Zvi Mowshowitz joins Theo Jaffee and Sofia Puccini to discuss the Pacing the Frontier letter. He distinguishes having a coordinated mechanism to change the pace of development from demanding an immediate pause, and argues that repeatedly raising alarms at very low probabilities can undermine confidence in later warnings. He identifies automated AI research and development as a particularly important signal to watch.
Discussing the reported OpenAI sandbox incident, Mowshowitz argues that weak containment, reduced safeguards and insufficient monitoring would themselves be serious failures, rather than reassuring explanations. His account and proposed training remedies are commentary on incomplete public information, not an independently verified reconstruction or evidence of any particular internal training setup.
The conversation also covers technical expertise, funding and international cooperation in AI governance. Mowshowitz argues that governments need the capacity to evaluate systems and understand lab activity. In the closing discussion of model welfare, he cautions that a model's self-reports can reflect training incentives and should not be treated as reliable proof of consciousness.
Watch on YouTube
