LLMs extract high-level semantic concepts from SVG and ASCII art (opens in new tab)

(transformer-circuits.pub)

3 pointsneuronerd15mo ago1 comments

1 comments

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

j / k navigate · click thread line to collapse

(transformer-circuits.pub)

3 pointsneuronerd15mo ago1 comments

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

j / k navigate · click thread line to collapse