~/software
[ 25 repos ] platform · main dependency · install under each // hover to preview
Research code09
- Claim-correctness probes / ProbeMonactivation-probes-claim-correctnessResearch code and two domain-specific probes (legal QA, biographies) for comparing Llama 3.1 8B’s hidden-state correctness signal with its stated confidence.
- Qwen Scope Labqwen-scope-labA browser-based workbench for sparse-autoencoder interpretability on Qwen3.5-2B: inspect features token by token, steer generation, and test detectors against controls. The model and its SAE run on the Mac through MLX.
- Block-sparse steeringbsf-steeringTrains a block-sparse dictionary on Qwen3.5-2B-Base’s layer-14 activations and tests steering with it against SAE features, difference-of-means and random directions, under a preregistered plan. Of five concepts, only medical vocabulary beat the difference-of-means and random baselines on held-out words, and 40% of those continuations fell into repetition loops. The SAE baseline failed its liveness check on every concept, so it gave no valid comparison.
- Panel-style reasoningmulti-modelTraining and evaluation code comparing a panel-style Qwen3-30B-A3B adapter with Qwen’s native thinking mode. Written up in two posts.
- Society of Thoughtsociety-of-thought-benchA preview adapter and synthetic evaluation suite (Countdown, evidence verdicts) for inspecting visible multi-persona reasoning.
- Hypothesis Forgehypothesis_forgeAn RL environment that scores proposed hypotheses on grounding, novelty, and the experiments that could falsify them.
- Adaptive retrieval experimentsadaptive_rag_rlmTest adaptive recursive search against fixed retrieval pipelines on procedurally generated document collections.
- autoresearch-evoAdds run memory, novelty-oriented search, and agent review to autoresearch’s training-code experiments.
- ProofgradeA first-pass reviewer for olympiad-style proofs, using an official solution and rubric. Includes benchmark results, a command-line tool, and a local API.
ML on Apple Silicon04
- nanochat-mlxTrain a small chatbot from scratch on Apple Silicon, from tokenizer training to a chat interface.
- Gemma 4 on a 24GB MacBookgemma4-m4-proMeasured runtime configurations, memory use, and fallback options for Gemma 4 26B A4B on an M4 Pro MacBook.
- Gemma 4 Sudoku on MLXtrain-gemma4-sudoku-on-your-macbookRun a Sudoku reinforcement-learning notebook for Gemma 4 E2B locally on Apple Silicon.
- TTT-Discover on MLXttt-discover-autoresearch-mlxSearch for improvements to training code with an RL outer loop, evaluating candidates locally on Apple Silicon.
Agent tooling01
macOS utilities11
- SunShiftAdjust Night Shift warmth directly from the menu bar.
- TabPilotFind duplicate and stale Safari tabs, search open pages, and group tabs by domain.
- SafariMarkdownTurn the current Safari page into Markdown using Codex.
- GhostLabelGive Ghostty tabs descriptive titles using Codex.
- PasteForgeReformat, encode, hash, and inspect clipboard text from the menu bar.
- TextDropPaste text into a minimal editor and save it as a file. You type the extension in the save dialog.
- ClipDropOpens with the clipboard’s text already loaded. Edit it and save it as Markdown, code, JSON, or one of the other supported text types.
- DiskPulseCheck used and available space on each connected drive.
- PortSentrySee which processes own listening TCP ports and stop them from the menu bar.
- ProcessBeaconGet notified when a long-running process finishes.
- BrewPilotStart, stop, and restart Homebrew services from the menu bar.