Introduction to miStudio
MechInterp Studio (miStudio) is an end-to-end mechanistic interpretability platform designed to replace the fragmented tooling typically associated with AI safety research — Jupyter notebooks, custom scripts, and manually tracked experiments — with a professional, database-backed workbench.
By providing a unified environment for data management, SAE training, feature discovery, and causal intervention testing, miStudio allows researchers to move from hypothesis to proven intervention in a fraction of the time required by traditional methods.
The Scalability Spectrum: Edge to Cluster
miStudio is engineered with a "scale-agnostic" architecture:
- At the Edge: Run on an NVIDIA Jetson Orin or a laptop with a single RTX 3060. The software optimizes memory via quantization (4-bit, 8-bit) and micro-batching for 1B–3B parameter models.
- In the Lab: Deploy on multi-GPU workstations. miStudio detects CUDA devices and distributes extraction and training jobs across all available VRAM via its Celery/Redis task queue.
- In the Cloud: Deploy on GCP/AWS with Kubernetes for auto-scaling GPU access.
Jupyter notebooks suffer from "hidden state" — results depend on cell execution order. miStudio enforces structured workflows where every experiment is recorded in PostgreSQL with exact hyperparameters, making research reproducible by default.
Modern LLMs represent more concepts than they have neurons through superposition — neurons become "polysemantic," firing for unrelated concepts. Sparse Autoencoders (SAEs) "unpack" these neurons into individual, monosemantic features. miStudio is the workbench for this science.