Unlike the Internet, Microsoft did recognise the pivotal moment. But they made the wrong play… that and they’ve annoyed pretty much everybody with their products, I don’t think I’ve ever seen such a successful company with such little good will from their community.
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.).
Put a decision model in my bike that decides when it is time to automatically shift up/down based on signals (FTP, Watt avg, HR, cadence, upcoming grade) and I will buy a subscription to your SaaS.
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.
This made me realize it's been a long time since I did some brainstorming and how I missed it.
The pelican meme is so much better than benchmaxxed stuff. Unfortunately we are likely now in pelicanmaxxing territory.
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.
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.
https://huggingface.co/baselabs/mammoth-1