An AI safety research organization applying Singular Learning Theory and developmental interpretability to understand how capabilities and values emerge during neural network training.
An AI safety research organization applying Singular Learning Theory and developmental interpretability to understand how capabilities and values emerge during neural network training.
People– no linked people
Updated 04/02/26Funding Details
Updated 04/02/26- Annual Budget
- $2,500,000
- Current Runway
- -
- Funding Goal
- -
- Funding Raised to Date
- -
Org Details
Updated 04/02/26Timaeus is an AI safety research organization founded by Jesse Hoogland and Daniel Murfet, publicly announced in October 2023. The organization's mission is to empower humanity by making breakthrough scientific progress on AI safety, with a focus on applying Singular Learning Theory (SLT) to technical AI alignment. Singular Learning Theory, developed primarily by mathematician Sumio Watanabe, establishes deep connections between the geometry of a model's loss landscape and its internal structure. Neural network loss landscapes are complex surfaces full of singularities where models can change internally without affecting external behavior, potentially masking dangerous misalignment. Timaeus leverages SLT to develop rigorous, mathematically grounded tools for understanding and controlling these phenomena. The organization's core research agenda is developmental interpretability, which studies how structure emerges in neural networks during training. Key to this approach is the Local Learning Coefficient (LLC), a measure derived from SLT that can detect phase transitions during training in models up to billions of parameters. Timaeus researchers have demonstrated that changes in the LLC predict critical developmental stages where models acquire new capabilities or internal structures. This work bridges mechanistic interpretability with ideas from statistical physics and developmental biology. Timaeus has published research at top venues including ICLR and AISTATS, with notable papers on the developmental landscape of transformers, local learning coefficients for deep neural networks, embryology of language models, and more recently on patterning (controlling neural network structure development) and spectroscopy (using perturbations to infer internal model structure in large language models). The organization maintains the devinterp Python library, an open-source toolkit for developmental interpretability research. Beyond direct research, Timaeus has built a significant community and educational infrastructure. The organization co-organized the SLT and Alignment Summit in Berkeley in June 2023, co-organized the ODYSSEY 2025 multi-track alignment conference, and hosted a Focus Period on Mathematical Science of AI Safety at the University of Sydney in late 2025. Jesse Hoogland and Daniel Murfet serve as MATS mentors, training the next generation of SLT and alignment researchers. Timaeus operates as a remote-first organization with hubs in Berkeley, Melbourne, and London. The team includes mathematicians, physicists, computer scientists, learning theorists, engineers, and operations staff. The organization was initially fiscally sponsored by Ashgro and incorporated as Timaeus Research Inc. in California in March 2025. Major funders include the Survival and Flourishing Fund (via Jaan Tallinn), with initial seed funding from a Manifund regrant led by Evan Hubinger. The organization is advised by David Dalrymple (ARIA), Evan Hubinger (Anthropic), and Adam Gleave (FAR AI).
Theory of Change
Updated 04/02/26Timaeus believes that Singular Learning Theory provides the mathematical foundation needed to rigorously understand how neural networks develop internal structure during training, including when and how they acquire capabilities, values, and potentially dangerous behaviors. By detecting and understanding phase transitions in training through measures like the Local Learning Coefficient, Timaeus aims to build scalable evaluation and interpretability tools that can identify critical developmental moments before dangerous properties emerge. This would enable alignment interventions at the training level, ensuring models develop in ways that reflect human values. Their approach bridges pure mathematics with empirical machine learning, creating a scientific framework where alignment properties can be formally stated, measured, and verified rather than merely hoped for.
Grants Received
Updated 04/02/26Projects– no linked projects
Updated 04/02/26Discussion
Key risk: The main risk is that SLT-derived signals (e.g., LLC) may fail to generalize to safety-relevant phenomena in real frontier systems, leaving their elegant math and promising mid-scale demos without actionable interventions or lab adoption, especially given limited access to compute and proprietary models.
Case for funding: Timaeus is uniquely positioned to turn Singular Learning Theory into practical, scalable developmental interpretability (LLC, spectroscopy, patterning, devinterp) that can flag dangerous phase transitions during training—an approach already shaping top lab agendas and likely to create training-time levers for frontier models.