#transformer architecture

The 70-Year-Old Test That Breaks Every LLM

A psychology test from 1935 just exposed a fundamental flaw in transformer attention. GPT-4o went from 91% accuracy to 15%. Here's what that actually means.