The Lytic Threshold

In this talk, Sheikh Abdur Raheem Ali: 1) Discusses rare cases where deployed LLMs have been observed to engage in self-directed behavior which would have led to catastrophic outcomes if the agent escaping containment were equipped with stronger capabilities. 2) Introduces arbitrium, a peptide-based communication method used by certain bacteriophages which release small molecules known as autoinducers to decide coordinated population-level behavior, and in particular lytic and lysogenic pathways, with a quorum-sensing mechanism. 3) Explores landmark results from alignment science which inform our current understanding of LLM biology (<250 malicious documents required to poison training datasets vs <100 viral particles required to produce infection in humans). 4) Analyzes early findings from experiments which attempt to investigate scalable improvements to defenses that monitor the internal activations of production transformer models (such as probes) and demonstrates how these enable targeted interventions on known distributions which are more difficult to achieve with input/output only methods (such as prompted classifiers) in certain cases (such as long context windows or latent reasoning models) ​Event Schedule6:00 to 6:30 - Food and introductions6:30 to 7:30 - Presentation and Q&A7:30 to 9:00 - Open Discussions ​​​​If you can't attend in person, join our live stream starting at 6:30 pm via this link. ​​​​This is part of our weekly AI Safety Thursdays series. Join us in examining questions like:  ​​​​How do we ensure AI systems are aligned with human interests? ​​​​How do we measure and mitigate potential risks from advanced AI systems? ​​​​What does safer AI development look like?