Guaranteed Safe AI Seminars 2024 review

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