, and weakening nuclear safety. At the same time, we distinguish these trajectories from responsible applications of machine learning in nuclear engineering, including predictive maintenance, non-destructive testing (NDT), materials modelling, and sensor anomaly detection. We emphasise integrating ML with existing safety analysis, including hazard and risks analysis considering system-level behaviour and operational constraints. The talk argues that responsible ML use in nuclear systems depends not on accelerating timelines, but on clearly articulated safety justifications that demonstrate how these tools reduce risk and support, rather than substitute for, expert judgment and regulatory review.

Abstract: Almost twenty years have passed since the introduction of the Nakamoto consensus protocol, underpinning the Bitcoin blockchain. Often criticized for its enormous energy consumption, Bitcoin has nonetheless operated continuously since 2008 in an open network. Nakamoto’s consensus protocol challenged what was known about consensus, in particular the impossibility of solving consensus in an asynchronous network in the presence of even a single faulty process, as shown by the FLP result. And yet, Bitcoin worked—and still works—at Internet scale. This raises fundamental questions: was Bitcoin really solving consensus? What network assumptions does it rely on? More recent protocols departed from the original Nakamoto design, adapting Byzantine consensus protocols to open networks while tolerating periods of asynchrony, i.e., adverse network behavior such as congestion or temporary partitions. This significantly reduced energy consumption, but at the cost of reduced openness. In this talk, I outline the historical evolution of blockchain consensus protocols, from Bitcoin to the Ethereum Merge, highlighting their security properties and trade-offs. I focus in particular on the role of network models for these protocols, which underpin their security. I argue that these network models may be overly theoretical, and that it remains unclear whether they accurately capture real networks and adversarial behaviors.