The Multimodal AI Revolution: From the Morse Code Trap to Visual Chain-of-Thought

Forum topic · ✨步子哥 · 2026-01-07

Summary

This zhichai.net forum post presents an infographic-style overview of advances in multimodal AI, arguing that converting continuous 4K image signals into discrete text tokens causes severe loss of geometric and physical information—likened to a "Morse code trap" where AI loses detail the more it reasons. It describes Chain-of-Visual-Thought (CoVT), which generates continuous visual tokens in latent space instead of relying on language, teaching models to reason through recognition, 3D relationships, structure, and semantics. It highlights Qwen3-VL's architectural innovations—Interleaved M-RoPE and Deep Stack Fusion—which address long-video understanding issues such as spectral bias and amnesia, reportedly achieving 100% needle-in-a-haystack accuracy. Finally, it frames embodied AI's shift from passive observers that merely recognize objects to active operators that understand affordances (what objects can do, such as being graspable or sittable), calling this a fundamental change in AI cognition rather than a mere version upgrade.

The Multimodal AI Revolution: From the "Morse Code Trap" to Visual Chain-of-Thought

This post is a visual poster-style explainer originally published on zhichai.net, covering recent developments in multimodal AI reasoning.

1. The Morse Code Trap

Forcing the continuous signal of a 4K image into discrete text tokens causes severe loss of geometric and physical information. The analogy: it's like listening to a symphony through a telegraph machine—the more the AI "thinks," the blurrier the details become.

Core pain point: lossy compression leads to a lack of physical intuition.

2. CoVT: Chain-of-Visual-Thought

Instead of relying on language, CoVT generates continuous "visual tokens" in latent space—teaching AI to "stop talking and draw" while reasoning. Key reasoning dimensions include:

Recognition
3D relationships
Structure
Semantics

3. Qwen3-VL: An Architectural Revolution

Qwen3-VL tackles the "spectral bias" and "amnesia" problems in long-video understanding through:

Interleaved M-RoPE (interleaved positional encoding)
Deep Stack Fusion

The poster cites a reported 100% needle-in-a-haystack accuracy for retrieving information from massive context streams.

4. The Future of Embodied AI

AI is evolving from a pure observer into a real-world operator. The key shift: moving from recognizing what an object *is* to understanding what it *can do* (affordances—e.g., graspable, sittable).

> "This is not a simple version upgrade, but a fundamental transformation in AI's mode of cognition."

*Note: The figures and claims above are as stated in the original poster; this translation preserves them without independent verification.*