Nature does the heavy lifting that pure deduction cannot.
A paper posted to arXiv argues that human mathematical reasoning has never been purely formal. The authors trace the history of tools like the Fourier transform, showing that at each key moment a physics problem — a vibrating string, a heat equation — forced mathematicians to accept ideas that strict logical derivation had failed to produce, or had actively resisted. Their broader claim: even decidable logical systems carry worst-case computational costs so large they are effectively unusable, making physics-inspired pattern matching a cognitive necessity rather than a shortcut.
The argument matters because it tries to put a theoretical floor under something the AI field has mostly justified empirically. If human-level mathematical creativity requires a vast store of cross-domain patterns rather than a deduction engine, that is a principled reason — not just a scaling bet — for why large language models trained on enormous corpora might behave the way they do. It shifts the conversation from "bigger models happen to work" to "bigger models work because of how cognition itself is structured."
The case is interesting, but the conclusion conveniently flatters the current state of the industry — a paper that ends by justifying the architecture everyone is already building should be read with at least one eyebrow raised.