---
title: "problem with mathematically proven claims about llms"
source_url: https://webdirections.org/blog/the-problem-with-mathematically-proven-claims-about-llms/
ingested: 2026-05-08
sha256: pending-rehash-2026-06-01
review_value: 8
review_confidence: 8
review_recommendation: strong
review_stars: 4
source_feed: TLDR AI (newsletter)
source_published: 2026-05-07
type: raw
created: 2026-05-10
updated: 2026-05-10
tags: []
tags: [raw-status:stub]

---
# The Problem with "Mathematically Proven" Claims About LLMs
John Allsopp 批判三类被误用的"数学证明"LLM 局限性声明：1) AI 不能自我改进（Zenil 论文的 KL-flow 收敛不适用于有外部验证器的系统） 2) 幻觉不可避免（定义过于宽泛，论文明确说不适用于 knowledge-enhanced LLM） 3) 数学天花板（Hartmanis-Stearns 定理不适用于多 LLM 协作+工具使用）。共同模式：取最大版本声明→证明→推广时丢弃假设→修饰性语言包装。