It becomes increasingly common to evaluate frontier models for dangerous capabilities. But what should developers do if sufficiently dangerous capabilities are actually discovered? In our new paper, we explore one possible response: coordinated pausing. arxiv.org/abs/2310.00374

Frontier AI developers should be prepared to respond to a ange of issues that arise in models that are already deployed. But what should this look like? I explore this question with my colleagues Shaun K.E. Ee and Zoe Williams at @_IAPS_ in our new paper. iaps.ai/research/deplo…

Check out this paper to better understand the incentives and scale of AI chip smuggling into China, concrete estimates for what this might look like in future, and what steps Congress and BIS can take to reduce it.

Check out this new paper for concrete recs on how actors like NIST, MITRE, CISA and frontier labs can effectively identify gaps in their risk management of frontier AI - even as those risks rapidly evolve.

International safety dialouges could be a great way to reduce AI risks - this paper gets down to the nuts and bolts of how to make these events as successful at that goal as possible.

External scrutiny is a key tool in governing AI. Check out this new paper for six things it needs to go well, and how to make them happen.

A fantastic explainer of the factors relevant to the US export controls on China regarding AI chips, and what they mean in practice - if you're looking to understand this area, I recommend it. Thanks to Christoper Phenicie and @_IAPS_ Erich Grunewald for authoring it.

Fantastic new report examining the role security features on AI chips could play in mitigating national security risks and protecting user privacy, with concrete recs for policymakers.

Excited for our first AI policy fellowship! Applications are open now, take a look if you're interested in moving into AI policy.

If you're involved in running multilateral events on topics related to advanced AI, you might find this new issue brief interesting - it discusses how to create agendas in ways that minimize risks of participants leaking sensitive information.

New paper: Future AI systems may be capable of enabling offensive cyber operations, lowering the barrier to entry for designing and synthesizing bioweapons, and other high-consequence applications. If these capabilities are discovered, who should know first, and how? More in 🧵