Iliad × MAISI · MISO
MAISI-Iliad Safety Orientation
A two-week school on the mathematics of AI alignment, for research mathematicians new to the field. Berkeley, 9 to 20 November 2026.
The school
Ten days on the mathematics of AI alignment
The Mathematical AI Safety Institute brings a group of senior mathematicians to the Bay Area to work on problems relevant to AI safety. They start in January; the school runs first, from 9 to 20 November, so that members can start working on problems as soon as possible. Iliad is putting the programme together with Jacob Tsimerman.
Senior researchers give introductory lectures: the basic definitions, what is known, and which mathematics is involved. The lectures build on each other along five threads, so that ten days read as one course, and participants leave able to read the literature, with open problems suited to their own mathematics.
Each day
Three lectures, at 10:00, 11:30 and 14:00, each an hour plus questions, with coffee and lunch between. Most days carry two threads, for example learning in the morning and agency in the afternoon, so no day is a single formalism.
The 15:30 slot is never a lecture: office hours with the day's lecturers and Iliad researchers, on some days a discussion on how the week's theories relate, or a research talk. The last afternoon is a panel with senior researchers who came to alignment from mathematics or physics.
Who it is for
Research mathematicians who are strong in their own field and new to alignment. No background in deep learning is needed; the lectures start from the definitions.
Lots of senior safety researchers, from ARC, Simplex, Resolution and others, are around during the two weeks. Mathematicians from outside MAISI are welcome too.
What Iliad brings
The programme is a shortened version of the Iliad Intensive, Iliad's course on AI safety theory. Every lecturer is told what the earlier days have covered and what the later days need. On every day, Iliad researchers who wrote the corresponding Intensive materials are in the room and available beyond the lecture.
Five threads
What the lectures cover
- Learning
Why deep learning works: classical learning theory, learning dynamics, singular learning theory, data attribution.
- Structure
What is inside a trained network: computational mechanics, the physics of deep learning, interpretability.
- Agency
What an agent is: reinforcement learning, world models, Solomonoff induction, agent foundations.
- Guarantees
What can be proven about a model: heuristic arguments, debate and scalable oversight.
- Verification
Cryptography for safety: backdoors, steganography, zero-knowledge proofs.
Programme
Ten days, forty sessions
Week 1 Mon 9 Nov to Fri 13 Nov
Mon 9 Nov
Orientation, and the science of deep learning
- 10:00 · Orientation The alignment problem for mathematicians Louis Jaburi, Jacob Tsimerman
- 11:30 · Orientation Alignment in practice: how frontier models are trained and where alignment fails
- 14:00 · Learning Science of deep learning: the mysteries of approximation, generalisation and optimisation Zach Furman
- 15:30 Q&A with the Iliad team, opened by a short talk on timelines: why the urgency, and what "feeling the AGI" means for a mathematician Louis Jaburi, Leonard Bereska
Tue 10 Nov
Learning theory, reinforcement learning, computational mechanics
- 10:00 · Learning Classical learning theory, and what deep learning broke
- 11:30 · Agency Foundations of reinforcement learning: from preferences to utility to reward
- 14:00 · Structure Computational mechanics I: belief states and their geometry Paul Riechers
- 15:30 Computational mechanics open problems, run by Simplex Paul Riechers, Adam Shai
Wed 11 Nov
Reinforcement learning in practice, and singular learning theory
- 10:00 · Agency RL as used to train frontier models: RLHF, RLVR, and what is known about them
- 11:30 · Learning Singular learning theory I: Bayesian learning, free energy, singular models
- 14:00 · Learning Singular learning theory II: susceptibilities and linear response
- 15:30 Office hours Kai Ogden
Thu 12 Nov
Physics and interpretability
- 10:00 · Structure Interpretability from a physicist's perspective
- 11:30 · Structure Statistical mechanics of learning: mean-field theory, phase transitions, emergence under RL
- 14:00 · Structure Physics-inspired interpretability
- 15:30 Optional research talk; office hours
Fri 13 Nov
Learning dynamics
- 10:00 · Learning Learning dynamics in deep networks: an introduction Andrew Saxe
- 11:30 · Learning Deep linear networks and saddle-to-saddle dynamics; from lazy to rich: NNGP, the NTK, feature learning Daniel Kunin, Jamie Simon
- 14:00 · Learning Continuation of the joint session if needed; otherwise problem session on the morning's material with the day's lecturers Andrew Saxe, Daniel Kunin, Jamie Simon
- 15:30 Fixed discussion 1: how the learning theories relate (classical, training dynamics, SLT, Bayesian), doubling as the week-one synthesis; office hours Daniel Kunin, Zach Furman, Kai Ogden, Andrew Saxe
Week 2 Mon 16 Nov to Fri 20 Nov
Mon 16 Nov
Agency: world models and Solomonoff induction
- 10:00 · Agency World models and transducers
- 11:30 · Agency Solomonoff induction and AIXI Cole Wyeth
- 14:00 · Agency Open slot
- 15:30 Fixed discussion 2, opening week two: agent foundations as a pre-paradigmatic field Louis Jaburi, Cole Wyeth
Tue 17 Nov
Agent foundations, and mechanistic interpretability
- 10:00 · Agency Agent foundations I: logical induction
- 11:30 · Agency Agent foundations II: condensation
- 14:00 · Structure Mechanistic interpretability, the high-level view: sparse coding, activation geometry, causal abstraction Joshua Batson
- 15:30 Automating research: what a mathematician does with a thousand research assistants
Wed 18 Nov
Open-source game theory, and heuristic arguments
- 10:00 · Agency Open-source game theory Jacob Tsimerman
- 11:30 · Guarantees Heuristic arguments I Jacob Hilton, Victor Lecomte
- 14:00 · Guarantees Heuristic arguments II: open problems session
- 15:30 Office hours Louis Jaburi
Thu 19 Nov
Data attribution, debate, and the physics of deep learning
- 10:00 · Learning Data attribution and influence functions Roger Grosse
- 11:30 · Guarantees Debate and scalable oversight
- 14:00 · Structure Physics of deep learning
- 15:30 Office hours Louis Jaburi
Fri 20 Nov
Verification, interpretability, and the closing panel
- 10:00 · Verification Backdoors
- 11:30 · Verification Steganography
- 14:00 · Structure Physics-inspired and ambitious interpretability; or zero-knowledge proofs if RESI sends a third lecturer
- 15:30 Closing panel: what the field needs from mathematicians, how to get involved, how to continue Alexander Gietelink Oldenziel
Preliminary. Sessions may still move; lecturers are named once they have confirmed. The full schedule has abstracts and readings.
People
Who is involved
Lecturers
Adam Shai
Simplex
Andrew Saxe
UCL
Cole Wyeth
AIXI Labs
Daniel Kunin
UC Berkeley
Jacob Hilton
ARC
Jamie Simon
UC Berkeley
Joshua Batson
Anthropic
Paul Riechers
Simplex
Roger Grosse
Vector Institute and Anthropic
Victor Lecomte
ARC
Zach Furman
Iliad
MAISI
Jacob Tsimerman
MAISI
Iliad
Alexander Gietelink Oldenziel
Iliad
Kai Ogden
Iliad
Leonard Bereska
Iliad
Louis Jaburi
Iliad
For participants · Starts Mon 9 Nov
Orientation, and the science of deep learning
The school opens on Mon 9 Nov at 10:00.
- 10:00 The alignment problem for mathematiciansLouis Jaburi and Jacob Tsimerman
- 11:30 Alignment in practice: how frontier models are trained and where alignment fails
- 14:00 Science of deep learning: the mysteries of approximation, generalisation and optimisationZach Furman
- 15:30 Q&A with the Iliad team, opened by a short talk on timelines: why the urgency, and what "feeling the AGI" means for a mathematicianLouis Jaburi and Leonard Bereska