AI Alignment and Mecanistic Interpretability researcher. Interested in building agentic systems for real-world problems

Chetan Talele  • Peerlist

<ul><li>NeuralDoodle uses intermediate activations of an LLM to understand how information flows through its layers.</li><li>Given a text prompt and a selected layer, it extracts the post-attention residual stream activations, projects them into the vocabulary space, and identifies the top 5 most probable next tokens along with their probabilities.</li><li>Additionally, the project provides an interactive UI for selecting a layer and generating images based on the top 2 predicted tokens for a specific layer.</li><li>So,you can basically see what an LLM thinks is the next possible token after every layer of input processing.</li></ul>