Weekly insights from leading experts in trustworthy & secure AI

Julian Stastny presented "Plan B: Training LLMs to fail less severely" at FAR.Labs, as part of the FAR.AI Seminar series.

Fabien Roger presented "Eliciting the capabilities of scheming LLMs" at FAR.Labs, as part of the FAR.AI Seminar series.

Hidenori Tanaka (Harvard) presented "3 Mechanisms Underlying Emergent Abilities in Generative Models," linking AI creativity and safety with insights from physics, neuroscience, and psychology.

How can strange bedfellows shape the future of AI policy? Andrew Freedman, founder of fathom.org, shares lessons from cannabis regulation to build unlikely coalitions for AI governance.

Evan Hubinger dives into “Deceptive Instrumental Alignment,” revealing how AI might appear aligned during training while secretly harboring harmful intentions for deployment.

Peter Barnett describes the potential of hardware enabled mechanisms to provide verification and confidence to international coordination schemes for AI.

Alex Pan discusses strategies for modeling and mitigating immediate deployment risks associated with large language models (LLMs).

Paul Riechers and Adam Shai, founders of Simplex, discuss their approach to developing a principled science of AI cognitive capabilities and associated risks.

Shaowei Lin discusses formal methods in AI-assisted code specification and synthesis, emphasizing formal verification as a strategy to protect sociotechnical systems from human-led and AI-enabled attacks.

Can Go AIs be adversarially robust? Tom Tseng explores defending Go bots against adversarial attacks, using adversarial training and architecture change.

Kris Brown Brown discusses the role of category theory in creating interpretable and adaptable AI systems, aiming for enhanced safety and stakeholder alignment.

Pioneering AI researcher Yoshua Bengio explores approaches to designing AI systems that are both aligned with human values and provably safe.

Ethan Perez shows how large language models can hide harmful behaviors, even under safety training.