Claude Can Plan a Rhyme Before It Writes the Line #Shorts

A language model writes one token at a time, but its computation can prepare for words further ahead. In selected Claude 3.5 Haiku poetry experiments, Anthropic researchers found signals for candidate endings before the next line began. Changing those signals changed how the model built the line.

This is evidence of a planning mechanism in those cases, not proof of human-like thought or a complete explanation of Claude. The researchers used an approximate, more interpretable replacement model and tested interventions in the original. This is a 2025 study, not a newly announced 2026 finding.

Primary research:
https://transformer-circuits.pub/2025/attribution-graphs/biology.html
Methods:
https://transformer-circuits.pub/2025/attribution-graphs/methods.html
Independent coverage:
https://www.technologyreview.com/2025/03/27/1113916/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model/

Original simplified animations, not recordings of actual activations. Poem examples are from the published research.

#AI #Claude #MachineLearning #AIResearch #ScienceCorner

Видео Claude Can Plan a Rhyme Before It Writes the Line #Shorts канала The Science Corner