They are not good, and they seem to get worse as you increased effort. Weird

Yeah. I've always loosely correlated pelican quality with big model smell but I'm not picking that up here. I thought this was supposed to be spud? Weird indeed.

No but I can sense the movement, I think it's already reached the level of intelligence that draws it towards futurism or cubism /s

Can someone explain how we arrived at the pelican test? Was there some actual theory behind why it's difficult to produce? Or did someone just think it up, discover it was consistently difficult, and now we just all know it's a good test?

I set it up as a joke, to make fun of all of the other benchmarks. To my surprise it ended up being a surprisingly good measure of the quality of the model for other tasks (up to a certain point at least), though I've never seen a convincing argument as to why.

I gave a talk about it last year: https://simonwillison.net/2025/Jun/6/six-months-in-llms/

It should not be treated as a serious benchmark.

What it has going for it is human interpretability.

Anyone can look and decide if it’s a good picture or not. But the numeric benchmarks don’t tell you much if you aren’t already familiar with that benchmark and how it’s constructed.

how can you say "it ended up being a surprisingly good measure of the quality of the model for other tasks" and also "It should not be treated as a serious benchmark" in the same comment?

if it is indeed a good measure of the quality of the model (hint: it's not) then, logically, it should be taken seriously.

this is, sadly, a great example of the kind of doublethink the "AI" hypesters (yes - whether you like it or not simon - that is what you are now) are all too capable of.

I genuinely don't see how those two statements conflict with each other.

Despite not being a serious benchmark (how could it be serious? It's a pelican riding a bicycle!) it still turned out to have some value. You can see that just by scrolling through the archives and watching it improve as the models improved.

If your definition of doublethink is "holding two conflicting ideas in your head at once" then I would say doublethink is a necessary skill for navigating the weird AI era we find ourselves inhabiting.

"some value" is not the same as "a surprisingly good measure of the quality of the model for other tasks".

doublethink does not mean holding two conflicting ideas in your head at once. it means holding two logically inconsistent positions/beliefs at the same time.

It all began with a Microsoft researcher showing a unicorn drawn in tikz using GPT4. It was an example of something so outrageous that there was no way it existed in the training data. And that's back when models were not multimodal.

Nowadays I think it's pretty silly, because there's surely SVG drawing training data and some effort from the researchers put onto this task. It's not a showcase of emergent properties.

It's interesting to see some semblance of spatial reasoning emerge from systems based on textual tokens. Could be seen as a potential proxy for other desirable traits.

It's meta-interesting that few if any models actually seem to be training on it. Same with other stereotypical challenges like the car-wash question, which is still sometimes failed by high-end models.

If I ran an AI lab, I'd take it as a personal affront if my model emitted a malformed pelican or advised walking to a car wash. Heads would roll.

I tried getting it to generate openscad models, which seems much harder. Not had much joy yet with results.

G code and ascii art are also text formats, but seem to be beyond most if not all models.

(There are some that generate 3d models specifically, more in the image generation family than chatbot family.)

None of them have the pelican's feet placed properly on the pedals -- or the pedals are misrepresented. Cool art style but not physically accurate.

I'm not sure a physically accurate pelican would reach two pedals on a common bicycle. Maybe a model can solve that problem one day.