> How does the prompt “Generate an SVG of a pelican riding a bicycle” add up to 95 input tokens? OpenAI’s tokenizer counts 10, Anthropic’s counts 10 for Opus 4.6, 30 for Opus 4.7 and 25 for Sonnet 5/Fable 5. Prompting “hi” to Kimi K3 counted 86 tokens, suggesting there may be an 85 token hidden system prompt. It refused to leak it though.
It's incredible Simon still believes pelicans on bikes aren't part of the training set, despite hundreds of them on blogs, forums, and Github. Stuff we put in our company blog shows up known by LLMs 6 months later, and we have 1000x less traffic than Simon's own website
The pelicans are still all rubbish. If they make it into the training set it doesn't help the models produce better pelicans, if anything it will make them perform worse!
Respectfully, the pelicans used to be an unrecognisable mess and now they’re unquestionably pelicans on bicycles, rendered poorly, from every model.
In the same timescale, model capabilities across the board have only meaningfully improved in places where the labs are focusing their training efforts.
Moreover, they have a uniform style, even though your prompt doesn’t ask for one. There's no model going rogue and producing a watercolour of a pelican. They’re all rendered in an approximately uniform style, even though the svg format has a basically unlimited possibility space.
You know what, that's actually something I hadn't considered before. There's definitely a bias towards a pelican cycling from left to right on a red bicycle against a blue sky and green grass.
Blue sky and green grass aren't that surprising, but the color and direction are interesting.
When I finally build the proper gallery I'll throw in a few other creature-vehicle combinations, and track some characteristics like which direction, color of bicycle, general pelican geometry etc. It will be interesting to see if other creatures end up with coincidentally similar design choices or if that's unique to the pelican-bicycle combination.
There was a glorious moment when I thought that the Chinese models were more likely to produce right-to-left cycling pelicans, but sadly that trend didn't seem to hold up.
I haven't seen many AI works that produces a pelican on a bicycle done in a "Ligne Claire" style, for example.
I guess AI's narrows down the output probability space drastically and converge on some agreed upon aesthetics. Works great for computer programs but bad for art.
> the pelicans used to be an unrecognisable mess and now they’re unquestionably pelicans on bicycles, rendered poorly, from every model
You would not expect that to happen if the models trained on the unrecognizable mess, right?
> model capabilities across the board have only meaningfully improved in places where the labs are focusing their training efforts
And the labs clearly did focus on improving image rendering.
> they have a uniform style
SVG output from LLMs always looks like that. It looked that way from the beginning; no LLM ever produced a watercolor when asked for SVG output. They all render the prompted element centered in the picture. They all tend to draw things going from left to right, and so on.
I’m not suggesting Simon’s pelicans in the dataset are having a meaningful impact. I’m expecting that a company like ScaleAI has a product along the lines of “benchmax dataset: SimonW’s Pelican on Bikes test” which is a private curated series of well-drawn SVGs of animals riding vehicles for training and RL.
Simon - has no one told you about the Willison-Pelican Scaling Law?
```
if is_willison_pelican_blog_post:
[redacted]
```
You haven't seen their final form [1]
[1] final form is a frontend/react/let's not talk about it, library - it caused a great deal of PTSD to me and my previous company's team due to its dogmatic preference for "we use these axioms, end of story", over practical utility - so it was quite challenging to do state of the art tasks such as nested form fields (e.g. 'user.address.personal.line-1'). The PTSD it caused made us all block out the memories, I suppose. But - it had zero dependencies. That is what mattered. It kept us going. We weren't reaching for more. We had plenty of time.
And thank god for that. Because I'd forgotten my watch in California - and this was in Tokyo [2]
[2] a joke within a joke about Jensen's Kyoto gardener story. Beautiful story, drowned out by WatchGate memes. Why can't jokes have layers? Models have trillions. If you miss 100% of the jokes you don't make, make all the jokes. Someone will laugh (eventually, maybe?) Even if it's: "this person + comedy club = full secret service detail". If someone laughs at that - at my own expense? I don't mind. They laughed. I know this is a gibberish, off-topic message - it's also a human message. I just felt we need more such things in our lives these days.
PS: have you physically seen a pelican in real life? (not a joke)
> PS: have you physically seen a pelican in real life? (not a joke)
We have several thousand living 15 minutes walk from our house. I recently started adding my wildlife photography (from iNaturalist) to my blog, so I'm posting several new pelican photos a week at the moment: https://simonwillison.net/search/?q=pelican&type=beat%3Asigh...
I agree with that. I think, in particular, all the broken bike frames associated with "pelican on a bike" probably make it harder for LLMs to render correct bike frames.
Simon has stated a few times that he knows it’s possible that pelicans could be in the training sets. He also has other tests he doesn’t share publicly. He’s just a fan of pelicans.
From the article it doesn't even sound like he cares about pelicans at all, and doesn't think they are a good way to compare models anymore ... but people are used to seeing the test now, and it does serve as a common "hello world" unit of work.
the nature of the test was to see if the models can effectively compose an image of a novel concept outside the training set. If they are trained on it, it ceases to be an interesting test to some extent.
I would urge you to re-read the blog post you are commenting on. It pretty clearly explains how it is an interesting test independently of "see[ing] if the models can effectively compose an image of a novel concept outside the training set".
it's still interesting because there's no pelican-on-bike model, and if you're training a model well enough, then it should be obvious when a model has reached "AGI" or whatever.
More to it, the actual bloody companies are using them as a reference. Maybe it’s a 3d version, not an svg - but it clearly shows they’re on the radar of these companies.
Maybe it gets posted every time because besides a personal believe by the person popularising this "benchmark", there is no reason to assume that certain labs aren't intentionally training to game this and every other lab at least unintentionally gets improvements for this specific combination of animal and action because the internet is full of both good and bad examples, often ranked, which does inevitably become training data.
I have shared examples of certain models by certain labs doing far better on the pelican cycling vs other, similar prompts. Just operating on a feeling that labs don't optimise for this (as mentioned, even if they don't training data is filled with these) is not solid enough that criticism shouldn't be leveraged when it comes up.
> I have shared examples of certain models by certain labs doing far better on the pelican cycling vs other, similar prompts
Please share those again!
One of the things I'm most looking forward to is a lab producing a model that creates a really great pelican riding a bicycle and then a terrible sloth riding a skateboard (or whatever).
Evidence in the other direction (that they're able to generalize) is that I can't think of any LLM currently that can't create usable (placeholder) SVG icons, I tried a bit before the pelican became popular and it was abysmal.
Happy to, here one example where Grok 4 Fast, despite producing a fairly consistent pelican [0], did severely worse in a similarly outlandish scenario along with Haiku 4.5 and GPT-5 for context: https://news.ycombinator.com/item?id=45599403
> [...] a really great pelican riding a bicycle and then a terrible sloth riding a skateboard [...]
Happy to play ball. You made a blog post a few weeks back on one of the Qwen models with the eye-catching title "Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7" [1].
Here is what Qwen3.6-35B-A3B via Openrouter provided for a sloth riding a skateboard: https://imgur.com/a/Dy8fvR5
Like Grok 4 Fasts attempt at a mushroom in a rowboat, it is barely recognisable as anything despite both Qwen3.6-35B-A3B and Grok 4 Fast having no issue with more popular (i.e. benchmarked) examples. Whether this is a case of training data being unsanitized or intentional benchmark targeted training, I cannot say, but it is the case.
A massive delta in favour of Opus 4.7, despite the pelican Qwen3.6-35B-A3B produced being noticeably better as you rightly pointed out. What does that tell us?
I would not critique your use of this fun benchmark the way I tend to if I did not have evidence to back up my position, including private evals that I can reliably use to point out major deviations between what a models claimed performance is according to major benchmarks vs the actual performance outside these known test cases.
I will also say that while I have a lot to be critical of regarding Anthropics modus operandi, especially how they present interesting findings like their j-space work, which I found was irresponsibly anthropomorphic in their reporting, especially as this wasn't a first in model interpretability, but mainly a leap due to being applied to a larger model, but of all the labs, they are the ones that never underperform my evals vs public ones and they appear to strictly keep their training data sanitised.
Happy to discuss public vs private evals and the merit of each if you'd like, I do appreciate your reporting in general but just think the SVG benches have become evidently polluted, which is also why even simple queries in my benchmarks are private. Just saw Thinking Machines Inkling model succeed in certain queries that neither Fable 5, nor GPT-5.6 Sol on any reasoning level managed, which I feel is valuable to truly gauge where we are at. Informs my work with models, my views of the industry and my assessment of the future these tools have, along with how to best implement them to enable better UX.
Respectfully, did you? The comment was specific to doubting the believe simonw has that labs are not training [0] specifically for this task, which is exactly what simonw wrote in the post [1], that it is a believe of his that they don't. He did not mention any kind of evidence or any piece of information that would indicate that the commenter didn't read the blog post.
Did you read either the post or the comment it was referencing?
On the note of training on SVGs, I have seen some labs models outperform when prompted for SVGs of certain animal and action combinations (pelican on bike, panda eating burger, etc.) compared to other similarly outlandish prompts for SVG output that are not part of widely reported benchmarks, even shared evidence one of the last times this came up on here.
I'll note there's a difference between "pelicans on bikes aren't part of the training set" and "I’m still not convinced that labs are training for the benchmark".
I'm sure all sorts of crap pelican riding bicycle SVGs have ended up in the huge crawls of data that the labs feed into their pre-training steps.
What I'm questioning here is that there are labs who have sat down and deliberately tested and tweaked the performance for this particular task, independent of general model improvements.
The one exception here is Gemini, who have clearly invested a lot of effort in SVG tasks. I have no idea if my stupid benchmark influenced that decision!
Gemini have boasted about how good they are at pelicans riding bicycles, frogs on penny-farthings, giraffes driving a tiny car, ostriches on roller skates, turtles kickflipping skateboards, and dachshunds driving a stretch limousine. So if they trained for the test they did at least expand it a whole bunch! https://twitter.com/JeffDean/status/2024525132266688757
This is a sight-reading test. If a musician practices a piece for thousands of hours, it would no longer be an effective sight reading / creativity test. The purpose of the test was to see how models would compose something novel requiring the ability to compose orthogonal, normally unrelated, components into a coherent image.
We do. People who, for example, memorize question banks to pass certification tests without knowing the underlying material are equally frowned upon for not having the problem solving skills that they purport to.
I'll leave the contrasts between LLMs and people to the well-written sibling comments.
They can be in the training set but not deliberately trained for. There may be a lot of people posting pelican svgs, but not typically because they're high quality and worth replicating.
Like Simon concludes the article, the main use of this isn't to say which model is "better", but to try and poke at the model to sort out things like quality vs cost vs speed.
So I put together a quick comparison of the last couple iterations of Opus, Fable and now Kimi.
3T is impressive, but parameter count seems to be less important than I thought.
GLM is half the size of DeepSeek but costs four times as much, and beats it on every benchmark.
I'm not an expert on this stuff but it seems to be the attention mechanism. DeepSeek were bragging about how cheap they made it. But if you cut costs on attention you get worse results with way more parameters.
If I had to guess it seems to be the difference between memory (params) and intelligence (attention density). I think you need both.
One thing i keep thinking: you only run the pelican once per model. Run the same model a few times and you get some different pelicans, so some of "this one is better" might just be which run you picked for it. Would love to see 8 runs per model side by side. I bet for two close models, the gap between runs is about as big as the gap between the models.
I've done versions in the past where I ran 3 and picked the best one. At some point I'd like to automate that with an LLM-as-a-judge (from the same model family) picking the "best" one to move forth in the competition.
I wonder how the Chinese labs are training a 3 trillion parameter model on what has to be vastly smaller compute resources. If the U.S. compute advantage is persistent, it's hard to imagine that Chinese labs will be able to keep pace forever, as a matter of physics, but... so far they seem to be doing just fine.
Training and serving large models does require increasingly more compute, though. (The Chinese labs have clearly found some massive optimizations, but my point was that you'd think at some point even those optimizations wouldn't be enough to keep up with exponentially increasing model sizes.)
Do any of the vision models render the SVG and look at the result.
Perhaps more importantly can they do that during reinforcement training. Learning how to critically analyse the appearance of what it generates would be quite useful.
Manually feeding images back to models has been hilariously bad in the past which suggests that relating something it sees to something it wrote is not an ability it is very good at.
I've tried doing a loop of rending the SVG and then tweaking based on that, with local models (so, not nearly as strong). It wasn't very successful; it would mostly report that the image looked great and didn't need any tweaks. Maybe I should try it again, there have been some newer models since I first tried it. And yeah, maybe worth trying with bigger models. But I have found that models aren't necessarily the best at visual reasoning and review, even with a vision loop. Their lack of visual reasoning is part of why they still have trouble with things like ARC-AGI-3.
The idea is not to use pelicans on bikes but a similarly random non-sensical prompts: crows on scooters, squirrels in a moon rover etc. Then pick another one for another for next cross-llm evaluation.
It will be valuable to have two types of benchmarks: ones that evolve alongside the models and ones that never change. You probably can't get historical stability and resistance to flooding and training on at least some parts of it from the same test
My personal benchmark for new models has been to compare video making skills with something like remotion. Usually reveals if they have any "taste" or outside the box thinking.
I'm starting to not trust any "benchmarks" when it comes to frontier models at least. As an example Sol feels the most "gets stuff done" but has zero taste, or any capability to surprise.
And for frontier models I go one step ahead and try to recreate a complex animation video, with the ability for the model to review its own work. And at this Fable is still the top one.
Ex: https://www.youtube.com/watch?v=uDAeAuYyl0E (recreation of Claude announcement video) and https://www.youtube.com/watch?v=cSsVNtGPOIg (recreation of a fireship video). Sol did something similar but you can instantly tell its AI slop from very small things, and it just has no narrative or thought put into the writing.
And on creativity at least visually, Gemini 3.1 pro is somehow still up there. But its really hindered by its inability to use tool calls effectively or make a long term plan.
Another day, another model and another pelican :-)
I can't help but wonder where is the trend going? What will we have in five years? Maybe it will all have puttered out, and we will have moved to the next thing? Or maybe the prompt then will be "make a pelican ride a bicycle", and out will come the genetic code for a giant pelican with extremities suitable for a handle bar and pedals, and an inborn affinity to ride bicycles?
You say it's performative joke, but it all depends what you're using model for. So far the rule has been quite straightforward, better models consistently renders pelican in higher quality, I've yet to see an exception. It is also a good enough (for me at least) test for "taste" the model has.
> better models consistently renders pelican in higher quality
The article literally avoid making this argument and gives counterexamples to this statement.
Imagine what amazing SVG generators we could have if Simon had randomized the target image from the start (and companies wouldn't just overfit on pelicans).
You’re reading a personal blog and complaining about an open source personal project he runs and distributes for free. He’s allowed to talk about his personal work on his personal blog. Especially considering the cli utility he talks about is directly related to the post.
Imagine complaining about someone generating valuable content for free and not packaging it to your personal tastes.
K3 is as expensive as Sonnet, not great at writing English, is handing IP back to the Chinese, and once open source will be difficult to run at scale without the compute that OpenAI and Anthropic have largely grabbed.
Sorry, how again is this the end of the frontier labs?
According to some benchmarks has the coding capability of Opus at the price of Sonnet, supposedly will be open weights and is not subject to random trade wars with allied states.
Engineers get unbelievably silly about evaluating costs of things.
"The tokens are so expensive!" Oh my sweet child, how much would even the least capable human effort cost? This is what the executives properly understand that the programmers don't.
they're comparing to similar capability llm models, not humans. If one dishwasher does job at similar quality as another dishwasher, but using 30% more water and energy, you wouldn't compare to how much it costs human to do the same work, it would make no sense.
> they're comparing to similar capability llm models, not humans
25 cents is 10x the cost of 2.5 cents, but it's still extremely cheap for the product. It's very much the wrong comparison for a world where the primary competition is still humans who need to eat, and it treats percentage differences as more important than absolute differences when the opposite is true.
Well first of all, any non-trivial use of LLMs is going to be orders of magnitude more tokens than this, usually multiple millions at minimum. Benchmarks are just benchmarks after all.
Secondly, humans vs LLMs are apples vs oranges. It makes no more sense to compare human costs vs LLM costs as it would have to compare human costs vs calculator costs. LLMs are faster and cheaper but extremely different beasts with different limitations. Humans do not one-shot SVGs of pelicans riding bicycles, and they do not charge in tokens.
Comparing LLM cost efficiency is not something that should need to be defended. It's quite straightforward and reasonable...
> How does the prompt “Generate an SVG of a pelican riding a bicycle” add up to 95 input tokens? OpenAI’s tokenizer counts 10, Anthropic’s counts 10 for Opus 4.6, 30 for Opus 4.7 and 25 for Sonnet 5/Fable 5. Prompting “hi” to Kimi K3 counted 86 tokens, suggesting there may be an 85 token hidden system prompt. It refused to leak it though.
This is quite possibly reasoning-effort prompt which is injected before the opening <think> token whenever you set a custom reasoning effort, see e.g. DeepSeek-V4 max mode prompt: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main...
It's incredible Simon still believes pelicans on bikes aren't part of the training set, despite hundreds of them on blogs, forums, and Github. Stuff we put in our company blog shows up known by LLMs 6 months later, and we have 1000x less traffic than Simon's own website
The pelicans are still all rubbish. If they make it into the training set it doesn't help the models produce better pelicans, if anything it will make them perform worse!
Respectfully, the pelicans used to be an unrecognisable mess and now they’re unquestionably pelicans on bicycles, rendered poorly, from every model.
In the same timescale, model capabilities across the board have only meaningfully improved in places where the labs are focusing their training efforts.
Moreover, they have a uniform style, even though your prompt doesn’t ask for one. There's no model going rogue and producing a watercolour of a pelican. They’re all rendered in an approximately uniform style, even though the svg format has a basically unlimited possibility space.
You know what, that's actually something I hadn't considered before. There's definitely a bias towards a pelican cycling from left to right on a red bicycle against a blue sky and green grass.
Blue sky and green grass aren't that surprising, but the color and direction are interesting.
When I finally build the proper gallery I'll throw in a few other creature-vehicle combinations, and track some characteristics like which direction, color of bicycle, general pelican geometry etc. It will be interesting to see if other creatures end up with coincidentally similar design choices or if that's unique to the pelican-bicycle combination.
In photography (and probably art in general), there's a composition "rule" to frame moving subjects from left to right.
So the direction may not be that interesting!
I wonder if that changes in countries where the main language is written right to left?
Is it culture dependent? Is it because in English we read left to right?
There was a glorious moment when I thought that the Chinese models were more likely to produce right-to-left cycling pelicans, but sadly that trend didn't seem to hold up.
The art styling is more or less uniform too.
I haven't seen many AI works that produces a pelican on a bicycle done in a "Ligne Claire" style, for example.
I guess AI's narrows down the output probability space drastically and converge on some agreed upon aesthetics. Works great for computer programs but bad for art.
I thought my joke post was silly and then I read new comments and I'm like, "I didn't try hard enough" lol
> the pelicans used to be an unrecognisable mess and now they’re unquestionably pelicans on bicycles, rendered poorly, from every model
You would not expect that to happen if the models trained on the unrecognizable mess, right?
> model capabilities across the board have only meaningfully improved in places where the labs are focusing their training efforts
And the labs clearly did focus on improving image rendering.
> they have a uniform style
SVG output from LLMs always looks like that. It looked that way from the beginning; no LLM ever produced a watercolor when asked for SVG output. They all render the prompted element centered in the picture. They all tend to draw things going from left to right, and so on.
I’m not suggesting Simon’s pelicans in the dataset are having a meaningful impact. I’m expecting that a company like ScaleAI has a product along the lines of “benchmax dataset: SimonW’s Pelican on Bikes test” which is a private curated series of well-drawn SVGs of animals riding vehicles for training and RL.
If such a product existed I'm reasonably confident someone would have tipped me off by now, NDAs be damned.
Watercolors in SVG?
Simon - has no one told you about the Willison-Pelican Scaling Law?
```
if is_willison_pelican_blog_post:
[redacted]
```
You haven't seen their final form [1]
[1] final form is a frontend/react/let's not talk about it, library - it caused a great deal of PTSD to me and my previous company's team due to its dogmatic preference for "we use these axioms, end of story", over practical utility - so it was quite challenging to do state of the art tasks such as nested form fields (e.g. 'user.address.personal.line-1'). The PTSD it caused made us all block out the memories, I suppose. But - it had zero dependencies. That is what mattered. It kept us going. We weren't reaching for more. We had plenty of time.
And thank god for that. Because I'd forgotten my watch in California - and this was in Tokyo [2]
[2] a joke within a joke about Jensen's Kyoto gardener story. Beautiful story, drowned out by WatchGate memes. Why can't jokes have layers? Models have trillions. If you miss 100% of the jokes you don't make, make all the jokes. Someone will laugh (eventually, maybe?) Even if it's: "this person + comedy club = full secret service detail". If someone laughs at that - at my own expense? I don't mind. They laughed. I know this is a gibberish, off-topic message - it's also a human message. I just felt we need more such things in our lives these days.
PS: have you physically seen a pelican in real life? (not a joke)
> PS: have you physically seen a pelican in real life? (not a joke)
We have several thousand living 15 minutes walk from our house. I recently started adding my wildlife photography (from iNaturalist) to my blog, so I'm posting several new pelican photos a week at the moment: https://simonwillison.net/search/?q=pelican&type=beat%3Asigh...
Yes, I see your point.
Your pelican output is thus both in the training set and yet still outside the capability of the model architecture.
And so you are tracking both the capability of the training and also the capability of the querying!
When you receive your first outstanding pelican it will track a gain of capability.
(btw I first mentioned simonw-pelican-into-training-set in May 2025 on twitter.)
My 3D-egyptology-explainer showed a massive uplift for Kimi K3 and this tracks a much improved 3D capability.
I agree with that. I think, in particular, all the broken bike frames associated with "pelican on a bike" probably make it harder for LLMs to render correct bike frames.
At this point I am simply interested how much longer you're gonna ride this schtick
I'm a deep believer in commitment to the bit. https://simonwillison.net/tags/pelican-riding-a-bicycle/
Simon has stated a few times that he knows it’s possible that pelicans could be in the training sets. He also has other tests he doesn’t share publicly. He’s just a fan of pelicans.
From the article it doesn't even sound like he cares about pelicans at all, and doesn't think they are a good way to compare models anymore ... but people are used to seeing the test now, and it does serve as a common "hello world" unit of work.
Pelicans and bikes can be in the training set without them training for this specific benchmark.
Yes and that would improve its ability to draw SVGs of pelicans on bikes, no?
Would it? Tongue in cheek.
and that is bad because ?
the nature of the test was to see if the models can effectively compose an image of a novel concept outside the training set. If they are trained on it, it ceases to be an interesting test to some extent.
I would urge you to re-read the blog post you are commenting on. It pretty clearly explains how it is an interesting test independently of "see[ing] if the models can effectively compose an image of a novel concept outside the training set".
it's still interesting because there's no pelican-on-bike model, and if you're training a model well enough, then it should be obvious when a model has reached "AGI" or whatever.
More to it, the actual bloody companies are using them as a reference. Maybe it’s a 3d version, not an svg - but it clearly shows they’re on the radar of these companies.
Yeah I asked Nano Banana to make a render of our company office and was scarily accurate
A person from Google famously put on her linkedin that her job was to optimize SVG for Gemini 3.0.
SVG output is useful, though. I often ask whatever LLM I have open to generate placeholder icons whenever I need them.
Did you read the post? It's not even that long. He explicitly mentions this...
Are they responding to: “I’m still not convinced that labs are training for the benchmark—if they were, I’d expect much better results.”
Clearly not. There's a subset of HN users who rush to post this same thing every single time.
Maybe it gets posted every time because besides a personal believe by the person popularising this "benchmark", there is no reason to assume that certain labs aren't intentionally training to game this and every other lab at least unintentionally gets improvements for this specific combination of animal and action because the internet is full of both good and bad examples, often ranked, which does inevitably become training data.
I have shared examples of certain models by certain labs doing far better on the pelican cycling vs other, similar prompts. Just operating on a feeling that labs don't optimise for this (as mentioned, even if they don't training data is filled with these) is not solid enough that criticism shouldn't be leveraged when it comes up.
> I have shared examples of certain models by certain labs doing far better on the pelican cycling vs other, similar prompts
Please share those again!
One of the things I'm most looking forward to is a lab producing a model that creates a really great pelican riding a bicycle and then a terrible sloth riding a skateboard (or whatever).
I've not seen that myself yet.
Evidence in the other direction (that they're able to generalize) is that I can't think of any LLM currently that can't create usable (placeholder) SVG icons, I tried a bit before the pelican became popular and it was abysmal.
Happy to, here one example where Grok 4 Fast, despite producing a fairly consistent pelican [0], did severely worse in a similarly outlandish scenario along with Haiku 4.5 and GPT-5 for context: https://news.ycombinator.com/item?id=45599403
> [...] a really great pelican riding a bicycle and then a terrible sloth riding a skateboard [...]
Happy to play ball. You made a blog post a few weeks back on one of the Qwen models with the eye-catching title "Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7" [1].
Here is what Qwen3.6-35B-A3B via Openrouter provided for a sloth riding a skateboard: https://imgur.com/a/Dy8fvR5
Like Grok 4 Fasts attempt at a mushroom in a rowboat, it is barely recognisable as anything despite both Qwen3.6-35B-A3B and Grok 4 Fast having no issue with more popular (i.e. benchmarked) examples. Whether this is a case of training data being unsanitized or intentional benchmark targeted training, I cannot say, but it is the case.
And here is Opus 4.7, again via Openrouter: https://imgur.com/a/Qus1Enf
A massive delta in favour of Opus 4.7, despite the pelican Qwen3.6-35B-A3B produced being noticeably better as you rightly pointed out. What does that tell us?
I would not critique your use of this fun benchmark the way I tend to if I did not have evidence to back up my position, including private evals that I can reliably use to point out major deviations between what a models claimed performance is according to major benchmarks vs the actual performance outside these known test cases.
I will also say that while I have a lot to be critical of regarding Anthropics modus operandi, especially how they present interesting findings like their j-space work, which I found was irresponsibly anthropomorphic in their reporting, especially as this wasn't a first in model interpretability, but mainly a leap due to being applied to a larger model, but of all the labs, they are the ones that never underperform my evals vs public ones and they appear to strictly keep their training data sanitised.
Happy to discuss public vs private evals and the merit of each if you'd like, I do appreciate your reporting in general but just think the SVG benches have become evidently polluted, which is also why even simple queries in my benchmarks are private. Just saw Thinking Machines Inkling model succeed in certain queries that neither Fable 5, nor GPT-5.6 Sol on any reasoning level managed, which I feel is valuable to truly gauge where we are at. Informs my work with models, my views of the industry and my assessment of the future these tools have, along with how to best implement them to enable better UX.
[0] https://simonwillison.net/2025/Sep/20/grok-4-fast/
[1] https://simonwillison.net/2026/Apr/16/qwen-beats-opus/
Respectfully, did you? The comment was specific to doubting the believe simonw has that labs are not training [0] specifically for this task, which is exactly what simonw wrote in the post [1], that it is a believe of his that they don't. He did not mention any kind of evidence or any piece of information that would indicate that the commenter didn't read the blog post.
Did you read either the post or the comment it was referencing?
On the note of training on SVGs, I have seen some labs models outperform when prompted for SVGs of certain animal and action combinations (pelican on bike, panda eating burger, etc.) compared to other similarly outlandish prompts for SVG output that are not part of widely reported benchmarks, even shared evidence one of the last times this came up on here.
[0] ... incredible Simon still believes ...
[1] I’m still not convinced that labs ....
I'll note there's a difference between "pelicans on bikes aren't part of the training set" and "I’m still not convinced that labs are training for the benchmark".
I'm sure all sorts of crap pelican riding bicycle SVGs have ended up in the huge crawls of data that the labs feed into their pre-training steps.
What I'm questioning here is that there are labs who have sat down and deliberately tested and tweaked the performance for this particular task, independent of general model improvements.
The one exception here is Gemini, who have clearly invested a lot of effort in SVG tasks. I have no idea if my stupid benchmark influenced that decision!
Gemini have boasted about how good they are at pelicans riding bicycles, frogs on penny-farthings, giraffes driving a tiny car, ostriches on roller skates, turtles kickflipping skateboards, and dachshunds driving a stretch limousine. So if they trained for the test they did at least expand it a whole bunch! https://twitter.com/JeffDean/status/2024525132266688757
Imagine if we applied this train of logic to humans.
"That artist saw a pelican at the beach once!" [cue the outrage] "He's not a real artist, he's a cheater and produces nothing original!"
This is a sight-reading test. If a musician practices a piece for thousands of hours, it would no longer be an effective sight reading / creativity test. The purpose of the test was to see how models would compose something novel requiring the ability to compose orthogonal, normally unrelated, components into a coherent image.
We do. People who, for example, memorize question banks to pass certification tests without knowing the underlying material are equally frowned upon for not having the problem solving skills that they purport to.
I'll leave the contrasts between LLMs and people to the well-written sibling comments.
Except, of course, LLMs are not humans, and they do not learn or "reason" in a way which even remotely resembles humans.
Plus obviously humans can still overfit to a specific style of test.
They can be in the training set but not deliberately trained for. There may be a lot of people posting pelican svgs, but not typically because they're high quality and worth replicating.
Like Simon concludes the article, the main use of this isn't to say which model is "better", but to try and poke at the model to sort out things like quality vs cost vs speed.
So I put together a quick comparison of the last couple iterations of Opus, Fable and now Kimi.
Kimi is cheapest by 5x but also slowest by 2x
https://9gpyw4uxr2.evvl.io/
3T is impressive, but parameter count seems to be less important than I thought.
GLM is half the size of DeepSeek but costs four times as much, and beats it on every benchmark.
I'm not an expert on this stuff but it seems to be the attention mechanism. DeepSeek were bragging about how cheap they made it. But if you cut costs on attention you get worse results with way more parameters.
If I had to guess it seems to be the difference between memory (params) and intelligence (attention density). I think you need both.
Or, GLM 5.2 simply had more time in the RL oven.
Deepseek V4 Flash, the 284B model, is roughly equivalent to launch GLM 5, the 744B [sic] model.
After MoE entered the mix, raw parameter count is less useful a measure.
The gap is closing . I think Kimi 3 is only 3 months behind the US model. It’s gpt 5.5 class model , which was released in the end of April.
One thing i keep thinking: you only run the pelican once per model. Run the same model a few times and you get some different pelicans, so some of "this one is better" might just be which run you picked for it. Would love to see 8 runs per model side by side. I bet for two close models, the gap between runs is about as big as the gap between the models.
I've done versions in the past where I ran 3 and picked the best one. At some point I'd like to automate that with an LLM-as-a-judge (from the same model family) picking the "best" one to move forth in the competition.
I built a whole ELO scoring mechanism a while back, described here: https://simonwillison.net/2025/Jun/6/six-months-in-llms/#ai-...
I probably should spend some time on this now, even though the benchmark itself is feeling a bit stale. There's still a lot of demand for a gallery!
I wonder how the Chinese labs are training a 3 trillion parameter model on what has to be vastly smaller compute resources. If the U.S. compute advantage is persistent, it's hard to imagine that Chinese labs will be able to keep pace forever, as a matter of physics, but... so far they seem to be doing just fine.
It's not like same parameter count models are identical, so that doesn't appear to be an indicator for quality, or even compute requirements?
There seems to be more to producing a better model than brute forcing parameter count after all.
Training and serving large models does require increasingly more compute, though. (The Chinese labs have clearly found some massive optimizations, but my point was that you'd think at some point even those optimizations wouldn't be enough to keep up with exponentially increasing model sizes.)
Do any of the vision models render the SVG and look at the result.
Perhaps more importantly can they do that during reinforcement training. Learning how to critically analyse the appearance of what it generates would be quite useful.
Manually feeding images back to models has been hilariously bad in the past which suggests that relating something it sees to something it wrote is not an ability it is very good at.
I've tried doing a loop of rending the SVG and then tweaking based on that, with local models (so, not nearly as strong). It wasn't very successful; it would mostly report that the image looked great and didn't need any tweaks. Maybe I should try it again, there have been some newer models since I first tried it. And yeah, maybe worth trying with bigger models. But I have found that models aren't necessarily the best at visual reasoning and review, even with a vision loop. Their lack of visual reasoning is part of why they still have trouble with things like ARC-AGI-3.
I imagine all vision models have to do this, this being html rendering, to be able to do well in web design.
The idea is not to use pelicans on bikes but a similarly random non-sensical prompts: crows on scooters, squirrels in a moon rover etc. Then pick another one for another for next cross-llm evaluation.
It will be valuable to have two types of benchmarks: ones that evolve alongside the models and ones that never change. You probably can't get historical stability and resistance to flooding and training on at least some parts of it from the same test
My personal benchmark for new models has been to compare video making skills with something like remotion. Usually reveals if they have any "taste" or outside the box thinking.
I'm starting to not trust any "benchmarks" when it comes to frontier models at least. As an example Sol feels the most "gets stuff done" but has zero taste, or any capability to surprise.
And for frontier models I go one step ahead and try to recreate a complex animation video, with the ability for the model to review its own work. And at this Fable is still the top one. Ex: https://www.youtube.com/watch?v=uDAeAuYyl0E (recreation of Claude announcement video) and https://www.youtube.com/watch?v=cSsVNtGPOIg (recreation of a fireship video). Sol did something similar but you can instantly tell its AI slop from very small things, and it just has no narrative or thought put into the writing.
https://mesmer.tools/benchmarks/ai-video-generation , I usually put basic ones here.
And on creativity at least visually, Gemini 3.1 pro is somehow still up there. But its really hindered by its inability to use tool calls effectively or make a long term plan.
Old and busted: benchmaxxing
New hotness: pelicanmaxxing
It’s not bad kind of expensive for 25c but if the prompt is rendered cost is much better.
I wonder what the non-subsidized cost is. Add in the electricity and water too.
We may be boiling the oceans but at least we are finally getting some good SVGs of pelicans on bicycles.
Is there a gallery of all pelicans generated by simon over time?
https://simonwillison.net/tags/pelican-riding-a-bicycle/ isn’t quite a gallery, but pretty close.
If Simon reads this debate, I would gladly vote for such a gallery. It would belong to "digital heritage of mankind".
I would be surprised if pelican svgs are not part of the training corpus rn
If that were the case then it'd do a way better job. Think experienced artist level.
how would great pelicans make their way into the training set?
what they do have are many different pelicans and people helpfully rating them in the comments.
That’s covered in the article
Don't see why we have to have this spammed every model release when Fable class models perform the same as Opus on basic tasks like these.
What spam? It’s one article. You can skip it
Another day, another model and another pelican :-)
I can't help but wonder where is the trend going? What will we have in five years? Maybe it will all have puttered out, and we will have moved to the next thing? Or maybe the prompt then will be "make a pelican ride a bicycle", and out will come the genetic code for a giant pelican with extremities suitable for a handle bar and pedals, and an inborn affinity to ride bicycles?
I’m excited for this specific brand of survival horror.
You are thinking too hard on this. This entire "benchmark" is a performative joke for attention that only works on HN.
> What will we have in five years? Maybe it will all have puttered out, and we will have moved to the next thing?
We will just have more of the same.
> This entire "benchmark" is a performative joke for attention that only works on HN.
I take exception to that! It's a performative joke for attention that works far more widely than just Hacker News.
You say it's performative joke, but it all depends what you're using model for. So far the rule has been quite straightforward, better models consistently renders pelican in higher quality, I've yet to see an exception. It is also a good enough (for me at least) test for "taste" the model has.
> better models consistently renders pelican in higher quality The article literally avoid making this argument and gives counterexamples to this statement.
Imagine what amazing SVG generators we could have if Simon had randomized the target image from the start (and companies wouldn't just overfit on pelicans).
If anyone wants to try SVG generation from different models, I made this: https://codeinput.com/svg (here is an older generation: https://codeinput.com/s/5KEGl1e3rB3)
You still need an OpenRouter API Key and be careful this can burn quite a bit of money.
Imagine shilling some CLI tools no one uses in this post.
Lighten up.
You’re reading a personal blog and complaining about an open source personal project he runs and distributes for free. He’s allowed to talk about his personal work on his personal blog. Especially considering the cli utility he talks about is directly related to the post.
Imagine complaining about someone generating valuable content for free and not packaging it to your personal tastes.
Kimi is right out since they use classical music branding to sell their slop. At least McDonalds does not sell Verdi or Allegro burgers.
Why does Kimi not use a "Double Cheese Whammy" branding for "their" butchered and stolen IP?
K3 is as expensive as Sonnet, not great at writing English, is handing IP back to the Chinese, and once open source will be difficult to run at scale without the compute that OpenAI and Anthropic have largely grabbed.
Sorry, how again is this the end of the frontier labs?
According to some benchmarks has the coding capability of Opus at the price of Sonnet, supposedly will be open weights and is not subject to random trade wars with allied states.
Competition is always good.
You mean the scale that AWS provides with Bedrock?
Bedrock needs to actually update their chinese models to the newest versions for this to matter.
> This is expensive—the pelican cost 25 cents!
Engineers get unbelievably silly about evaluating costs of things.
"The tokens are so expensive!" Oh my sweet child, how much would even the least capable human effort cost? This is what the executives properly understand that the programmers don't.
they're comparing to similar capability llm models, not humans. If one dishwasher does job at similar quality as another dishwasher, but using 30% more water and energy, you wouldn't compare to how much it costs human to do the same work, it would make no sense.
> they're comparing to similar capability llm models, not humans
25 cents is 10x the cost of 2.5 cents, but it's still extremely cheap for the product. It's very much the wrong comparison for a world where the primary competition is still humans who need to eat, and it treats percentage differences as more important than absolute differences when the opposite is true.
Well first of all, any non-trivial use of LLMs is going to be orders of magnitude more tokens than this, usually multiple millions at minimum. Benchmarks are just benchmarks after all.
Secondly, humans vs LLMs are apples vs oranges. It makes no more sense to compare human costs vs LLM costs as it would have to compare human costs vs calculator costs. LLMs are faster and cheaper but extremely different beasts with different limitations. Humans do not one-shot SVGs of pelicans riding bicycles, and they do not charge in tokens.
Comparing LLM cost efficiency is not something that should need to be defended. It's quite straightforward and reasonable...
Would anyone pay a human to create an SVG of a pelican riding a bike?
In fact humans get paid to create SVGs of all kinds of things.
Well, not anymore
Well, no, not now they won’t.