transformer-universal-turing-machine-simulation
IN premise — summaries/2026/08/24/wiki-Transformer_deep_learning_architecture-chunk-6-chunk-2.md
Created 2026-08-24T17:11:26+00:00
Phuong & Hutter (2022, arXiv:2207.09238) proved that transformers with finite precision can simulate universal Turing machines, linking the architecture to theoretical computer science.
Summary
Transformers are not just clever pattern-matching tricks; in principle, they can perform any computation a digital computer can, meaning the architecture itself is as general-purpose as a CPU. This gives the system a theoretical floor on what transformer-based reasoning can express, so limitations seen in practice come from finite precision, depth, or training rather than from the architecture being fundamentally incapable.