What's this about?
A developer's personal benchmark asks 14 models to draw an SVG of a frog with a Habsburg jaw — the famous protruding chin of European royalty. The trap: 7 of 14 models silently dressed the frog in crowns and royal robes even though the prompt named only an anatomical feature, revealing how models leak associated concepts into outputs.
What exactly is a "Habsburg jaw"?
The Habsburg dynasty ruled much of Europe for centuries and intermarried so heavily that a distinctive jutting lower jaw became hereditary — it's now shorthand for that anatomical feature. A model that truly understands the prompt draws a frog with a big chin; a model pattern-matching on "Habsburg" draws royalty. That gap between understanding and association is exactly what the test measures.
What's the community saying?
HN 122 pts, 56 comments
- Several argue it's better than the pelican test because it has a built-in trap: two models even admitted they were extrapolating royalty ("because Habsburg").
- Opus 5 "frogmaxxed" — consensus best output; commenters ranked the gallery like an art show.
- A genuine puzzle thread: how do LLMs even approach SVG drawing — reasoning about coordinates like a language, or something closer to learned visual priors?
- The recurring objection — "what does this measure, practically?" — met the usual answer: unsaturated, un-gameable tests beat leaderboards.