I understand prose and code and raster images, but how does SVG work? The SVG structure and visual result seem so disconnected, I would have thought current LLMs would have a harder time on arbitrary descriptions of imagery.
Right, but rendering the visual understanding is based on neighboring pixels normally, which is much more directly related to the visual than plaintext like SVG tags.
I kind of disagree, because describing an image is more like writing an svg than rendering pixels.
If I ask you to describe how would you draw a cat head, you could do it in text like: "a big circle as the head, 2 small circles as the eyes, 2 triangles for ears, 3 lines on each side of the mouth as moustaches, etc..."
Hmm, sure. I'm still surprised - it also has to say where they are in coordinate space. It feels like the way genAI works ("what comes next") is not amenable to this use case (demonstrably I'm wrong, of course).
I remember when I was learning WebGL/OpenGL and had to draw some test shapes to test my shader, I would manually think what the vertex positions should be to draw a triangle, a pyramid, etc. I think for AIs it's quite easy, because most are quite decent at math, and have been trained on many geometry problems, and probably also a lot of OpenGL code too, and 3D assets.
I like how the Gemini 3.5 Flash (medium) one[0] added a ladder, so the hamster can get on the table.
[0]: https://aibenchy.com/showcase/#showcase=eb4878dbff331c67
Not sure how I feel about Grok... [0]
[0]: https://aibenchy.com/showcase/?q=grok
I understand prose and code and raster images, but how does SVG work? The SVG structure and visual result seem so disconnected, I would have thought current LLMs would have a harder time on arbitrary descriptions of imagery.
Svgs are just shapes, they (models) mostly only use sqaures, circles, etc. Rarely paths.
Plus most models are multi-modal, so they have a visual understanding too.
Right, but rendering the visual understanding is based on neighboring pixels normally, which is much more directly related to the visual than plaintext like SVG tags.
I kind of disagree, because describing an image is more like writing an svg than rendering pixels.
If I ask you to describe how would you draw a cat head, you could do it in text like: "a big circle as the head, 2 small circles as the eyes, 2 triangles for ears, 3 lines on each side of the mouth as moustaches, etc..."
Hmm, sure. I'm still surprised - it also has to say where they are in coordinate space. It feels like the way genAI works ("what comes next") is not amenable to this use case (demonstrably I'm wrong, of course).
I remember when I was learning WebGL/OpenGL and had to draw some test shapes to test my shader, I would manually think what the vertex positions should be to draw a triangle, a pyramid, etc. I think for AIs it's quite easy, because most are quite decent at math, and have been trained on many geometry problems, and probably also a lot of OpenGL code too, and 3D assets.
the DeepSeek V4 Flash is impressive for the cost
I use V4 Flash as my main model, it really is exceptionally capable for the price.
It can follow instructions quite well, if you have a plan, it usually executes it to completion and you get code that works.
My only concern with it, is that it's mostly served by Chinese companies on OpenRouter.
Sorry, accidentally posted the Romanian version, lol, here is en: https://aibenchy.com/showcase/hamster-playing-table-tennis-s...
If mods could update the link please.
If I were English, I would not love the US flag beside "English".
Well, it is US English I think, I could add UK English too.
If you were a former colony, it would be kinda funny ;)
What is this?
If the title is anything to go by, SVG’s of a hamster playing table tennis.
A visual way to compare model outputs
AI Slop.