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

Plainsight Systems LLC released Dolittle, an open-source C++20 engine that runs Qwen3 8B alongside marco-bazzani's natural language autoencoder, capturing the output of block 24 (zero-based) so users can read text descriptions of the model's internal activations, edit those descriptions, and compare the model's continuation with and without the change. The tool runs generation, interpretation and steering on CUDA and Metal through one browser interface, requires Python 3.11+ and Git, and downloads about 45 GB of model files on first launch; it is available on Linux/NVIDIA (tested on an RTX 3090 Ti with 24 GB VRAM) and macOS/Apple Silicon (macOS 26+, tested on an M3 Max with 128 GB unified memory), with a native Windows/NVIDIA build planned and AMD undecided. A historical 4B/8B reference comparison run in PyTorch used 15.4 GiB peak allocation with the models in separate BF16 phases, and the project is described as R&D under Plainsight Systems LLC.

Dolittle puts Qwen3’s generated thinking beside natural language explanations of its internal activations. Enter a prompt, select a token and read what the NLA describes. Edit that explanation, then compare the model’s continuation with and without the change. The C++20 engine runs Qwen3 8B with marco-bazzani’s natural language autoencoder , capturing the output of block 24 zero-based . Generation, interpretation and steering run on CUDA and Metal through the same browser interface. Each experiment keeps the original response, interpretation samples and steering comparisons available to explore. Use Python 3.11+ and Git, then run the same command on either supported host: git clone https://github.com/plainsight-systems/dolittle.git cd dolittle python3 scripts/run.py | Platform | Status and launch path | Tested hardware | |---|---|---| | Linux/NVIDIA | Available. Builds and runs in Docker with NVIDIA Container Toolkit. | RTX 3090 Ti, 24 GB VRAM | | macOS/Apple Silicon | Available. Builds and runs natively with Metal. Requires macOS 26+, Xcode 26 with Metal toolchain, CMake 3.24+ and OpenSSL 3. | M3 Max, 128 GB unified memory | | Windows/NVIDIA | Native build planned. | — | | AMD | Undecided. | — | The launcher builds from source, downloads about 45 GB of model files and checks their hashes. Later launches reuse those files and the saved verification result. Use --verify-models to check every file again and repair the cache. Open the full Open Dolittle link printed by the launcher and keep that terminal running. The public project website explains the experiment; the local service runs it. A browser on another machine can connect to that service. See application setup https://github.com/plainsight-systems/dolittle/blob/main/app/README.md for prerequisites, cache options and remote access, then follow the rhyming walkthrough https://dolittle.plainsight-systems.com/writing/running-an-investigation/ . The engine uses BF16 tensors and FP32 reductions. It keeps model weights on the GPU between requests when there is room for the next model, its cache and working buffers. Each generation gets a fresh KV cache. CUDA builds include native Ampere code and PTX that a compatible NVIDIA driver can compile for newer cards. Development checks: python3 scripts/check.py host On Linux: build CUDA sources in the pinned Docker environment. python3 scripts/check.py cuda-build See build and verification commands https://github.com/plainsight-systems/dolittle/blob/main/tests/README.md and VS Code setup https://github.com/plainsight-systems/dolittle/blob/main/tests/README.md vs-code-and-compiler-settings . The historical 4B/8B reference comparison https://github.com/plainsight-systems/dolittle/blob/main/docs/research/2026-10-16-nla-comparison/README.md used PyTorch to compare the model pairs before building this engine. The 8B run used 15.4 GiB peak allocation with the models running in separate BF16 phases. The project site can be previewed independently: python3 scripts/site.py serve Open http://127.0.0.1:1315/ http://127.0.0.1:1315/ . Python 3.11+ and Hugo 0.154.2 are required; see site setup https://github.com/plainsight-systems/dolittle/blob/main/web/README.md . The production URL is https://dolittle.plainsight-systems.com/ https://dolittle.plainsight-systems.com/ . GitHub Pages builds and deploys the site on pushes to main . Doctor Dolittle learned to communicate with animals. This project takes its name from that idea: making an unfamiliar intelligence understandable. A natural language autoencoder translates an internal activation vector into text, then reconstructs a vector from that text. Its descriptions are interpretations that can be wrong. Dolittle distinguishes those interpretations from the reasoning text the target model generates itself. The purpose is to understand the machinery by implementing it and making it easy to interact with. Dolittle is R&D under Plainsight Systems LLC. | Path | Purpose | |---|---| | src/ | C++20, CUDA and Metal inference engine | | app/ | Browser interface served with the engine | | web/ | Public project website | | tests/ | Numerical, GPU, and application verification | | scripts/ | Source-build, setup, launch, and check tooling | | docs/architecture/ | Logical design, change axes, and file mapping | | docs/process/ | Project memory, work queue, and local workflow | | docs/research/ | Source investigations and measurements | | docs/writeups/ | Explanations, explorations, and milestone articles | | docs/assets/ | Project artwork and its provenance | Downloaded weights live in ignored models/ ; build output and local experiments also stay out of Git. Start with project memory https://github.com/plainsight-systems/dolittle/blob/main/docs/process/MEMORY.md , the work queue https://github.com/plainsight-systems/dolittle/blob/main/docs/process/QUEUE.md , and the local workflow https://github.com/plainsight-systems/dolittle/blob/main/docs/process/workflow.md . Agent entry points are AGENTS.md https://github.com/plainsight-systems/dolittle/blob/main/AGENTS.md and CLAUDE.md https://github.com/plainsight-systems/dolittle/blob/main/CLAUDE.md . The architecture overview https://github.com/plainsight-systems/dolittle/blob/main/docs/architecture/README.md explains the computation and how the CUDA and Metal backends share it. Verification https://github.com/plainsight-systems/dolittle/blob/main/tests/README.md covers numerical comparisons, tests and build checks. - Anthropic's natural language autoencoder article https://www.anthropic.com/research/natural-language-autoencoders - Natural Language Autoencoders paper https://transformer-circuits.pub/2026/nla/index.html - Original NLA implementation and checkpoints https://github.com/kitft/natural language autoencoders - Selected Qwen3 8B NLA pair https://huggingface.co/marco-bazzani/Qwen3-8B-nla Code and configuration use Apache-2.0; documentation and original artwork use CC BY 4.0. See LICENSE https://github.com/plainsight-systems/dolittle/blob/main/LICENSE for scope and NOTICE https://github.com/plainsight-systems/dolittle/blob/main/NOTICE for attribution. Third-party weights and components retain their own licenses.