Decoy Font, made by Eric Lu at Mixfont—a type foundry and AI research lab—is a typeface that exploits the hybrid image optical illusion. Each character overlays two letterforms: a blurry low-spatial-frequency shape in the background, and a crisp high-spatial-frequency outline in the foreground. Human eyes, which read in context and shift focus naturally, tend to land on the fuzzy background forms; a squint or a step back from the screen lets you slide between the two. Large language models, which process images by reading pixels at close range, reliably lock onto the foreground outlines instead. The result: one image reads "SLEEP EARLY" to a person and "BINGE SHOWS" to GPT-5.6 Sol or Gemini 3.5. Same letterforms, two audiences, two messages that know nothing of each other.

The beauty here is structural, not conceptual. What Decoy Font reveals—and makes literally usable—is the gap between two fundamentally different visual regimes: the embodied, context-sensitive attention of human seeing, and the pixel-literal, frequency-naive scanning of current LLM vision. This gap is not a metaphor. It is a physical and computational fact about how two systems process spatial information. The hybrid image as optical illusion is already-construct—a good sixty years of vision science, and the Marilyn Monroe / Albert Einstein composite image has been reproduced so many times it's become a classroom example. The mechanism is named, understood, exhausted. But using that mechanism to write simultaneously to two different readerships—human and machine—choosing your reader by choosing your spatial frequency—is a logic that has not yet been named. It is still growing. That gap between what is known and what is not yet named is the remainder.

In SAE terms: the chisel-construct cycle describes how every genuine chisel-out generates new remainders, which subsequent eras digest into construct. Decoy Font stands at the open end of this cycle. "Anti-AI art" is beginning to crystallize as a category—IMAGO proposes locally-trained AI models as an ethical alternative to scraped training data, Ghost Font hides messages inside moving animations—but the specific move of spatial-frequency address, embedding two readers inside a single glyph, has no complete conceptual shell yet. It has not been absorbed into any discourse. It is alive as remainder precisely because the art world is slowly approaching it without having fully arrived.

Lu says openly that the font may stop working "in just a few months" as models advance. This is not a disclaimer. It is the exact temporal shape of 余项之美. The remainder exists in the window before the chisel has fully struck—before the technology has closed the gap, before the theory has given it a complete name. Today's LLMs process images pixel by pixel, zoomed in, reading only high-frequency edges; they miss the low-frequency shape that human eyes, at normal distance, perceive as the primary message. That "miss" is real, physical, determined by architecture. When multimodal models learn to perceive the way human vision does—with context, with distance, with the capacity to defocus—the gap closes. The remainder disappears. What is now a live question becomes a footnote in a history of the AI transition.

Seeing it now means seeing the fossil record of this mismatch while it's still forming. Decoy Font is not an art manifesto. It is a downloadable TTF file and a web encoder, published as a research experiment by an AI researcher at a type foundry—positioned between typography, machine perception, and quiet resistance to the reach of LLMs into every act of reading. That positioning between disciplines, before any discipline has claimed it, is exactly where 余项之美 tends to appear: not in the gallery, not in the paper, but in the gap that is still too new to have a room of its own.

Han Qin (秦汉) · @hqinjarsy · SAE Judgment & Aesthetics