RESI Reading Group on AI Safety

The RESI reading group is a weekly forum, meant as a platform for open discussion. Each week a set of reading materials and guiding questions will serve as a basis for conversation, and all participants are encouraged to read and take part. The topics will touch on AI safety in the broad sense: from historical approaches, to modern debates and ideas. Sessions will encompass both technical and non-technical works, and no particular background is required.

On top of being an open space for discussing AI safety, the reading group will also serve as a door to the RESI community, and will accept visitors and guests who are interested to take part in the conversation and get to know the organization. A number of spots are available for outside guests each session. If you are currently not a RESI member and wish to participate in the meeting, please fill out the interest form linked below.

The group is organized by Miranda Christ, Tal Herman, and Jonathan Shafer. For questions, feel free to contact Jonathan.


đź“… When: Wednesdays at 11:00 am ET

📍 Where: RESI offices (details shared with participants)

đź“§ Mailing list: sign up via the interest form

Fall 2026

The Emergence of Modern Cryptography

Discussion leader: Tal Herman

October 7, 2026

In the spirit of RESI, facing the definitional crisis in AI safety, we take inspiration from the successful project of modern cryptography. We chose to devote the first reading group session to discuss the breakthroughs that established modern cryptography as we know it: public-key cryptography, semantic security, interactive proofs and zero knowledge. In this session we will explore the landscape prior to these revelations and try to understand how the new ideas came about, focusing on the definitional project rather than on the definitions themselves.

Heuristic Defenses

Discussion leader: Miranda Christ

October 14, 2026

Currently, much of AI safety is akin to cryptography before the period of definitional progress in the 1970s and 1980s. Perhaps this is because AI safety properties are more difficult to define, schemes achieving these properties are more difficult to construct, or both.

In this session, we will explore heuristic methods in AI safety, focusing on defenses that initially appeared secure but were later shown to fail.

AI Safety Incidents in the Wild

Discussion leaders: Miranda Christ and Greg Gluch

October 21, 2026

As AI systems become increasingly capable and are deployed in more consequential settings, we are beginning to observe real-world incidents involving misuse, unintended behavior, and failures of safety mechanisms. In this session, we will examine documented AI safety incidents, with the goal of understanding what went wrong, why existing safeguards were insufficient, and what these incidents can teach us about the challenges of AI safety. Participants are encouraged to come prepared to briefly present an incident of their choice, either from the suggested readings or elsewhere.

Sandbagging: Capability Hiding and Capability Elicitation

Discussion leaders: Anastasiia Struss and Miranda Christ

November 4, 2026

Sandbagging is strategic underperformance on an evaluation. A developer deliberately trains a model that appears less capable than it really is. For example, a model capable of solving difficult cybersecurity tasks might answer them incorrectly during evaluation but recover its full performance when given a secret password or deployed in a different environment. Sandbagging could therefore cause evaluators to underestimate a system’s risks and approve a model that requires stronger safeguards.

We will read about empirical evidence of capability hiding and discuss an attempt to provide rigorous guarantees for capability hiding. Our goal is to identify the limitations of current definitions, constructions, and threat models.

Existential Risk I

Discussion leader: Jonathan Shafer

November 11, 2026

We will discuss one of the most extraordinary and disturbing claims in AI safety: that artificial superintelligence will cause human extinction.

What are the strongest arguments for this claim, and do they withstand scrutiny? We will examine the reasoning and assumptions behind it, drawing on readings by Yudkowsky, Soares, and Bostrom.

Slides

Existential Risk II

Discussion leader: Jonathan Shafer

November 18, 2026

Safety Incidents in the Wild

Discussion leader: TBD

December 2, 2026