What Timaeus does
Timaeus (timaeus.co) is an AI safety research organization that focuses on applying singular learning theory (SLT) to developmental interpretability and technical AI alignment. Its core research direction is to understand how training data shapes model behavior and how internal neural-network structure emerges during training (e.g., through stages/“phase transitions” in the SLT sense), with the stated goal of building interpretability and evaluation tools that can support safer model development and alignment.
As of 2026-06-10, Timaeus says it has “now [become] Resolution,” and the site points to Resolution’s open roles and launches. The Timaeus program is therefore best understood as the SLT/developmental-interpretability research component that fed into Resolution’s broader nonprofit research effort.
Technically, Timaeus’s differentiation (vs. many interpretability approaches centered on activations/probes) is its emphasis on SLT-grounded measurements and estimators (e.g., susceptibilities and related “spectroscopy” tooling) and “developmental” analysis of how internal structures form over training. The organization maintains an open-source research package, DevInterp, intended to let external researchers perform SLT/DevInterp-style research on large language models.