Post

Did Astra improve the SVG structure too, or mainly the visual result in terms of editability and path complexity?

That’s encouraging. Clean, well-commented SVG structure matters almost as much as the visual result if the asset is meant to stay editable.

Aw no, it's getting more boring and less funny... Ask it to do an impression of GPT-3.5's style of pelican

Too much training and it's no longer outsider art. Just like in real life :(

Well, there are biases in human art, when the subject isn't facing forward. Paintings tend to have the subjects facing left. Coins by contrast strongly favour right when there's no protocol about how people should face. For left-to-right readers, facing right means "moving forward"

Been using Astra this morning. Finding that it’s useful at one area most models struggle— good overview diagrams. The other models lack the metacognition and the diagrams are always literal. Astra seems to be able to step back and summarize better.

FYI: Link to pelican transcript in your post is missing. Observation: Pelicans for Astra are more consistent in the drawing choices from low->high. Arbitrary details like position of the sun, number of tail-feathers, shape of wing/arm is the same, but for other models no consistency between levels.

It _feels like_ Astra remembered choices it made at lower reasoning levels (assuming those ran first) or Astra just makes more consistent artistic choices when approaching the task from scratch.

If there is one thing this comparison proves is that the 5.6 Terra model should not exist. In almost every level it fails even Luna in basic expectations.

It's funky that they each have, like, a *style*. (Which incidentally suggests they're not spitting them out from training set memory, or at least not only.)