Our Mission
To empower humanity by making breakthrough scientific progress on AI safety.
We are first and foremost a research organization, working on advancing our understanding of advanced AI systems, so we can achieve safe and positive outcomes.
AI Safety
Humanity is building highly powerful AI systems but we still do not know how to ensure those systems are robustly aligned with human values. We conduct scientific research to ensure the development of advanced AI systems goes well.
Singular Learning Theory
At Timaeus, we’re using singular learning theory to develop the science of understanding
how training data determines model behavior.
- Data
- Geometry
- Learning Process
- Algorithms
We study how patterns in training data determine…
…the geometry of the loss landscape, which determines…
…the learning trajectory followed by SGD, which determines…
…the ultimate algorithms and behaviors learned by neural networks.
Our Research
Evals
The aim of our research on model evaluations is to advance the science of generalization, to predict how large learning machines will behave.
Learn moreInterpretability
The aim of our research on interpretability is to understand how large learning machines work on the inside.
Learn moreAlignment
The aim of our research on alignment is to advance the reliability and robustness of our techniques for aligning advanced AI systems with human values.
Learn morePublications
Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient
Wang et al.
October 4, 2024
Loss landscape geometry reveals stagewise development of transformers
Wang et al.
June 16, 2024
Estimating the Local Learning Coefficient at Scale
Furman and Lau
February 6, 2024
The Developmental Landscape of In-Context Learning
Hoogland et al.
February 4, 2024
Quantifying Degeneracy in Singular Models via the Learning Coefficient
Lau et al.
August 23, 2023