Other AI Project · Python

atlas-inference

Can a static SAE feature atlas of a frozen LLM be driven as a runnable, legible surrogate? Open research.

Rick@JuiceB0xC0de · Atlas 73.3 · #2= builder

What it is

The repository investigates whether a static feature atlas of a frozen language model can serve as a runnable surrogate by testing three gates: bridging features to vocabulary, concept directions, and layer-to-layer dynamics. It finds that while the atlas can produce coherent vocabulary and predict aggregate feature transitions, concept-specific information does not transport well through the layers.

Scored from the code at its last scan, 22 Oct 2026. How merit is scored

Where it counts

NLP & LLMs90%
Machine Learning80%
Data Science40%
MLOps10%

How much of the project is about each field, as scoring counts it. It ranks in the field it is most about.

Around it on the NLP & LLMs board

Whole board

RankProjectBuilderFieldMerit

19, shared rankaios-claude-codeML Application · TypeScriptAIOS is a starter kit for configuring Claude Code as a local AI operating system. It includes scripts to convert business documents (docx, xlsx, pdf) and media (images, videos) into a format consumable by Claude, along with a configuration system and an optional dashboard.58.4/100AdvancedAngelescu Andrei-CiprianNLP & LLMs
19, shared rankatlas-builder-review-and-upgradesAI Developer Tool · PythonThis repository provides a toolkit for running interpretability analyses on Hugging Face causal language models, including activation census, OV-circuit analysis, and compliance-behaviour axis extraction. It is designed as a developer tool for debugging and understanding model…58.4/100AdvancedRickNLP & LLMs
21atlas-inferenceOther AI Project · Python58.1/100ImplementationRickNLP & LLMs
22orionOther AI Project · TypeScriptOrion is a local-first desktop application for tracking hunting sessions in the game Entropia Universe. It monitors the game's chat log to extract events (loot, skill gains, combat, etc.), stores them in a local database, and provides analytics through a React/TypeScript…57.0/100ImplementationNepheriusNLP & LLMs
23HuggingFace-Project-LearningEducational Implementation · Jupyter NotebookThe repository is a collection of 38 Jupyter notebooks that follow the Hugging Face NLP and Reasoning courses, with added reimplementations of tokenizers, GRPO, and a causal LM training loop from scratch, and an end-to-end Romanian QA project (though the key evaluation notebook…56.2/100ImplementationPop AlexandruNLP & LLMs

What it is

Where it counts

Around it on the NLP & LLMs board

More from Rick