Research Engineer
Atul Saxena
I build the environments models are trained in, and the tools that show what they learned.
- RL Systems
- Interpretability
- ML Infrastructure
Research Engineer
Atul Saxena
I build the environments models are trained in, and the tools that show what they learned.
Watch a language model think, token by token
Hooks GPT-2’s residual stream at layer 8, runs each token’s hidden state through a sparse autoencoder, and streams the top-K firing features into a live 3D force graph in the browser.
Mechanistic interpretability results normally arrive as static figures in a paper. You can read that a feature fired without ever developing an intuition for what the activation pattern looks like as text is actually being generated.
status: Active · research tool