Tutorials

Step-by-step guides for running interpretability experiments with Seer.

Getting Started

1. Local Mode - Run locally without Modal (no GPU, API-based investigations)
2. Sandbox Intro - Spin up a GPU, load a model, let an agent explore it
3. Scoped Sandbox - Expose specific functions instead of full notebook access

Example investigations

1. Hidden Preference - Discover hidden bias using activation analysis
2. Introspection - Can models detect injected concepts?
3. Checkpoint Diffing - Find behavioral changes between model versions
4. Petri Harness - Auditor/target/judge pattern for probing behaviors