Hacker News Reader

Comment by simonw | Hacker News Reader

I tried my "Generate an SVG of a pelican riding a bicycle" prompt against Gemma 3n 7.5GB from Ollama and 15GB for mlx-vlm and got a pleasingly different result for the two quantization sizes: https://simonwillison.net/2025/Jun/26/gemma-3n/

View on HN

So interesting, in the end it (accurately) describes the image. SVG is hard. Made me think we could try introducing a feedback loop until it reaches a close enough representation of what’s asked.

Is that actually a useful benchmark, or is it just for the laughs? I've never really understood that.

For me, it shows if LLM are generalising from their training data. LLM understand all of the words in the prompt. they understand the spec for svg better than any human. They know what a bird is. They know what a bike is. They know how to draw (and given access to computer-use could probably ace this test). They can plan and execute on those plans.

Everything here should be trivial for LLM, but they’re quite poor at it because there’s almost no “how to draw complex shapes in svg” type content in their training set.

It was supposed to be a joke. But weirdly it turns out there's a correlation between how good a model is and how good it as at my stupid joke benchmark.

I didn't realize quite how strong the correlation was until I put together this talk: https://simonwillison.net/2025/Jun/6/six-months-in-llms/

Always loved this example, what do you think of ASCII art vs SVG?

Since it's not a formal encoding of geometric shapes, it's fundamentally different I guess, but it shares some challenges with the SVG tasks I guess? Correlating phrases/concepts with an encoded visual representation, but without using imagegen, that is.

Do you think that "image encoding" is less useful?

It's a thing I love to try with various models for fun, too.

Talking about illustration-like content, neither text-based ASCII art nor abusing it for rasterization.

The results have been interesting, too, but I guess it's less predictable than SVG.

I've had disappointing results with ASCII art so far. Something I really like about SVG is that most models include comments, which give you an idea of what they were trying to do.

Yes, the comments part makes sense, you also included it in the talk (I read the transcript but forgot to mention it in my comment, sorry :)

It makes sense, since it works adds associations between descriptions and individual shapes / paths etc., similar to other code.

No way *wink*, our DevRel don't push for a good outcome on this one test case to get positive coverage on the top independent blog read by LLM people!

https://simonwillison.net/2025/May/20/google-io-pelican/

I think in 5 years we might have some ultra-realistic pelicans and this benchmark will turn out quite interesting.

And then the author will try the "Pelican tries to swallow the capybara as-is". And it will fall apart again.

It's this part where it gets interesting... how exactly it falls apart :D

like a quack

It’s been useful though given the authors popularity I suspect it’s only a matter of time new LLMs become “more aware” of it

It's useful because it's SVG so it's different than other image generation methods.

Given how primitive that image is, what's the point of even having an image model at this size?

This isn't an image model. It's a text model, but text models can output SVG so you can challenge them to generate a challenging image and see how well they do.

>Multimodal by design: Gemma 3n natively supports image, audio, video, and text inputs and text outputs.

But I understood your point, Simon asked it to output SVG (text) instead of a raster image so it's more difficult.

It can handle image and audio inputs, but it cannot produce those as outputs - it's purely a text output model.

Yeah you're right. Also, you're Simon :)

Reply on news.ycombinator.com