LLMs extract high-level semantic concepts from SVG and ASCII art

1 comments

• neuronerd1 a day ago

  Anthropic interpretability researchers find neural representations of the semantic components of text-based visual depictions.