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

transformer-circuits.pub

3 pointsneuronerd11 comment

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