Tools & Infrastructure
The platforms and libraries we build so that interpretability research does not depend on owning the hardware. Open source and open access wherever possible.
mlxterp: Mechanistic Interpretability on Apple Silicon
Interpretability tooling assumes you have an NVIDIA GPU. mlxterp brings activation capture and interventions to MLX on Apple Silicon, so a MacBook is enough to start.
Read more ↗
eDIF: A European Deep Inference Fabric for Remote Interpretability of LLMs
Inspecting a 70B model's internals needs GPUs most European labs do not have. eDIF is a shared remote fabric that lets researchers run interpretability experiments without owning the cluster.
Read more ↗Why we build this
Understanding what happens inside a large model requires running it and capturing its internal state — which, at frontier scale, means hardware most research groups do not have. That quietly decides who gets to scrutinise these systems.
For a field whose purpose is making AI legible and accountable, keeping the tools of scrutiny narrowly held is a poor arrangement. So we build shared infrastructure and open libraries, and publish what works alongside what does not.