Why aren't there more AlphaFolds?

According to LessWrong, AlphaFold stands uniquely as a deep learning model that transformed science—despite frontier LLMs being trained on virtually all human scientific output. Why haven't we seen more AlphaFolds? Author Nim Keivan argues the answer hinges on a fundamental difference: LLMs learn from human-generated text, which is secondhand knowledge, downstream of discoveries already made and published. But AlphaFold works directly with observations at science's frontier, where definitive answers don't yet exist in textbooks. To illustrate: in fourteen ninety-seven, Vasco da Gama's sailors contracted scurvy—swollen limbs, failing gums after months at sea. They recovered after eating local oranges but attributed it to the climate. Two hundred fifty years later, physician James Lind ran a controlled trial proving citrus cured scurvy. Yet he blamed "noxious sea air" and lack of fresh vegetables—not vitamin C, which he'd never heard of. When the Royal Navy finally mandated lemon juice in seventeen ninety-five, it worked through pattern-matching alone, not mechanism. Decades later, lacking lemons, they switched to limes with lower vitamin C, and scurvy returned. The insight: scientific breakthroughs come at the frontier—where observation precedes understanding. That's where AlphaFold operates. Most AI models, trained on the safe knowledge already written down, stay in familiar territory. Real scientific transformation requires models that can tackle problems where the answers aren't yet in print.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton