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.
> All 21 pelican-bicycle images, across all seven labs, face right. No other animal/vehicle combination does that.
> However, facing right is common: 60% of all 1,008 images do it. How common depends on the animal and the vehicle, and bicycles are one of the two vehicles where it’s strongest
Of course the pelican on the bicycle is facing right. The drivetrain on a bicycle is on the right side. If you want any representation of a bicycle that shows the drivetrain you're going to show the right side of it if you want to do so without the frame occluding it. It's an excellent bet that their training data reflects this.
As near as I can tell, all of the bicycles are shown facing right, regardless of the direction the animal is facing (GPT 5.6-Terra, Sample 1/3). Also, in every case where the rider has legs (i.e. not the whale) both of the rider's legs are on the right side of the bicycle. This suggests a pretty serious lack of actual understanding of how a bicycle works.
I've had the feeling that labs aren't pelicanmaxxing specifically but that they do have some sort of RL environment for SVGs that they are letting the AIs overcook in. Specifically I'm thinking of the gemini 3.1 pro annoucnement that seemed to have a huge leap in animated SVG performance but not much else impressive about it.
So they aren't pelicanmaxxing but they are benchmaxxing in a way. The benefit of the pelican was originally that uplift on the pelican signaled an overall uplift on intelligence. I don't believe that is the case anymore and it is just another jagged edge of model intelligence.
This is funny, I actually did a similar experiment just yesterday.
Looking for evidence of the same, but with another twist: checking if the models would choose to create a pelican on a bicycle, if no specific bird or method of transportation was specified.
I'm glad someone ran the numbers on this. Every single Simon Willison post of an SVG is followed with someone dismissing it saying "I'm sure they train on it by now." This is despite a good blog post with sound logic on how easy that is to catch. [1] Glad to see someone took the time for a quantitative analysis of dumb little animals riding dumb little bikes.
I feel like getting LLMs to spit out an SVG is akin to getting a human artist to draw something by just reciting a list of coordinates. It's insanely hard and unnatural.
Image generation models nowadays can easily generate a photorealistic pelican riding a bicycle, where the bicycle has perfect structure. But it is, of course, only a raster image.
It seems that we're missing a kind of step to decompose an image into a list of instructions (say, SVG paths, or even brush strokes with a real brush) to reproduce it properly. Doing so would probably need a true understanding of the structure of the scene, which is something that AI still struggles with to this day.
I know. I'm just thinking about how to make AI create SVGs better... in theory, a sufficiently smart AI could "generate an image in its head", think about it, and then output the SVG paths to produce said image. Intuitively that would be somewhat closer to how human artists convert artistic visions into a sequence of arm movements while holding a brush (obviously, humans don't hold a fully formed, photorealistic image in the head while drawing, but rather vague concepts, but still).
Image models that support text output like Image2, or general text models that can read images like Claude can vectorize raster images. But they aren't very good at it, doing it manually in Inkscape still produces better quality even when done by non-artists.
Pelicans aside, we need to remember that benchmarks are the only good quantitative way we have of comparing models. If someone has complaints about “benchmaxxing”, please ask them to contribute a better benchmark! It is valuable work and very appreciated.
He didn't. That's why the article exists. You have to do the science to see if it does.
He was asking the question - do we see gains across other tasks? The underlying question was: Is the additional attention given to this specific task creating a false impression of progress?
So this is not my experience at all for asking about simple SVG icons for web-pages. Here is one of the examples I have tried for in the past, make a simple cartoon SVG knife for a map icon for a crime map.
Can see the images for ChatGPT/Claude (Sonnet 5), and Gemini are all quite bad.
Jagged edge of LLMs. How do you explain being able to generate very complicated shapes in the Pelican example but cannot make a much simpler icon without just alluding to it is in the training data?
I find it humorous that the animal + plane combo appears to be such an outlier. I assume this is due to the models assuming the user mean plain and misspelled it in the prompt.
OK, so we've done animal_vehicle, how about new SVG ideas each time? I just tried "make an SVG of a man sitting in a chair at a computer behind a desk" which gives more interesting results than the animalVehicle test.
Asking questions that are absurd is like interviewing developers and asking absurd questions on the grounds that it tests creative and critical thinking.
Remember these Microsoft interview questions designed to identify the best developers?
"If you could eliminate one U.S. state, which one would it be?"
"How would you move Mount Fuji?"
Absurd interview questions have an air of legitimacy due to the quasi sophisticated justifications put forward for why they are good tests.
Absurd interview questions are not good tests of people or LLMs.
It was never intended as a meaningful benchmark. The surprising thing was that for the first ~12 months performance on the stupid pelican benchmark did seem to correspond to the performance of the models on other tasks.
That pattern no longer holds - Fable 5 and GPT-5.6 have both been out-pelicaned by lesser models now.
That’s an impressive image, but what a mistake it was to click the second link (on mobile without an ad blocker). I wouldn’t send it to anyone I respect ...
Just click through the models. At a glance (and highly subjective) I don't see anything jumping out as oom worse than anything else. I only noticed a model placing the animal inside a plane (with seat and small window) but other than that, they all seem similar inside each model to me.
If an AI researcher was going to pelicanmaxx, they would almost certainly apply the augmentations mentioned in the article during training, e.g. randomly selecting animals and conveyances. You’d want a model that generalizes well, just sfting in that specific prompt would be pretty bush league for a frontier lab.
I don’t have any reason to believe they are gaming the benchmark, just saying. I do find the idea of a data labeller having to generate thousands of svgs of different animals on different modes of transportation quite funny though.
At this point, I think there are so many pelican images in the pretraining data that drawing a pelican no longer makes sense as a model evaluation task.
The models definitely seem to pay attention to the tests.
Since the tests can be generally gamed with directing descriptions at it non-deterministically, there's a greater chance the questions solution can be found.
Of course, hopefully the models are instead adding patterns and types of questions as well and it makes the models more capable, but it may be limited in how it transfers to other types of questions in breadth or depth.
> Every single one is a pelican, on a bicycle, first try. When every student gets an A, the exam has stopped grading.
Numerous pelicans and their bikes are clearly horribly malformed. In fact none of the bike frames are correct. Fable and Opus come close, but the top of the diamond is disconnected in Fable's case and the head tube is misaligned with the front fork in Opus's case.
And of course, as the parent post shows, labs don't actually seem to be training on the pelican bike case.
Agreed. And more; the Macbooks are pretty much the same - some are god approximations, some are terrible, all of them are recognisably a MacBook. And if you start using it they can train on it.
The problem isn't the test, its that is a public test.
Simon has previously said he has a list of secret prompts (at least one of which he "burned" as a demonstration a while ago). That's what makes it a good test - his commentary on the public test is something of a proxy for non-public tests. This makes it a good benchmark.
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.
> However, facing right is common: 60% of all 1,008 images do it. How common depends on the animal and the vehicle, and bicycles are one of the two vehicles where it’s strongest
Of course the pelican on the bicycle is facing right. The drivetrain on a bicycle is on the right side. If you want any representation of a bicycle that shows the drivetrain you're going to show the right side of it if you want to do so without the frame occluding it. It's an excellent bet that their training data reflects this.
Citation: https://www.rei.com/c/bikes
Edited to add:
As near as I can tell, all of the bicycles are shown facing right, regardless of the direction the animal is facing (GPT 5.6-Terra, Sample 1/3). Also, in every case where the rider has legs (i.e. not the whale) both of the rider's legs are on the right side of the bicycle. This suggests a pretty serious lack of actual understanding of how a bicycle works.
So they aren't pelicanmaxxing but they are benchmaxxing in a way. The benefit of the pelican was originally that uplift on the pelican signaled an overall uplift on intelligence. I don't believe that is the case anymore and it is just another jagged edge of model intelligence.
Looking for evidence of the same, but with another twist: checking if the models would choose to create a pelican on a bicycle, if no specific bird or method of transportation was specified.
My version of it: https://www.modelbias.ai/pelican-on-a-bicycle-test
1. https://simonwillison.net/2025/Nov/13/training-for-pelicans-...
Image generation models nowadays can easily generate a photorealistic pelican riding a bicycle, where the bicycle has perfect structure. But it is, of course, only a raster image.
It seems that we're missing a kind of step to decompose an image into a list of instructions (say, SVG paths, or even brush strokes with a real brush) to reproduce it properly. Doing so would probably need a true understanding of the structure of the scene, which is something that AI still struggles with to this day.
Exactly--and you have to ask yourself at this point what "maxxing" really means, since "get better at drawing SVGs" is a useful skill.
Pelicans aside, we need to remember that benchmarks are the only good quantitative way we have of comparing models. If someone has complaints about “benchmaxxing”, please ask them to contribute a better benchmark! It is valuable work and very appreciated.
He was asking the question - do we see gains across other tasks? The underlying question was: Is the additional attention given to this specific task creating a false impression of progress?
https://x.com/CrimeDecoder/status/2080008114615537766
Can see the images for ChatGPT/Claude (Sonnet 5), and Gemini are all quite bad.
Jagged edge of LLMs. How do you explain being able to generate very complicated shapes in the Pelican example but cannot make a much simpler icon without just alluding to it is in the training data?
Your next iteration will need different animals and different transportation options. You'll run out after a few iterations.
Test the LLLM against things you want it to do.
Asking questions that are absurd is like interviewing developers and asking absurd questions on the grounds that it tests creative and critical thinking.
Remember these Microsoft interview questions designed to identify the best developers?
"If you could eliminate one U.S. state, which one would it be?"
"How would you move Mount Fuji?"
Absurd interview questions have an air of legitimacy due to the quasi sophisticated justifications put forward for why they are good tests.
Absurd interview questions are not good tests of people or LLMs.
Relevant questions are good tests.
Yes, deliberately so.
It was never intended as a meaningful benchmark. The surprising thing was that for the first ~12 months performance on the stupid pelican benchmark did seem to correspond to the performance of the models on other tasks.
That pattern no longer holds - Fable 5 and GPT-5.6 have both been out-pelicaned by lesser models now.
I don’t have any reason to believe they are gaming the benchmark, just saying. I do find the idea of a data labeller having to generate thousands of svgs of different animals on different modes of transportation quite funny though.
Since the tests can be generally gamed with directing descriptions at it non-deterministically, there's a greater chance the questions solution can be found.
Of course, hopefully the models are instead adding patterns and types of questions as well and it makes the models more capable, but it may be limited in how it transfers to other types of questions in breadth or depth.
I think we should stop using pelican benchmark.
> Every single one is a pelican, on a bicycle, first try. When every student gets an A, the exam has stopped grading.
Numerous pelicans and their bikes are clearly horribly malformed. In fact none of the bike frames are correct. Fable and Opus come close, but the top of the diamond is disconnected in Fable's case and the head tube is misaligned with the front fork in Opus's case.
And of course, as the parent post shows, labs don't actually seem to be training on the pelican bike case.
The problem isn't the test, its that is a public test.
Simon has previously said he has a list of secret prompts (at least one of which he "burned" as a demonstration a while ago). That's what makes it a good test - his commentary on the public test is something of a proxy for non-public tests. This makes it a good benchmark.