Large language models succeeded before world models, and that needs explaining. By raw information volume an image carries far more than a paragraph of the same storage size — pixels, spatial relations, texture — and video carries more still, embodied interaction the most. If intelligence scaled with how much information goes in, vision and embodiment should have arrived first. The opposite happened.
Our claim: what determines intelligence is not the volume of information but its quality. Every symbolic use in a language has been filtered by billions of speakers over centuries, and the three properties that survive that filtering — low entropy, density, self-consistency — make language the highest-quality information carrier humans have produced. Pretraining a large model is, in effect, the largest low-entropy pretraining run in history.
The framework catches four things at once: why vision and video models lag behind language models, why AI developed in the order language, image-text, video, 3D and embodiment, how to re-read the Chinchilla scaling laws, and what emergent abilities actually are. We also ran the one test we could: a cross-modal contamination propagation experiment, 2 vision-language models by 3 tasks by 4 conditions by 10 repeats, 240 calls in all, asking which channel a model trusts when text and image disagree. The result splits: one model follows text 96.7% of the time, consistent with the prediction; the other follows text only 33.3% of the time, in the opposite direction. The experiment therefore does not validate the central claim; it leaves an anomaly to be explained.
This is a theoretical construction, not an empirical report. The boundaries, stated plainly: we do not give the functional form of information quality, we do not compare information quality across modalities quantitatively, and the emergence threshold is characterised only qualitatively. The experiment covers two models with ten repeats per condition, so percentages resolve to 3.3 points. More importantly, the two models disagree in direction, so the experiment does not establish that text is more trusted than image.
Keywords: Large Language Models, Low-Entropy Information, Nature of Language, Emergent Abilities, Multimodal Learning, Intelligence Theory