Scoping Developmental Interpretability (Timaeus first funding)

The Scoping Developmental Interpretability grant was Manifund's late-2023 regrant to Jesse Hoogland and the founding team of what later became Timaeus. The award funded a six-month period of work to test whether singular learning theory (SLT) could be turned into a research program for neural-network interpretability. Manifund's own 2025 regrants retrospective lists this ticket as one of the higher-impact regrants in the cohort, on the grounds that it predated and seeded the formal lab.

Timaeus was incorporated as an alignment lab shortly afterward, with co-founders Jesse Hoogland (CEO) and Daniel Murfet (a mathematician at the University of Melbourne who is one of the principal academic developers of singular learning theory). The lab now has hubs in Berkeley, Melbourne, and London, and runs streams within the MATS program.

The technical thesis is that the geometry of the loss landscape carries information about what a network is actually computing. SLT, originally due to Sumio Watanabe, formalizes the observation that neural-network loss surfaces are not smooth bowls but are full of singularities, that is, directions in which parameters can change without changing model behavior. The Local Learning Coefficient (LLC) is the SLT-derived invariant Timaeus uses to measure model complexity at a point. Subsequent papers (Furman and Lau, 2024, on scaling LLC estimation; Hoogland et al., 2024, on developmental landscapes of in-context learning; refined LLC variants applied to attention heads) show that the LLC can detect phase transitions during training, including transitions that are invisible from the loss curve, at scales up to 7B parameters.

In the stack the grant and the lab sit at the evaluation and weights layers. The position is complementary to Anthropic-style mechanistic interpretability: rather than finding circuits inside a trained model, developmental interpretability tries to characterize the training trajectory and locate where new structure is acquired. The Manifund regrant amount is not stated on the public project page; what is stated is that it was the first funding for the program and accelerated subsequent fundraising by months.