09Catho

09Catho / axon

Public

Real-time 3D visualisation of SAE feature activations inside GPT-2, token by token

3d-visualization deep-learning fastapi gpt2 llm-interpretability
10
0
94% credibility
Found May 20, 2026 at 10 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
JavaScript
AI Summary

AXON is a research visualization tool that lets you watch a text-generating AI think in real-time. When you type a prompt, the AI generates words one by one—and a 3D interactive graph shows which concepts inside the AI are activating at each moment. You can click on glowing nodes to learn what each concept means, explore how different prompts trigger different patterns, and watch the AI's 'thought process' unfold as colorful, breathing connections in space. It's designed for understanding how large language models work internally.

How It Works

1
🔍 You hear about a way to watch AI think

A friend tells you about a tool that visualizes what concepts an AI activates as it writes text.

2
📦 You install the tool in one command

You download a small package and run an install script. The computer sets everything up automatically.

3
🚀 Everything starts with one click

You run the launcher script. The AI model loads in the background and your browser opens to a waiting canvas.

4
✍️ You type any prompt you like

You enter a sentence, phrase, or question into the sidebar. Maybe something about France, or mathematics, or a poem.

5
Watch the magic unfold
🔢
For factual prompts

Geographic, historical, and logical patterns light up on the graph as the AI responds with facts.

💻
For creative prompts

Different patterns emerge—language, emotion, and storytelling concepts glow on the graph.

6
🔎 You click a node to explore

Curious about a concept? You click any glowing node. The camera flies to it and shows you what it means.

✨ You see the AI's mind at work

You watch concepts connect and fade as text flows. What was once invisible becomes a living, colorful map of thought.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 10 to 10 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 axon?

Axon is a real-time visualization tool that shows you what concepts inside GPT-2 are firing as it generates text. The JavaScript frontend renders a live 3D force graph where each node represents a Sparse Autoencoder feature (pulled from Neuronpedia with human-readable labels), and edges connect features that co-activated. You type a prompt, hit Generate, and watch the graph light up with glowing nodes breathing and fading as tokens stream in, with a sidebar panel showing the top active concepts alongside the output text.

Why is it gaining traction?

Mechanistic interpretability has moved from academic papers into developer tooling, and axon puts a tangible window onto what a model is "thinking." The hook is the immediacy: instead of reading activation heatmaps post-hoc, you watch concepts activate in real time as the model reasons. GPU acceleration is available for RTX 3090 (~18ms/token) or CPU fallback for Apple M2 (~400ms/token), and the WebSocket API lets you drive generation from any client. The ability to inspect individual nodes and see co-activating neighbors gives researchers a fast way to hypothesize about circuit behavior without heavy instrumentation.

Who should use this?

Interpretability researchers debugging whether specific concepts (geography, code syntax, sentiment) live where they expect. ML practitioners who want a visual gut-check on model behavior before deploying. Educators teaching mechanistic interpretability who need a live demo that makes the abstract concrete. Not for production inference pipelines—this is a research/education tool, not a serving layer.

Verdict

Axon delivers a genuinely novel visualization for anyone curious about LLM internals, with a clean architecture (Python backend, Three.js frontend, WebSocket streaming) that "just works" on startup. At 10 stars the project is young and unproven at scale, and test coverage beyond the benchmark script is thin, so expect to read the source when something breaks. That said, with 1842 cached Neuronpedia labels and a model-agnostic design, it is a credible starting point for exploring SAE-based interpretability right now.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.