Georgia Tech's Transformer Explainer Runs Live GPT-2 in Your Browser
Georgia Tech's Transformer Explainer runs a live GPT-2 model in your browser, letting you watch every internal operation unfold as you type — no setup required.
Takeaways
- Transformer Explainer runs a live GPT-2 model in the browser with zero setup, showing every internal operation in real time.
- Users can type custom prompts and watch tokenization, attention weights, MLP activations, and output probabilities update live.
- Interactive controls for temperature, top-k, and top-p sampling let you directly observe how hyperparameters shift the output distribution.
- The tool has reached 125,000 users, ~8,000 GitHub stars, and was accepted to CHI 2026.
- Built with Svelte and D3.js on top of Andrej Karpathy's nanoGPT, converted to ONNX Runtime for client-side execution.
- MIT-licensed and open source; forkable for building explainers for other architectures.
Transformer Explainer is an interactive, browser-based visualization tool from Georgia Tech's Polo Club of Data Science that lets you watch a real GPT-2 model think in real time. Type any text prompt, and the tool animates every step of the transformer pipeline , tokenization, embeddings, multi-head self-attention, MLP layers, and final output probabilities , as they actually execute. No Python environment, no GPU, no sign-up required.
The gap it fills
The Transformer architecture underpins modern large language models, yet its complexity makes it difficult for non-experts to learn. Existing resources often lack interactivity, rely on static descriptions of simplified architectures, or fail to reflect models' behavior with real data. Blog posts and static diagrams can explain the theory, but they can't show you how attention weights actually shift when you change a single word in your prompt.
To address this gap, the team built Transformer Explainer as an interactive visualization tool. It integrates an overview of the Transformer's data flow with on-demand explanations that gradually reveal mathematical details, and smooth transitions across abstraction levels highlight the interplay between high-level structures and low-level operations.
A live model, not a simulation
The key design choice here is that nothing is faked. Running a live GPT-2 instance directly in the browser, Transformer Explainer empowers learners to experiment with custom input and hyperparameters without setup, observing next-token predictions in real time. The model is derived from Andrej Karpathy's