Comment by simonw
8 hours ago
This is fantastic
I've been casually spot-checking other animals in other vehicles, because my absolute dream situation here is to catch an AI lab that's demonstrably better at pelicans on bicycles than other combinations.
Catching a lab cheating specifically on my one dumb benchmark would be really funny.
Dylan's methodology here - generating 1008 SVGs across an 8x6 combination - is significantly more robust than anything I was considering.
His conclusion:
> Nothing jumped out at me. I couldn’t find a case where the pelican-bicycle images looked noticeably better than the rest of that model’s grid.
What if they’re not pelicanmaxxing, but svgmaxxxing in general?
Because otherwise using a LLM to generate complex svgs is pretty niche and what I thought made this a good benchmark when it was new - generalized programming and spatial knowledge.
Obviously image gen in svg format is not a particularly hard problem if tackled directly on its own.
But that's a genuine worthwhile capability. It's like benchmarkmaxxing on a weightlifting competition by getting really strong.
Yes, but the original purpose of the benchmark (simonw, please correct me if I'm wrong!) was to test whether new models were good at novel problem solving. Things they haven't been trained on. So yes, getting better at generating SVGs is great news (and it seems they have been) but this particular benchmark still strikes me as largely worthless now, unless SVGs happen to be what you care about in particular when a new model drops.
Instead, it feels like a more appropriate benchmark for the original purpose would be to come up with new, novel problems each time, and compare across all models (including previous ones).
Throwing my hat in the ring: Generate a pelican shaped crossword where all the clues are related to bicycles.
(Haiku 4.5: https://imgur.com/a/N112Nxo, I'm trying some others but it's very slow! Opus has been at it for about 20 minutes.)
2 replies →
https://www.youtube.com/watch?v=jgYYOUC10aM
reminds me of this Key and Peele skit
How useful actually is this? It generates SVGs of pelicans on bicycles, sure, and some of them are (almost) spatially correct. But, none of them look good.
AI image generation suffers from this more generally. You can generate pictures of pelicans, sure. Newer models clearly generate images with more pelican-ness than before. But all of it is still uglier than sin. Drawing things accurately is one thing, making results that someone might actually want to use (without embarrassing themselves) is something else.
8 replies →
The fear is that the SVGmaxxing is limited to "X doing Y". If such 'template maxxing' exists, it will break for other templates, e.g. "X not doing Y", "X and Y doing Z", "X doing Y doing Z", etc.
2 replies →
well that holds IF svgmaxxing is 100% "code-writing-maxxing"
...which.. hmm I dunno if they are same or not
7 replies →
If they're optimizing for SVG generation, then that's an excellent outcome in my opinion. Vector images shouldn't be "pretty niche".
Please it's almost my primary bench for drafting understanding.
I'm almost about to post that xkcd 810
Gemini have absolutely been SVGmaxxing. They've openly talked about it.
Then it's great. A year ago I couldn't get any AI to draw a simple company logo in SVG or even convert from raster. No doubt the Pelicans put pressure on the labs to fix the awful SVG situation. Now I can even make a decent Peli in 3D.
Not an expert on this, so won’t speculate regarding what traits would create robustness specifically attributed to SVG visual representation capability, but felt you might find this paper on reasoning models trained on physical world video data becoming better at general reasoning interesting:
https://arxiv.org/abs/2210.05359
Then they’re legitimately getting better at svg which is a valuable tool.
Addressed in the article, in case you're curious.
Addressed by three inconclusive sentences. When someone goes into deeper detail on the topic, we shouldn't assume they failed to read the article just because they said "what if" instead of mentioning those three sentences.
Well, it’s mentioned as a limitation of the analysis, very much not ruled out (or in.)
That simonw is causing labs to do extra fine-tuning runs for this seems highly probable :)
I think svg is a balanced test because of the level of indirection and the required 'conceptualization' of physical elements then expressed through code.
Funnily enough, not that niche, because I have tried many times to do it as part of a wider project.
SVG are really useful you can create images that don't have the AI look
I agree, other formats, both textual and binary should be tested.
> What if they’re not pelicanmaxxing, but svgmaxxxing in general?
Mission fucking accomplished. https://xkcd.com/810/
Simon I hope from this day hence, your bio always includes:
"Simon Willison, among other things, is an advocate for the inclusion of pelican geometry in LLM training datasets."
“Today, the NASDAQ fell by 18% moments after Simon Willison published his latest blog evaluating the Legend 7.4 model that rendered an animated SVG of a pelican on a bicycle and the pelican fell off”
I asked Gemini 3.6 Flash for a pūkeko on a unicycle yesterday and the results were very stylish. The thinking notes were cool too, making sure it got all the high points of what it means to be a pūkeko (distinctive swamp hen found in NZ and Australia for anyone not familiar) and then figuring out how the legs would work with the pedals.
I worked with image analytics years ago. In 2024 and cynical about LLM image capability, I tested different models to create "Two female octopi in a bar each having their own drink. One is wearing a yellow hat. The other is wearing a red hat. They have no human features." It went badly I think because the training visual training data associated with the word "female" was overwhelmingly biased in quantity and obviated the 'no human features' instruction even after iterated instructions to remove.
So I tried "Make an image of a Djibouti cab with a camel sitting in the passenger seat. Give the camel no human anatomical features." Still bad.
Two years later the graphics and perspectives of the tools are so much better. But that isn't the big story. What really stands out is the models are no longer adding human female anatomies to camels and octopi. The gates and filtering based on instructions have matured enough so that the pelican/bicycle deductive reasoning puzzle is less problematic. But until an LLM anticipates something like impressionism from Parisian artists rebelling against the rules of the French Academy, continue to reserve a place for human artistry.
tl;dr simonw, your spot-check nailed it just as well as any extensive methodology. LLM's no longer need to cheat this part of the test (better to hack the question than the tool).
What if they’re just unsuccessful at it and something about a model grokking how to create a realistic visual representation of a pelican on a bicycle ends being key to the next OOM of capability unhobbling.
And since we have established this silly routine once, I must keep going and ask - yes, you have quite a collection of bona fide pelicans you’ve seen and photographed.
Have you physically seen all the other animals you’ve evaluated as well?
I was born in Soviet Russia, so trust but verify and if still around, maybe there pelican make LLM draw Simon ride bicycle and notice Python code improve.
It's a shame you started with pelicans on a bicycle and not snakes on a plane.
It's better to be original
> Catching a lab cheating specifically on my one dumb benchmark would be really funny.
Similar thing happened when TPC came up with SQL benchmarks.
If you're not good at TPC, your engineering team is no good.
If you're good at TPC, then (as a customer) we will actually include you in a bake-off benchmark for our specific problem.
Winning on it is the price of admittance into the game, especially in a crowded market.
But how narrowly you benchmarket matters, you can't just hard-code that specific scenario & not fix anything adjacent while you're at it.
For example when it comes to GPUs, the "Quack3" (sic) benchmark on ATI cards comes to mind.
Yea, I worked for a competitor to ATI back in the day and we were definitely Quakemaxxing. We were not so unethical as to try to detect the .EXE name so to put the GPU into a "cheating" mode only when that benchmark was run, but we did run Quake 3 pretty much constantly while trying to eke out a few more FPS...
Wasn’t that almost the norm for both new OS drivers and subsequently GPU SoC firmware of ‘gamer cards’ to literally optimize for the latest AAA titles? I feel in that case, the interests aligned - in a world where I had a chance to actually play Crysis, I wouldn’t care why the frame rate was decent, no?
I think a more fundamental test is SVG art creation in general. Perhaps a pipeline to take any image, caption it, ask the LLM for an SVG, rasterize to an image, and finally either use a deterministic visual similarity check or ask another LLM to be the judge and score how close the SVG is to the original image.
Fidelity to the original is definitely not how humans would measure "art" in this context.
True, maybe we can call the generated SVG something else than art.
Perhaps also vary the bird? Wikipedia tells me pelicans are in the order _Pelecaniformes_ so shoebills or herons might do.
> I've been casually spot-checking other animals in other vehicles
Snakes on a plane, weasels on a diesel, spiders on a glider, baboons on a balloon, goats on a boat.
[dead]