Show HN: Dolittle, explore and steer Qwen3's internal activations

Category: ai-ml

Tags: interpretability, llm, activations, steering, cuda, metal

Score: 7.0/10 (Innovation: 7, Technical: 8, Documentation: 7, Utility: 6)

Dolittle is a C++20 inference engine that runs Qwen3 8B with a natural language autoencoder to interpret and steer internal activations via a browser interface. It allows users to read explanations of token-level activations, edit them, and compare model behavior with and without the changes, supporting CUDA and Metal backends.

Target audience: AI researchers, ML engineers, interpretability researchers

Repository: https://github.com/plainsight-systems/dolittle · C++ · NOASSERTION · 1 stars

View on Hacker News