Timaeus is a nonprofit research organization devoted to AI safety, with a mission it describes as empowering humanity by making breakthrough scientific progress on AI safety. Its distinctive angle is mathematical: it applies frameworks from algebraic geometry and statistical physics, especially singular learning theory, to understand how training shapes the behavior of neural networks.
The organization’s flagship research agenda is developmental interpretability, which studies how capabilities and values emerge over the course of a network’s training rather than only inspecting a finished model. Timaeus argues that understanding these developmental dynamics enables interventions that help ensure models end up aligned with human values. Singular learning theory provides the mathematical lens, modeling how the geometry of a network’s loss landscape governs what the model learns and when.
Timaeus maintains research hubs in Berkeley, Melbourne, and London, runs fellowship programs and seminars, and hosts events such as its Mathematical Science of AI Safety focus period and the ODYSSEY conference.
For a general reader, Timaeus represents a bet that deep mathematical theory, not just empirical probing, will be needed to truly understand and steer how advanced AI systems form their abilities.