LLMs extract high-level semantic concepts from SVG and ASCII art (transformer-circuits.pub)

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