EAG Talks

Join the organizers for a brief welcome and a group photo of attendees, followed by a short talk from Ben West.

Watch on Youtube: https://www.youtube.com/watch?v=XqwE9RyxxQs&t=4s

In this session, Ryan Ritterson will use examples of his and others’ work at Gryphon to tell stories about how to effectively influence public policy. Based on Gryphon’s data-driven approach, he’ll provide key lessons learned and takeaways for others interested in influencing policy.

Watch on Youtube: https://www.youtube.com/watch?v=0ohVtc5Vnps

This talk examines whether advanced AIs that perform well in training will be doing so in order to gain power later — a behavior Joe Carlsmith calls "scheming" (also often called "deceptive alignment"). This talk gives an overview of his recent report on the topic, available on arXiv here: https://arxiv.org/abs/2311.08379.

Watch on Youtube: https://www.youtube.com/watch?v=AxUTiGS6BHM

Should you be able to order smallpox DNA in the mail? Biosecurity professionals have argued for almost 20 years that synthesis companies should screen orders so that pathogen and toxin sequences are only sent to people with a real scientific use for them. Now, it seems like fears of AI-engineered pandemics may spur governments to make screening mandatory.

Watch on Youtube: https://www.youtube.com/watch?v=q-dVdRe3oco

Media coverage of catastrophic risk is going mainstream; our panel of media professionals discuss the ways it's going right vs going wrong, and what actions they are taking to focus the conversation on well-reasoned risk models and effective interventions.

Watch on Youtube: https://www.youtube.com/watch?v=33XDXk6wBgg

Hayley Clatterbuck will summarize key findings from Rethink Priorities' ""Causes and uncertainty: Rethinking value in expectation"" (CURVE) project, which evaluated the cost-effectiveness of existential risk mitigation projects under both standard expected utility maximization and risk aversion. Persistence of effect and future growth trajectory are the strongest contributors to expected utility. Different risk models often deliver different recommendations about which existential risk projects to pursue, but there are some actions for which these models agree. Clatterbuck will use these findings to draw lessons for cause prioritization within the global catastrophic risk space.

The five years since Toby Ord wrote The Precipice have seen dramatic changes to the landscape of existential risk. Ord will explore the biggest changes to the biggest risks, showing how new developments have upended key assumptions and pushed some of these risks into new phases. And we’ll see how the world has woken up to the very idea of existential risk, with it becoming an explicit priority and talking point on the global stage.

Watch on Youtube: https://www.youtube.com/watch?v=vQ3ml6wcsn4

If an AI system learned a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? That's the question that Evan and his coauthors at Anthropic sought to answer in their work on ""Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training"", which Evan will be discussing.

Evan Hubinger leads the new Alignment Stress-Testing team at Anthropic, which is tasked with red-teaming Anthropic's internal alignment techniques and evaluations. Prior to joining Anthropic, Evan was a Research Fellow at the Machine Intelligence Research Institute and worked on a variety of theoretical alignment work, including ""Risks from Learned Optimization in Advanced Machine Learning Systems"". Evan will be talking about the Anthropic Alignment Stress-Testing team's first paper, ""Sleeper Agents: Building Deceptive LLMs that Persist Through Safety Training"".

Watch on Youtube: https://www.youtube.com/watch?v=BgfT0AcosHw

The final session of the conference will include some closing words, followed by a talk and fireside chat with Dan Zimmer.

Watch on Youtube: https://www.youtube.com/watch?v=E6Fe1iPCgfU

Pandemic security aims to safeguard the future of civilisation from exponentially spreading biological threats. In this talk, Kevin will outline two distinct scenarios–"Wildfire" and "Stealth"–by which pandemic-causing pathogens could cause societal collapse. He will then explain the ‘Delay, Detect, Defend’ plan to prevent such pandemics, including the key technological programmes his team oversees to mitigate pandemic risk: a DNA synthesis screening system that prevents malicious actors from synthesizing and releasing pandemic-causing pathogens; a pathogen-agnostic wastewater biosurveillance system for early detection of novel pathogens; AI/bio capability evaluations and technical risk mitigation strategies; and pandemic-proof PPE.