Could one reverse the test? Feed ai generated svgs into the model and ask it to decide which of 10 animals (dog, cat, pelican, crocodile, etc) doing which of 10 actions (moonwalking, paddleboarding, sitting, riding a bicycle, etc.).
It's a meme. Simon Willison started doing it because, I'm paraphrasing, it's something that (at the time) was just sort of at the edge of what you could expect an LLM to be able to do, and also it's somewhat of an impossible task since, of course, Pellicans can't ride bicycles. From there, he started writing about it along with every major LLM release he covered. His coverage is generally well regarded (hi, Simon!) and so it sticks around. It's never been intended as a serious benchmark.
already , i self hosted Deekseek 4.1 Flash , Qwen 3.8 Flash Next , Glm 5.3 flash , tested it , all benchmaxxed to the point that all of them can draw that well , much better than other logos .
I recall that when this benchmark came out there were concerns that AI companies were gaming benchmarks to fake their progress. The purpose was to find something silly and outlandish that AI companies would definitely not have trained their models on, and thus have a way to check a model's capabilities.
A pelican riding a bicycle is pretty random. That was the whole point.
I haven't evaluated this one, but for what I have, it seem the pplx-decider-v1.2 (and its -flash variant) are pretty good. Anyone evaluated all these open-source decision models on their own tasks? (I evaluated on MonitoringBench).
I didn't look specifically at open-source but I evaluated as many decision models as I could on openrouter against how good are they at detecting phishing, which isn't the best use of these models but I was interested to see how they compared. https://emdashcyber.com/phishnchips-decision-models/
Have you checked mammoth-1/elephant-1 models? They are benchmarking quite well compared and are currently free on openrouter. I currently use mammoth-1 for a small internal app and it works flawlessly for what I am doing.
Reads like an ad for Microsoft and its model because it's an article about a decision model that doesn't mention Jev or other decision models at all.
No pelicans riding bicycles on this one, eh?
Could one reverse the test? Feed ai generated svgs into the model and ask it to decide which of 10 animals (dog, cat, pelican, crocodile, etc) doing which of 10 actions (moonwalking, paddleboarding, sitting, riding a bicycle, etc.).
AlexNet (2012) can do the first part :)
Instead the right app is hotdog / not hotdog
Pelicans are deciding.
Pelicans deciding if they are going to ride a bicycle.
Bikes with integrated decision models deciding which pelican needs a ride the most.
This made me realize it's been a long time since I did some brainstorming and how I missed it.
Yeah the most usesless and stupidest bench .. I don't know why LLM Community thinks its good idea to use that.. It can't work on decision systems.
It's a meme. Simon Willison started doing it because, I'm paraphrasing, it's something that (at the time) was just sort of at the edge of what you could expect an LLM to be able to do, and also it's somewhat of an impossible task since, of course, Pellicans can't ride bicycles. From there, he started writing about it along with every major LLM release he covered. His coverage is generally well regarded (hi, Simon!) and so it sticks around. It's never been intended as a serious benchmark.
The 3D visualizations have always made it starkly apparent the difference in quality between quantizations and thinking.
The pelican meme is so much better than benchmaxxed stuff. Unfortunately we are likely now in pelicanmaxxing territory.
already , i self hosted Deekseek 4.1 Flash , Qwen 3.8 Flash Next , Glm 5.3 flash , tested it , all benchmaxxed to the point that all of them can draw that well , much better than other logos .
Sorry I didn't knew it was started here and , lol I didn't know it was viralized by Simon of django! Now these days , many LLMs are benchmaxxed to it.
> Yeah the most usesless and stupidest bench ..
I recall that when this benchmark came out there were concerns that AI companies were gaming benchmarks to fake their progress. The purpose was to find something silly and outlandish that AI companies would definitely not have trained their models on, and thus have a way to check a model's capabilities.
A pelican riding a bicycle is pretty random. That was the whole point.
I haven't evaluated this one, but for what I have, it seem the pplx-decider-v1.2 (and its -flash variant) are pretty good. Anyone evaluated all these open-source decision models on their own tasks? (I evaluated on MonitoringBench).
I didn't look specifically at open-source but I evaluated as many decision models as I could on openrouter against how good are they at detecting phishing, which isn't the best use of these models but I was interested to see how they compared. https://emdashcyber.com/phishnchips-decision-models/
I had my first use case recently, and did some very non-systematic benchmarking of models on OpenRouter.
The only one that came close to Jev itself was the one from respan (respan/span-02), which is also open source and cheaper too.
Have you checked mammoth-1/elephant-1 models? They are benchmarking quite well compared and are currently free on openrouter. I currently use mammoth-1 for a small internal app and it works flawlessly for what I am doing.
https://huggingface.co/baselabs/mammoth-1