---
date: '2024-01-30'
description: well, here we go again.
id: Turing-complete Transformers
modified: 2026-06-05 15:08:22 GMT-04:00
tags:
  - pattern
  - ml
title: Turing-complete Transformers
created: '2024-01-30'
published: '2024-01-30'
pageLayout: default
slug: thoughts/Turing-complete-Transformers
permalink: https://aarnphm.xyz/thoughts/Turing-complete-Transformers.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
<blockquote class="twitter-tweet" data-lang="fr" data-dnt="true"><p lang="en" dir="ltr">Turing Complete Transformers: Two Transformers Are More Powerful Than One<br>"We prove transformers are not Turing complete, propose a new architecture that is Turing complete, and empirically demonstrate that the new architecture can generalize more effectively than transformers."… <a href="https://t.co/LGVlZt0afu">pic.twitter.com/LGVlZt0afu</a></p>&mdash; Burny - Effective Curiosity (@burny_tech) <a href="https://x.com/burny_tech/status/1744100637187461455?ref_src=twsrc%5Etfw">7 janvier 2024</a></blockquote>



The idea is to combine two small [[thoughts/Transformers|transformers]] rather than one [[thoughts/large models]]

More specialised on given tasks, and prove to be Turing-complete?

![[posts/images/shogoth-gpt.webp|Shogoth as GPTs]]

> Speculatively, people might think GPT-4 without any guardrails _could_ pass the Turing-test. A more important question is “What is the Turing-test equivalent for pseudo-intelligence system?”

John Searle famously said:

<blockquote class="quotes"><p>Turing machine is not to be found in nature. They're to be found in our <em>interpretations</em> of nature.</p><p>John Searle</p></blockquote>

the [paper](https://openreview.net/forum?id=MGWsPGogLH) detailed what is very similar to the setup of [recursive language model](https://alexzhang13.github.io/blog/2025/rlm/), ideally we want to deal with long-horizon tasks more effectively and economically viable.