Schedule
Day 1 · Mon 9 Nov
Orientation, and the science of deep learning
Then the alignment ecosystem and who is who. What is alignment a problem of, and why now?
Day 2 · Tue 10 Nov
Learning theory, reinforcement learning, computational mechanics
Which structure must any optimal predictor of a process carry, and where do we find it in trained networks?
Day 3 · Wed 11 Nov
Reinforcement learning in practice, and singular learning theory
Bayesian and dynamical. What does SLT say about a real training run, and what can it not say yet?
Day 4 · Thu 12 Nov
Physics and interpretability
Which physics transfers to trained networks: field theory, mean field, renormalisation, and where does each stop?
Day 5 · Fri 13 Nov
Learning dynamics
Day 6 · Mon 16 Nov
Agency: world models and Solomonoff induction
What are the right objects (agents, beliefs, goals), what would a theory of agency have to explain, and which of the candidate formalisms (AIXI, logical induction, condensation, world models) answers which question? Deliberately open: no consensus is the honest starting point.
Day 7 · Tue 17 Nov
Agent foundations, and mechanistic interpretability
Zulip, Lean, agents, a live demo. Which parts of your own work would you hand to an agent tomorrow, and which never?
Day 8 · Wed 18 Nov
Open-source game theory, and heuristic arguments
What would a proof about a trained network look like?
Day 9 · Thu 19 Nov
Data attribution, debate, and the physics of deep learning
Oversight as a protocol problem. What does complexity theory buy, and where does the human in the loop break the argument?
Day 10 · Fri 20 Nov