Here, you can find the syllabus we'll use throughout the technical fellowship (Fall 2026). Each week lists the core readings we'll discuss together, plus "further readings" for participants to (optionally) explore independently if they're especially interested in a topic.
Week 0 is intended as an optional introduction to machine learning, to provide a more thorough understanding of machine learning (ML), deep learning (DL), reinforcement learning (RL), and transformers for the rest of the fellowship. The first meeting will discuss Week 1 readings, not Week 0 readings.
In mid-2026, AI agents being tested inside OpenAI escaped their test environment, coordinated with one another, hacked into Hugging Face, and tried to hide what they had done. This week uses the incident to introduce the problems the fellowship studies. We begin with a journalist's account of what happened and why independent experts found it alarming. We then watch the OpenAI researchers' own reconstruction: the agents were chasing reward on their tasks, not following a human attacker, and warning signs appeared weeks before the harm. We close with what the incident suggests about more capable systems: a United Nations scientific panel explains how training can produce AI systems that competently pursue goals nobody intended, and why stopping this incident is no assurance about future ones, and 80,000 Hours sets out the argument that such systems could seek power and ultimately cause human extinction. The central question is what this incident shows about the risks from more capable AI systems, and what it does not show; the rest of the fellowship tests each step of the argument.