In this article, the author takes us on a journey inside the mind of R1—a neural model—by animating its thought process as it tackles questions like "Describe how a bicycle works." The clever approach involves saving chains of thought as text, transforming these into embeddings with the OpenAI API, and then sequentially plotting them using t-SNE. Along the way, the visualization reveals intriguing patterns in how R1 "thinks": from an initial dynamic "search" phase with large jumps between thoughts, through a steady "thinking" phase, and finally a "concluding" phase. As one insightful line from the article puts it, "By default we calculate cosine similarity between the embeddings and normalize across the set of all consecutive steps to 0, 1." The article even provides practical instructions, including a script called "pull_cot.js" for easy data extraction.

Key Points:

• The thought process of R1 is visualized by converting chains of text into embeddings.
• Embeddings are generated using the OpenAI API and plotted with t-SNE to reveal thought dynamics.
• Visualizations include graphs for consecutive distances (jumps between thoughts) and aggregate distances, highlighting different "thinking" phases.
• A direct quote underscores the methodology: "By default we calculate cosine similarity between the embeddings and normalize across the set of all consecutive steps to 0, 1."
• The analysis shows distinct phases: a rapid “search” phase, a stable “thinking” phase, and a final “concluding” phase.
• The article provides usage examples with thought-provoking prompts and instructions for running the visualization scripts, including the "pull_cot.js" script.
• Practical installation steps are outlined using a Pipfile and a run.py function to bring the visualization to life.

This summary introduces the innovative technique of rendering a neural network's internal thought process into visually engaging, interpretable graphs—a fascinating peek into the computational mind of R1!
Link to Article