Independent research lab · San Francisco
We study how reasoning, generalization, and failure emerge in large models — the discontinuities that show up between one checkpoint and the next. We publish what we find, including the results that complicate our own work.
Our agenda is narrow on purpose. Each thread has a team that owns it end to end, from the training run to the paper.
We reverse-engineer the internal structure of trained models — features, circuits, and the representations that survive scale — so that claims about model behavior can be checked against mechanism rather than inferred from benchmarks.
Long-horizon reasoning fails in structured, reproducible ways. We build evaluations that isolate the step where a trajectory goes wrong, and study which training interventions actually move that step.
Some capabilities arrive gradually and some arrive at once. We characterize which is which, and look for the measurable precursors that appear in a run before a capability does.
As models take on tasks their supervisors cannot fully check, evaluation becomes the bottleneck. We work on protocols that stay reliable when the system under review is stronger than the reviewer.
Before a question becomes a training run, it becomes a measurement. If we cannot say what result would change our mind, the project is not ready to start.
Findings are reproduced across at least three model scales before we believe them. Most results that look clean at 1B do not survive the second replication.
Failed replications and null results go out with the same care as headline findings. The field is slower when only the wins are legible.
Small teams, long horizons, and no separation between the people who run the experiments and the people who write them up. We hire for taste in problems more than for a specific background.