top of page

AI Safety Unconference @ conf

The purpose of the AI Safety Unconference is to foster connection and shared understanding between researchers interested in AI safety. It is a series of events, hosted alongside main AI/ML conferences. It welcomes both newly interested and established researchers. It features talks, moderated discussions, one-on-ones, free-form interactions, and participant-driven activities.

~85 participants, independents and from various affiliations: OpenAI, DeepMind, Cambridge, MIRI, Mila, Cornell, Anthropic, MIT, Columbia, Stanford, U Toronto, Waterloo, Cooperative AI, ...

Lighting talks:

  • Haydn Belfield: What standard-setting in EU + US might mean for AI safety

  • Esben Kran: Hackathons in AI safety research

  • Franziska Boenisch: Privacy attacks against federated learning

  • Aaron Tucker: Bandits with Costly Reward Observations

  • Lewis Hammond: Cooperative AI

  • Adam Dziedzic: Stealing and defending self-supervised models

  • David Lindner: Active Learning for Reward Modelling

  • Lauro Langosco di Langosco: An empirical demonstration of deceptive alignment

  • Zhijing Jin: Causally aligning language models

 

Facilitated discussions (1h each):

  • Haydn Belfield: AI governance

  • Adam Dziedzic: Is this model mine? On stealing and defending machine learning models

  • Lewis Hammond: Cooperative AI

  • Lauro Langosco di Langosco: Deceptive alignment

Testimonials

  • This was a fascinating event that was helpful for keeping up with the cutting edge of the field, and for launching collaborations. – Haydn Belfield

  • The AI safety unconference was very useful to meet and talk with the AI safety researchers at NeurIPS. — Esben Kran

  • It was very reassuring to hear that diverse perspectives on AI risk are being studied seriously, including criticism of the AI safety community. — Arvind Raghavan

~50 participants, independents and from various affiliations: OpenAI, DeepMind, Cambridge, MIRI, Mila, ...

Participant-driven discussions, multiple lighting talks, ...

~50 participants from various affiliations: UC Berkeley, Vector Institute, Mila, OpenAI, DeepMind, Oxford, CHAI, Mcgill, NYU, Partnership on AI, etc

Talks from Adam Gleave, Jan Leike, David Krueger, Dan Hendrycks, Aaron Tucker, Victoria Krakovna

Testimonials

> A great way to meet the best people in the area and propel daring ideas forward. — Stuart Armstrong
> The event was a great place to meet others with shared research interests. I particularly enjoyed the small discussion groups that exposed me to new perspectives. — Adam Gleave

bottom of page