The Future of AI Interaction: From Low-Bandwidth Language to High-Bandwidth Vision

Currently, AI is primarily LLM-based, relying heavily on language. However, human reading speed is naturally limited—especially when processing a non-native language, where reading speeds slow down even further. Reading and listening are fundamentally low-bandwidth forms of communication.

In contrast, vision is the highest-bandwidth medium. The input bandwidth of the human eye is exceptionally high, allowing the human brain to compress super-high-bandwidth visual information into extremely low-bandwidth conceptual summaries and efficiently extract critical insights.

The future of human-AI interaction must make a breakthrough in the visual dimension. Since AI is excellent at writing code, we should leverage it to generate diagrams, charts, and visualizations that compress and summarize information effectively. Large walls of text generated by AI are simply too overwhelming for humans to parse quickly.

From Code Diffs to Visual Diffs: The Ideal Interaction Scenario

Ideally, AI should be able to present a “visual-to-visual comparison” rather than code-level red-and-green additions and deletions. The final output should not be endless paragraphs of text, but rather an interactive HTML dashboard containing bar charts, violin plots, and other intuitive visual elements.

This represents the natural evolution of AI as an intelligent collaborator: progressing from “code-level red/green diffs” to “visual and semantic diffs,” and moving from “linear text dialogs” to “interactive dashboards.”

Why Is the Current Chart Generation Experience Lacking?

While this scenario is technically achievable today, users frequently encounter issues like unstable chart quality, collapsed layouts, poor aesthetics, or missed focal points. These problems highlight the current boundaries and pain points of AI technology:

1. AI Writes “Blindly” (Lack of a Visual Feedback Loop): Current LLMs generate code based on logical probabilities; they cannot “see” the final rendered pixels. They do not know if the X-axis labels overlap, if the color palette is jarring, or if a flex container collapses under specific screen resolutions. They output code, not images.
2. “Focal Points” Are Tacit Knowledge: While data is objective, “what we want to express” is purely subjective. Data structures can be easily parsed, but key insights—such as a “fat-tail effect” or specific business perspectives—are tacit human knowledge that AI cannot easily guess.
3. Fragile Layout Systems: Frontend UI, especially complex dashboards, is highly interconnected; a single styling conflict can break the entire layout. AI’s stability in writing pure logic is currently much higher than its stability in managing global CSS states.

The future of AI interaction must resolve these “last mile” experience issues, enabling AI not just to write code, but to “see” and comprehend the visual layouts it generates.