Guaranteed Safe AI Seminars 2024 review

[Orpheus Lummis]

Dear Guaranteed Safe AI enjoyers,

Thanks for having participated or being interested in the seminar series!

2024 in review:

• The series started in April, and as of December, it grew to ~230 subscribers. We had ~490 RSVPs. The recordings had ~76 hours of watch time and ~900 views.
• Each session had a solid presentation and insightful community discussion.
• We had the following sessions:
• Compact Proofs of Model Performance via Mechanistic Interpretability – Louis Jaburi
• Bayesian oracles and safety bounds – Yoshua Bengio
• Constructability: Designing plain-coded AI systems – Charbel-Raphaël Ségerie & Épiphanie Gédéon
• Proving safety for narrow AI outputs – Evan Miyazono
• Gaia: Distributed planetary-scale AI safety – Rafael Kaufmann
• Provable AI Safety – Steve Omohundro
• Synthesizing Gatekeepers for Safe Reinforcement Learning by Justice Sefas
• Verifying Global Properties of Neural Networks by Roman Soletskyi
• There was a 2 months hiatus because the main organizer had an accident.
• It started as Provably Safe AI Seminars but expanded in scope and rebranded to GS AI Seminars.
• We thank Long-Term Future Fund for supporting the series for a 6 months period.

Onward to 2025:

• We intend to continue the monthly series.
• We aim to launch a living GS AI bibliography at guaranteedsafe.ai.
• We are planning a moderated debate involving critics and proponents, leading to a LessWrong dialogue post.
• We will add occasional social mixer events.

Call to action:

• As of this post, we do not have funding for the upcoming year. To support the project:
• Donate, comment, or vote on our Manifund proposal (https://manifund.org/projects/horizon-events-2025 – only a few days left to contribute).
• Visit our Donation page for a subscription or one-off donation.
• Interested in speaking or know someone that might be? Visit or share https://www.horizonevents.info/guaranteedsafeaiseminars. We welcome speakers working on GS AI or related research agendas, including in relation to world models, verification methods, safety specifications, real-world applications, …
• Subscribe to the Luma calendar: https://lu.ma/guaranteedsafeaiseminars.
• Watch recordings: https://www.youtube.com/@HorizonEvents_.
• Give us feedback: https://airtable.com/appYgVucDqD7WkTyy/paghbgReScjYO69Dy/form.
• Subscribe to Horizon Events on Bluesky https://bsky.app/profile/horizonevents.info.

Thanks to you all.

Best,

– Orpheus Lummis + Horizon Events