norika1207-lab

norika1207-lab / llm-neuron-atlas

Public

3D explorable atlas of every neuron in a transformer LLM. Click neurons, trace residual highways, compare cross-architecture conservation. Live demo: https://charenix.com/qwen3b-atlas

interactive llm mechanistic-interpretability neuron-atlas qwen
14
4
85% credibility
Found May 27, 2026 at 14 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
HTML
AI Summary

LLM Neuron Atlas is a research visualization tool that creates a 3D explorable map of a large language model's internal structure. Think of it as Google Maps for an AI brain: you can zoom in to see 73,728 individual neurons across 36 layers, click on any neuron to see its top connections, and follow 'highways' showing how signals travel through the model. The project highlights eight specially named neurons with known functions—like one that handles negation or another that controls writing style. It runs entirely on a regular laptop without special hardware, making AI interpretability accessible to researchers and curious people who want to see inside how language models work.

How It Works

1
🔍 You hear about a 3D map of AI thinking

Someone tells you about a project that lets you explore inside a large language model like wandering through a city, clicking on buildings to see how they connect.

2
🖥️ You visit the live demo

You open the web page and see a colorful 3D tower with 36 floors, each floor holding 2,048 glowing points representing the AI's thoughts.

3
🎯 You click on a neuron and watch its signal travel

You click on a bright point and see glowing lines shoot forward through all 36 layers, showing how that one thought influences everything else.

4
You choose how deep to explore
🌈
Browse by region

Explore the five color-coded zones: early token processing (blue), surface features (green), structure tracking (yellow), deep meaning (pink), and output (orange).

🏷️
Find named neurons

Search for the INHIBITOR that handles negation, or the CONTROLLER that manages writing style, or other neurons with known roles.

🛣️
Follow a highway

Trace how a single signal travels vertically through all layers, branching forward and backward to show its influence.

5
📊 You build your own atlas

You download the tool, point it at a language model on your computer, and it creates a complete visualization file you can explore offline.

✨ You understand how AI thinks

You now have a tangible, visual sense of how millions of calculations flow through an AI model to produce answers, making the invisible visible.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 14 to 14 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is llm-neuron-atlas?

This is a browser-based 3D visualization tool that lets you explore the internal structure of a transformer LLM at the individual neuron level. You navigate a 36-layer tower, clicking on neurons to see their connections, tracing residual "highways" as signals propagate through layers, and comparing conservation across different model architectures like Phi-3 and Mistral 7B. The interface runs entirely in HTML and JavaScript using three.js, with no build step or server required beyond a simple static file host.

Why is it gaining traction?

Existing LLM visualization tools show flat 2D slices of model internals. This is the first tool that renders an entire multi-billion parameter model as a navigable 3D space. The hook is the "Google Maps for transformers" metaphor: you can zoom from a 73,000-foot view of all 36 layers down to a single neuron, see its top-10 outgoing connections, and follow named pathways like the INHIBITOR (dim 715) that fires on negation. The live demo loads in seconds and works on modest hardware, since the extraction pipeline runs on CPU only.

Who should use this?

ML researchers and interpretability enthusiasts who want to develop intuition about how information flows through transformer layers. Engineers verifying Mercury Paper B claims about single-dim control mechanisms. Anyone curious about whether structural patterns (like the CONTROLLER pathway) are conserved across model families. It is less useful for production workflows or debugging models.

Verdict

The concept is compelling and the demo is genuinely impressive for a 14-star project. The credibility score of 0.85% reflects its early stage, thin documentation, and lack of community validation. Try the live demo to see if this approach fits your research workflow, but do not depend on it for anything mission-critical yet.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.