Organisation
Timaeus
Developmental interpretability and singular learning theory — understanding how structure emerges in neural networks during training.
Timaeus is a non-profit research group pursuing "developmental interpretability", an approach that tries to understand how structure forms inside neural networks over the course of training rather than only in the finished model. Founded in October 2023 by Jesse Hoogland, Daniel Murfet and Alexander Gietelink Oldenziel, it draws on singular learning theory — a mathematical framework for learning machines developed by Sumio Watanabe — to identify the phase transitions the group argues organise how models learn. Its output is papers and tooling rather than models, and its bet is that a more fundamental theory of learning could make interpretability and alignment tractable; the programme remains early and its practical payoff unproven.
- Category
- Safety & alignment research
- Founded
- 2023
- Key people
- Jesse Hoogland, Daniel Murfet