Figures as Interfaces: Toward LLM-Native Artifacts for Scientific Discovery

The project addresses a critical bottleneck in AI-assisted science: the 'lossy' conversion of data into static images that LLMs then struggle to manipulate. While theoretically sound, the project currently exists as a fresh research implementation (0 stars, 7 forks, 1 day old). It faces extreme displacement risk from frontier labs; Anthropic's 'Artifacts' and OpenAI's 'Canvas' are already implementing the UX paradigm of generative, interactive code-based outputs. Specifically, if a frontier lab enables a 'Scientific Mode' in their UI that keeps the data-visualization link live, this project's unique value proposition disappears. The moat is currently purely conceptual/academic. To survive, it would need to integrate deeply into established scientific workflows (e.g., as a Jupyter/VS Code extension) or provide a specialized library of 'scientific widgets' that go beyond standard Plotly/Matplotlib capabilities. Its defensibility is hampered by the fact that the underlying 'interactive' logic is usually just a thin wrapper around existing reactive programming patterns applied to LLM tool-calling.