Show HN: LLM-Visualized – Interactive 3D and 2D Visualization of GPT-2
TL;DR
A developer built an interactive 3D and 2D visualization of GPT-2 Small (124M parameters) showing real activations and attention scores from an actual forward pass.
Key Points
- The 3D view uses Three.js; the 2D view is plain HTML/CSS/JS with no framework dependencies.
- The goal is to make transformer internals visually tangible without drowning in math.
- It is an open-source learning tool, not a commercial product.
Nauti's Take
Choosing GPT-2 as the teaching model is smart – old enough to be fully documented, relevant enough to still matter. The combination of real activations instead of synthetic examples with a 3D view is unusual and deserves attention.
Whether Three. js is the right foundation for serious interpretability tooling is debatable, but for a first visual impression it works well.
Anyone learning Mechanistic Interpretability will find this a solid starting point, better than most explainer videos out there.
Sources