Research institute · Research orgs · Est. 2023
Timaeus (timaeus.co) was an AI safety non-profit research organization focused on developing developmental interpretability and technical alignment research through Singular Learning Theory (SLT). Its core technical premise is that the geometry and stagewise structure of neural network learning—linked to SLT concepts—can support interpretability tools and interventions intended to improve alignment-relevant understanding of how capabilities and values emerge during training.
Strategically, Timaeus positioned itself at the intersection of mathematical learning theory and interpretability: it emphasized SLT-informed methods for analyzing and discovering internal structure in neural networks (rather than only activation-based probing), and it described plans to translate that understanding into “intervention” pathways for steering model behavior. On June 10, 2026, Timaeus announced it was merging into a new nonprofit research organization called Resolution, and indicated that open roles would move to Resolution’s hiring pipeline.
Publications are not yet indexed for this organization. Read them on timaeus.co ↗.
Timaeus publication describing Posterior Prefix Tuning (PPT) for eliciting model continuations by optimizing a prompt distribution via gradients estimated from prior samples, in a setting where Bayes-filtered transformers are exact. ([timaeus.co](https://timaeus.co/research/2026-07-21-elicitation-without-backprop))
Linear Response Estimators for Singular Statistical ModelsTimaeus publication defining susceptibilities as data-perturbation responses of observables for singular statistical models and introducing estimators with consistency and asymptotic unbiasedness claims. ([timaeus.co](https://timaeus.co/research/2026-05-08-linear-response-estimators))
Interpreting Reinforcement Learning Agents with SusceptibilitiesTimaeus publication generalizing susceptibilities to regret in deep reinforcement learning and discussing interpretability of stagewise development in a gridworld model. ([timaeus.co](https://timaeus.co/research/2026-05-08-rl2-interpreting-agents))
Guide for Sampling Hyperparameter SelectionTimaeus practical guide for using SGMCMC-based sampling and selecting hyperparameters in SLT-related workflows. ([timaeus.co](https://timaeus.co/research/2026-04-21-sampling-guide))
Timaeus Research Fellows ProgramTimaeus launched a Research Fellows Program intended for tenured faculty and senior researchers to collaborate on SLT-informed alignment research while retaining their existing academic positions. ([timaeus.co](https://timaeus.co/blog/updates/2026-04-09-fellows))