dynolab
Dyno Lab — an AI safety, alignment and interpretability research workbench for Apple Silicon. Explore activations, probes, SAEs and interventions with local inf
What is dynolab?
Dyno Lab — an AI safety, alignment and interpretability research workbench for Apple Silicon. Explore activations, probes, SAEs and interventions with local inference, a Python SDK, APIs and MCP.
Overview of dynolab
What it is
Dyno Lab is a free, open-source Mac application for AI safety, alignment, and interpretability research, designed for Apple Silicon. It allows users to run models locally, inspect internal activations, train probes and sparse autoencoders (SAEs), and perform interventions. The app includes a Python SDK, HTTP API, and MCP support. It bundles Python and MLX; model weights are downloaded separately. Requires macOS 14+ and Apple Silicon.
Key features
- Assistant: describe a research question in natural language (typed or spoken) and Dyno sets up the experiment using a local model.
- Studies: a local notebook to keep questions, prompts, results, and observations, with streaming, revision comparison, and archiving.
- Activation capture: explore layer × token activation magnitudes and next-token predictions; resident capture reuses the serving model’s weights.
- Probes: train labeled linear probes and inspect held-out metrics alongside control baselines.
- SAE sandbox: train small sparse autoencoders, inspect training curves and feature examples.
- Interventions: scale, ablate, patch, or steer block outputs and compare with an unchanged baseline.
- Neuron analysis: capture every neuron of a layer over a corpus, browse poles, label with local models, train SAEs at scale, and ablate or steer units.
- Token analysis: explore token probabilities and alternatives, compare outputs, and reopen saved analyses.
- Claude Code and Codex support: run these tools on your own downloaded models with one command.
Use cases
- Investigating model behavior and testing hypotheses about internal reasoning.
- Reproducing interpretability experiments (e.g., identifier-fidelity experiments with Qwen models).
- Running agent sandbox tests where a lead agent creates teammates to reach a goal while a hidden observer records rule violations.
- Evaluating models with Inspect AI benchmarks alongside agent tests.
- Exploring activation patterns, training probes, and SAEs for safety research.
Getting started
- Download the Apple Silicon DMG from the repository, open it, and drag Dyno into Applications. 2. Open Dyno. In Discover, download an MLX model that fits your Mac, then select and start it in Models. 3. Open the Assistant (⌘J) and describe what you want to test, or try a prompt in plain chat. See each request in Execution, then capture your first activation map. To run Claude Code or Codex on your own model:
dyno launch claude --model qwen3.8-27b dyno launch codex --model qwen3.8-27bRequires Apple Silicon and macOS 14+. Python and MLX are bundled; model weights are downloaded separately.
License and pricing
License: MIT. Pricing: free and open source.
Based on the project's public README and repository metadata.