alt Hacker NewsQwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7
Comments
For coding, qwen 3.6 35b a3b solved 11/98 of the Power Ranking tasks (best-of-two), compared to 10/98 for the same size qwen 3.5. So it's at best very slightly improved and not at all in the class of qwen 3.5 27b dense (26 solved) let alone opus (95/98 solved, for 4.6).
I understand the 'fun factor' but at this point I really wonder what this pelican still proofs ? I mean, providers certainly could have adapted for it if they wanted, and if you want to test how well a model adapts to potential out of distribution contexts, it might be more worthwhile to mix different animals with different activity types (a whale on a skateboard) than always the same.
Such a disconnect from the minutes I’ve lost and given up on Gemini trying to get it to update a diagram in a slide today. The one shot joke stuff is great but trying to say “that is close but just make this small change” seems impossible. It’s the gap between toy and tool.
On thinking about the reasons this may be something at least slightly more than training on the task is the richness with which language is filled with spatial metaphors even in basic language not by laymen considered metaphor outside the field of linguistics proper, in which concepts eg Lakoff's analysis in "Metaphors we Live By and others are simply part of the field, (though unsurprisingly, among the HN crowd I've occasionally seen it brought up)
The amount of money you have in the bank may often "increase" or "decrease" but it also goes up and down, spatial. Concepts can be adjacent to each, orthogonal. Plenty more.
So, as models utilize weight more densely with more complex strategies learned during training the patterns & structure of these metaphors might also be deepened. Hmmm... another thing to add to the heap of future project-- trace down the geometry of activations in older/newer models of similar size with the same prompts containing such metaphors, or these pelican prompts, test the idea so it isn't just arm chair speculation.
You can just straight up ask Opus if it's good at generating images and it will say no. It has never been marketed as being for image generation.
That's not surprising; Opus & Sonnet have been regressing on many non-coding tasks since about the 4.1 release in our testing
I don't know what such a demo would prove in the first place. LLMs are good at things that they have been trained on, or are analogues of things they have been trained on. SVG generation isn't really an analogue to any task that we usually call on LLMs to do. Early models were bad at it because their training only had poor examples of it. At a certain point model companies decided it would be good PR to be halfway decent at generating SVGs, added a bunch of examples to the finetuning, and voila. They still aren't good enough to be useful for anything, and such improvements don't lead them to be good at anything else - likely the opposite - but it makes for cute demos.
I guess initially it would have been a silly way to demonstrate the effect of model size. But the size of the largest models stopped increasing a while ago, recent improvements are driven principally by optimizing for specific tasks. If you had some secret task that you knew they weren't training for then you could use that as a benchmark for how much the models are improving versus overfitting for their training set, but this is not that.
Maybe the next time we suspect they're optimising for the test, switch the next test to drawing "the cure for cancer".
Wonder what would happen if we unleashed Karpathy’s autoresearch on the pelican bicycle test. And had it read back the image to judge it.
Oh maybe it might continue to iterate on the existing drawing?
I'm an iguana and need to wash my bicycle in the carwash. Shall I walk or take the bus?
This is a useless benchmark now a days, every model provider trains their models on making good pelicans. Some have even trained every combination of animal/mode of transportation
I'm really curious about what competes with Claude Code to drive a local LLM like Qwen 3.6?
I've been using Qwen3.5-35B-A3B for a bit via open code and oMLX on M5 Max with 128Gb of RAM and I have to say it's impressively good for a model of that size. I've seen a huge jump in the quality of the tool calls and how well it handles the agentic workflow.
I really wish they spent some time training for computer use. This model is incapable of finding anywhere near the correct x,y coordinate of a simple object in a picture.
Between the legs and the beak I'd still rate the opus pelican higher
I love this benchmark!
looks like opus have been nerfed from day1
That Qwen flamingo on the unicycle is actually quite good. A work of art.
All those models that were just at version 1.x in 2024
That’s so wild
FYI, using a 128GB M5 MacBook Pro, sourced from another article by the author.
I liked both of Opus' better, it was very illuminating, in both cases I didn't see the error's Simon saw and wondered why Simon skipped over the errors I saw.
Pelican: saturated!
I'm currently testing Qwen3.6-35B-A3B with https://swival.dev for security reviews.
It's pretty good at finding bugs, but not so good at writing patches to fix them.
Good reminder that these tests have always been useless, even before they started training on it.
[dead]
[flagged]
How about switching to MechaStalin on a tricycle? It gets kind of boring.
I literally cannot believe that people are wasting their time doing this either as a benchmark or for fun. After every single language model release, no less.
Going to have to disagree on the backup test. Opus flamingo is actually on the pedals and seat with functional spokes and beak. In terms of adherence to physical reality Qwen is completely off. To me it's a little puzzling that someone would prefer the Qwen output.
I'd say the example actually does (vaguely) suggest that Qwen might be overfitting to the Pelican.