What's this about?
Meta's Muse Spark 1.3. Simon Willison ran his standard test immediately: an SVG pelican riding a bicycle, at 4.2 cents and 38 seconds, and he judges it a clear improvement over 1.2.
Why does the same benchmark keep appearing here?
Because it is longitudinal and unfakeable. The same prompt, run on every model for years, produces a comparison no leaderboard offers — and because it is absurd, it is not in anyone's training data as a solved problem. The thread also surfaced something genuinely interesting about it: the pelicans almost always face right, on flat ground, in 2D side view, without any of that being specified. One commenter links a study finding that pelicans on bikes always face right while other animals on other vehicles sometimes face left.
What's the community saying?
HN 411 points, 274 comments
- The composition question — why do they all look the same? — is the thread's best sub-discussion, with a cultural explanation offered (left-to-right reading makes "right = forward" feel natural) and an actual study cited.
- Willison's own note: the flat 2D framing feels reasonable for SVG, which implies vector illustration.
- Almost none of the thread is about Meta or the model, which is its own comment on how these releases land now.