Vibe Coding Discoverdynolab
View on GitHubDyno 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.
A free macOS Apple Silicon workbench for AI safety and interpretability research. Run local MLX models, capture activations, train probes and SAEs, run interventions, and script or automate studies via a Python SDK, HTTP API and MCP bridge.
Use Cases
Capture layer-by-token activations from a local modelTrain and evaluate labeled linear probes on held-out dataTrain small sparse autoencoders and inspect featuresRun causal patch sweeps, ablation and steering interventionsCompare model outputs against unchanged baselinesKeep reproducible saved studies with provenanceChat with and inspect local MLX models on a MacServe an OpenAI-compatible local endpoint for other toolsScript interpretability experiments with the Python SDKDrive experiments from an assistant over a local MCP bridgeDownload and manage local model weights from Hugging FacePair a Mac with a Windows NVIDIA worker for larger GGUF modelsMonitor throughput, time-to-first-token and GPU memoryReproduce published safety/robustness experiments
Built With
- Language
- Swift
- Frameworks
- MLX · MLX LM · SwiftUI · MCP · Hugging Face Hub · zeroconf · OpenAI-compatible API · hatchling
Tags
ai-safety · interpretability · mechanistic-interpretability · alignment · activations · probes · sparse-autoencoders · activation-patching · steering · local-inference · apple-silicon · mlx · macos · swiftui · python-sdk · mcp