Show HN: LLM Thought Visualization (github.com)

A new tool called Subtext, developed independently, has been introduced as a real-time instrument for visualizing the internal workings of language models during live conversations. This innovation stems from recent research by Anthropic, which identified a subset of representations within language models—the J-space—that functions like a global workspace for reasoning and verbalization. Subtext leverages a method called the Jacobian lens to provide continuous visual feedback on the model's internal states as it reads and responds, allowing users to observe how several tokens of reasoning are formed before the model generates its output. This tool is significant for the AI/ML community as it bridges the gap between internal model processing and observable output, enhancing our understanding of language models' decision-making processes. By rendering intermediate computation steps, such as judgments and planning, Subtext allows for a more granular examination of how models like Qwen3.5-4B handle complex queries. Its design enables higher engagement and analysis during conversations, making it a notable advancement in interpretability tools for AI. While Subtext has some limitations, such as only displaying single vocabulary token concepts, it promises to enrich the discourse on AI reasoning and model transparency.