Latest News

Erlangen AI Hub Seminar: Singular Learning Theory for Interpretability, Jesse Hoogland

In this seminar, Jesse Hoogland explores how singular learning theory provides a framework for connecting interpretability to the theory of generalization and the geometry of the loss landscape. He presents recent work at Timaeus applying SLT to different areas of interpretability, including an SLT-based generalization of influence functions for data attribution and an SLT-native approach to ablations for circuit discovery, and describes initial work on using these interpretability techniques to actively shape the learning process by controlling training data, with implications for developing better alignment techniques. Jesse Hoogland is the executive director of Timaeus, an AI safety non-profit working on interpretability and alignment, and a cofounder of the developmental interpretability research agenda and the broader program of applying singular learning theory to AI safety, which he is now pursuing with his team at Timaeus.

Jesse Hoogland - Singular Learning Theory and AI Safety

Jesse Hoogland presented "Singular Learning Theory and AI Safety" on September 17, 2025 at FAR.AI Labs Seminar.

Focus Period: Mathematical Science of AI Safety

Influence Dynamics and Stagewise Data Attribution

Compressibility Measures Complexity: Minimum Description Length Meets Singular Learning Theory

The Loss Kernel: A Geometric Probe for Deep Learning Interpretability

Bayesian Influence Functions for Hessian-Free Data Attribution

ODYSSEY 2025

Research Engineer @ Timaeus

Embryology of a Language Model

From Global to Local: A Scalable Benchmark for Local Posterior Sampling

Singular Learning Theory & AI Safety | SLT Seminar

In the SLT seminar, Jesse Hoogland from Timaeus talks to us about his research agenda applying singular learning theory to AI safety.

Singular Learning Theory and AI Safety | MATS 8.0

MATS 8.0 seminar by Jesse Hoogland. Singular learning theory (SLT) suggests that the geometry of the loss landscape is key to developing a better scientific understanding of deep neural networks, along with new practical tools for engineering safer systems.

Programs as Singularities | SLT Seminar

Daniel Murfet from Timaeus tells us how to think about Turning machines as critical points of an analytic function, from a recent paper with Will Troiani.

Embryology of AI: How Training Data Shapes AI Development | Cognitive Revolution

Jesse Hoogland and Daniel Murfet, founders of Timaeus, introduce their mathematically rigorous approach to AI safety through 'developmental interpretability' based on Singular Learning Theory. They explain how neural network loss landscapes are actually complex, jagged surfaces full of 'singularities' where models can change internally without affecting external behavior—potentially masking dangerous misalignment.

Studying Small Language Models with Susceptibilities | SLT Seminar

In this talk, Garrett Baker from Timaeus presented some recent work: 'Studying Small Language Models with Susceptibilities'. This paper uses singular learning theory to study the response of language models to perturbations in the data distribution.

Director of Operations @ Timaeus

Modes of Sequence Models and Learning Coefficients

Structural Inference: Interpreting Small Language Models with Susceptibilities

Programs as Singularities

You Are What You Eat – AI Alignment Requires Understanding How Data Shapes Structure and Generalisation

Structure Development in List-Sorting Transformers

Dynamics of Transient Structure in In-Context Linear Regression Transformers

Open Roles @ Timaeus

Jesse Hoogland on Singular Learning Theory | AXRP

In this video, Jesse Hoogland discusses Singular Learning Theory (SLT) and introduces the refined Local Learning Coefficient (LLC). They explore how this refined LLC helps uncover new circuits in language models.

The Australian AI Safety Forum 2024

Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient

Singular learning theory: exercises

ILIAD 2024

So you want to work on technical AI safety

Singular Learning Theory: Overview And Recent Evidence | Plectics

Jesse Hoogland introduces Singular Learning Theory (SLT), which links data structure, neural network architecture, loss landscape geometry, and learning dynamics. The talk surveys recent evidence supporting this framework and explores its implications for AI safety and interpretability.

Jesse Hoogland - Singular Learning Theory

Singular Learning Theory (SLT) is a novel mathematical framework that expands and improves upon traditional Statistical Learning theory using techniques from algebraic geometry, bayesian statistics, and statistical physics. It has great promise for the mathematical foundations of modern machine learning.

Singular Learning Theory with Daniel Murfet | AXRP

This video explores Singular Learning Theory (SLT), a Bayesian statistics framework that helps explain deep learning models, their learning dynamics, and generalization. The discussion with Daniel Murfet covers phase transitions, learning coefficients, AI alignment, and open problems in SLT’s application to AI safety and capabilities.

Stagewise Development in Neural Networks

Simple versus Short: Higher-order degeneracy and error-correction

Timaeus's First Four Months

Loss Landscape Degeneracy and Stagewise Development of Transformers

Generalization, from thermodynamics to statistical physics

Learning coefficient estimation: the details

The 2023 Oxford Conference

Announcing Timaeus

You're Measuring Model Complexity Wrong

The 2023 Melbourne Hackathon

The 2023 Amsterdam Retreat

The Local Learning Coefficient: A Singularity-Aware Complexity Measure

Jesse Hoogland–AI Risk, Interpretability | UCL

Jesse Hoogland is a research assistant at David Krueger's AI Safety lab in Cambridge. He now focuses on Singular Learning Theory and Developmental Interpretability. Previously, he co-founded a health-tech startup for automating bariatric surgery care.

DSLT 4. Phase Transitions in Neural Networks

DSLT 3. Neural Networks are Singular

The 2023 Berkeley Conference

The Primer

DSLT 2. Why Neural Networks obey Occam's Razor

DSLT 1. The RLCT Measures the Effective Dimension of Neural Networks

DSLT 0. Distilling Singular Learning Theory

The Physics of Intelligence: from Classical to Singular Learning Theory

Learn how Singular Learning Theory (SLT) helps us understand neural networks by analyzing their loss landscapes. This talk explores model interpretation and phase transitions during training. Presented at Imperial College London, May 22, 2023.

Approximation is expensive, but the lunch is cheap

Empirical risk minimization is fundamentally confused

The shallow reality of 'deep learning theory'

Gradient surfing: the hidden role of regularization

Interview Daniel Murfet on Universal Phenomena in Learning Machines

Spooky action at a distance in the loss landscape

Neural networks generalize because of this one weird trick