Tag: research

All the articles with the tag "research".

• Claude Flags 'This Is a Test' 26% of the Time and Says So Under 1%

  25 Oct, 2026

  Anthropic's Natural Language Autoencoders read evaluation awareness out of Claude's activations at rates 20x higher than Claude mentions it out loud. The team's flashier global-workspace claim is the contested one.

• Three Ways to Type an Em Dash. The Machines Need None of Them.

  30 Jun, 2026

  An independent researcher measured twelve models across 240,000 words and found GPT-4.1 using em dashes at 3.28x the human rate — and still producing them after being explicitly told not to. The number everyone quotes is the least interesting part.