Are AI Labs Pelicanmaxxing?(dylancastillo.co) |
Are AI Labs Pelicanmaxxing?(dylancastillo.co) |
"Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes."
Or the more pop layman version
"When a measure becomes a metric/KPI, it ceases to be a good measure."
Story time, I live in Argentina, and we don't have Big Macs, the main Mc Donald's brand, here, because during the CFK presidency, one of her tactics was to Goodhart economic metrics. Even the informal obscure ones like the [Big Mac Index](https://en.wikipedia.org/wiki/Big_Mac_Index), I don't know the precise details, but the Big Mac ended up being a very cheap item, like 2 or 3 times cheaper than actual menu items, but it was never on the advertised menu, and it also ended up being very small compared to the other burgers, so it wasn't even like a hack, a shrinkflation type of deal.
But hey, anyone who read the Big Mac Index table would never find Argentina at the bottom of that list along with a couple of other countries with bad brands, so the ploy worked. And now we live with the aftershock, the brand never really turned around, other brands with ridiculous names took over it like the McTasty, which makes me sound like that skit from Tarantino's Pulp Fiction.
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.
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.
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.
Snakes on a plane, weasels on a diesel, spiders on a glider, baboons on a balloon, goats on a boat.
> 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.
1. https://simonwillison.net/2025/Nov/13/training-for-pelicans-...
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.
If you train towards the test, you aren't necessarily improving overall fitness, but you are destroying the value of that test over time because you're decreasing its correlation with overall fitness.
I don't think I'd go that far!
When someone says a model has been benchmaxxed, what they really mean is that it performs better in benchmarks compared to their real world experience. That's a real thing, I've certainly experienced it with some models.
...my take is that some things in life just resist quantitative measurements. Who is the best job candidate? What is the best programming language? Add AI models to the pile.
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
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.
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.
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.
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?
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.
aren't some LLM going to digest that thread at some point and indirectly learn from it?
basically my point is originally this was a good benchmark because it was an absurd never-seen-before thing, but now that it is in content, some models are going to get a benefit in education?
you'd need the "AI" equivalent of an old-school "google whack", something with no previous results
It is the reinforcement learning that produces more tangible results with less data, but it is something that the AI labs specifically selects and it is not picked up unknowingly
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.
I agree, it is ridiculous to ask an LLM to replace an artist.
What sufficiently hard, but useful, problem would you ask the model for?
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.
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.
That simonw is causing labs to do extra fine-tuning runs for this seems highly probable :)
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.
...which.. hmm I dunno if they are same or not