Text-Only Conservatism

From the AI Research Compass wiki

Text-Only Conservatism: Believes text alone carries enough of human knowledge to reach strong intelligence, so adding vision and other modalities is mostly surplus.

Beliefs

Text-Only Conservatism looks at the rush toward multimodality and asks a blunt question: what did the images add? Written language is humanity’s compressed record of the world, already abstracted, already labeled, already arguing with itself. Physics textbooks describe gravity better than a million videos of falling apples. A model trained on enough text learns about colors, spatial layouts and cause and effect secondhand, and secondhand is often enough. Pixels are expensive, redundant and mostly about textures nobody needed.

It sits at the far Language edge and in the upper half, aligned with one large model and the scaling program, but it refuses the multimodal expansion most of its neighbors embraced. Its paradigm level is Ultra-Orthodox: it defends the configuration that produced the first generation of strikingly capable language models and sees no need to complicate it.

Arguments

  • Text-only language models showed broad few-shot ability across tasks with no perceptual input at all, which is strong evidence that text carries a great deal of world knowledge.
  • Per token, text is far denser in information than images or video, so a fixed compute budget buys more knowledge from text.
  • Many multimodal models still reason mainly in their language backbone; vision often contributes recognition, not reasoning.

Critiques

  • World-side camps point to tasks such as spatial reasoning, counting and physical prediction where text-only models make mistakes no sighted child would.
  • High-quality text is finite, and the supply may run out before the knowledge it can teach does.
  • Users increasingly hand models screenshots, charts and photos; refusing modalities means refusing real work.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ −1.00, y ≈ +0.26, paradigm ≈ −0.93.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data