73 comments

[ 0.25 ms ] story [ 22.6 ms ] thread
Another day, another model and another pelican :-)

I can't help but wonder where is the trend going? What will we have in five years? Maybe it will all have puttered out, and we will have moved to the next thing? Or maybe the prompt then will be "make a pelican ride a bicycle", and out will come the genetic code for a giant pelican with extremities suitable for a handle bar and pedals, and an inborn affinity to ride bicycles?

I would be surprised if pelican svgs are not part of the training corpus rn
That’s covered in the article
As mentioned elsewhere, the benchmark introduces bad pelicans in the training set. What I'm curious about however, if it's possible for a human artist to "poison" the benchmark by releasing some really good pelicans svgs and have all future models output their version.
> How does the prompt “Generate an SVG of a pelican riding a bicycle” add up to 95 input tokens? OpenAI’s tokenizer counts 10, Anthropic’s counts 10 for Opus 4.6, 30 for Opus 4.7 and 25 for Sonnet 5/Fable 5. Prompting “hi” to Kimi K3 counted 86 tokens, suggesting there may be an 85 token hidden system prompt. It refused to leak it though.

This is quite possibly reasoning-effort prompt which is injected before the opening <think> token whenever you set a custom reasoning effort, see e.g. DeepSeek-V4 max mode prompt: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main...

My personal benchmark for new models has been to compare video making skills with something like remotion. Usually reveals if they have any "taste" or outside the box thinking.

I'm starting to not trust any "benchmarks" when it comes to frontier models at least. As an example Sol feels the most "gets stuff done" but has zero taste, or any capability to surprise.

And for frontier models I go one step ahead and try to recreate a complex animation video, with the ability for the model to review its own work. And at this Fable is still the top one. Ex: https://www.youtube.com/watch?v=uDAeAuYyl0E (recreation of Claude announcement video) and https://www.youtube.com/watch?v=cSsVNtGPOIg (recreation of a fireship video). Sol did something similar but you can instantly tell its AI slop from very small things, and it just has no narrative or thought put into the writing.

https://mesmer.tools/benchmarks/ai-video-generation , I usually put basic ones here.

> This is expensive—the pelican cost 25 cents!

Engineers get unbelievably silly about evaluating costs of things.

"The tokens are so expensive!" Oh my sweet child, how much would even the least capable human effort cost? This is what the executives properly understand that the programmers don't.

It's incredible Simon still believes pelicans on bikes aren't part of the training set, despite hundreds of them on blogs, forums, and Github. Stuff we put in our company blog shows up known by LLMs 6 months later, and we have 1000x less traffic than Simon's own website
Yeah I asked Nano Banana to make a render of our company office and was scarily accurate
It's incredible people still discuss the pelicans... But then again, the ad just works.
It's incredible you can't reason to see if pelican on a bike is a thing. It's not! This has been discussed to death. You can ask any model to generate anything. Generate an SVG of earthworm and a robin boxing. Guess what? The smarter the model the better the image, doesn't matter if it's a vision model or not. I rolled my eyes at this eval when I first saw it, then I tried various ridiculous things and noticed a very strong correlation. Things that are absolutely not in the training set.
This reminded me about the news cycle last year that we were running out of training data (and how silly that was)
(comment deleted)
(comment deleted)
Imagine what amazing SVG generators we could have if Simon had randomized the target image from the start (and companies wouldn't just overfit on pelicans).
Is there a gallery of all pelicans generated by simon over time?
K3 is as expensive as Sonnet, not great at writing English, is handing IP back to the Chinese, and once open source will be difficult to run at scale without the compute that OpenAI and Anthropic have largely grabbed.

Sorry, how again is this the end of the frontier labs?

Well, with the actions of the US government, for every business that does not exclusively operate in the US, they have now added _supplier risk_ to US companies.

Even as a paying customer, even as an enterprise, your access to US models may be turned off at any time for arbitrary reasons, including someone mis-understanding "Please fix this [open source] code" (which contained security vulnerabilities that were fixed) as a jailbreak.

Do any of the vision models render the SVG and look at the result.

Perhaps more importantly can they do that during reinforcement training. Learning how to critically analyse the appearance of what it generates would be quite useful.

Manually feeding images back to models has been hilariously bad in the past which suggests that relating something it sees to something it wrote is not an ability it is very good at.

I've tried doing a loop of rending the SVG and then tweaking based on that, with local models (so, not nearly as strong). It wasn't very successful; it would mostly report that the image looked great and didn't need any tweaks. Maybe I should try it again, there have been some newer models since I first tried it. And yeah, maybe worth trying with bigger models. But I have found that models aren't necessarily the best at visual reasoning and review, even with a vision loop. Their lack of visual reasoning is part of why they still have trouble with things like ARC-AGI-3.
Don't see why we have to have this spammed every model release when Fable class models perform the same as Opus on basic tasks like these.
What spam? It’s one article. You can skip it
I think the user should be banned. It’s insane spam
The idea is not to use pelicans on bikes but a similarly random non-sensical prompts: crows on scooters, squirrels in a moon rover etc. Then pick another one for another for next cross-llm evaluation.
I wonder how the Chinese labs are training a 3 trillion parameter model on what has to be vastly smaller compute resources. If the U.S. compute advantage is persistent, it's hard to imagine that Chinese labs will be able to keep pace forever, as a matter of physics, but... so far they seem to be doing just fine.
The Chinese just saved the world economy by draining their absurdly enormous oil storage reserves nobody knew they had, wouldn't surprise me if they had lots of hidden compute too.
Or they just don't actually have any compute access restrictions of significance? Chinese companies can just go use those GPUs in neighboring countries that aren't export-restricted, like Malaysia. Like ByteDance openly did: https://www.tomshardware.com/pc-components/gpus/chinas-byted...

and Tencent is rumored to have done via Japan: https://wccftech.com/china-tencent-gains-access-to-nvidia-bl...

And that's not even considering just smuggling the GPUs in by eg buying them in Singapore.

AI-specific chips also seem to be on the easier side to design & create relative to high performance CPUs & GPUs, so there's no particular reason to expect Chinese domestic designs to continuously lag behind. They have access to the same fabs, after all

Gamers Nexus has a good video where his team travelled to China and did some actual investigative journalism: https://www.youtube.com/watch?v=1H3xQaf7BFI

Firstly, the export-restricted GB202s (e.g. 5090, RTX 6000 Pro Blackwell) are fabled in TSMC, and then packaged/made in... China before they supposedly have to be sold out (by US law; but not by Chinese law). You can immediately see the problem there.

Secondly, despite the supposed 'crackdowns' and et al, NVIDIA and their channel partners pretty much will sell to anyone in countries like Singapore without any questions.

Third, there's human "smugglers" who just physically carry em on trips, and Chinese customs is obviously not going to care about the US's laws on Chinese soil.

Huawei Ascend chips were used to train DeepSeek v4 over 4 months ago, and they shared their kernel with the other Chinese labs. China also has their own DDR5 fabs.
It’s not bad kind of expensive for 25c but if the prompt is rendered cost is much better.
The gap is closing . I think Kimi 3 is only 3 months behind the US model. It’s gpt 5.5 class model , which was released in the end of April.
Like Simon concludes the article, the main use of this isn't to say which model is "better", but to try and poke at the model to sort out things like quality vs cost vs speed.

So I put together a quick comparison of the last couple iterations of Opus, Fable and now Kimi.

Kimi is cheapest by 5x but also slowest by 2x

https://9gpyw4uxr2.evvl.io/

Personally I'd consider the three middle ones to be failing, in the typical "Gemini/Google" fashion in that the model is doing more than what the prompt asks. The prompt asks for SVG, yet the model is providing more.

Edit: Actually, looking at the K2.6 response, that's borderline failing too, it's using HTML+CSS+SVG, not just SVG, again failing to follow the prompt properly.

By the way, that website seems like a black hole for information, it says "Expires in 6 days" in the top right which seems really weird for a page hosting couple of KB of data at most.

It's a free site, so I was trying to limit both the privacy and risk exposure.

Making the content auto expire after a short period of time greatly decreases the attractiveness of the site to lots of SEO spammers and other types of abuse, and if someone were to get something malicious or vile posted it will clean up after itself without me having to wade into things.

3T is impressive, but parameter count seems to be less important than I thought.

GLM is half the size of DeepSeek but costs four times as much, and beats it on every benchmark.

I'm not an expert on this stuff but it seems to be the attention mechanism. DeepSeek were bragging about how cheap they made it. But if you cut costs on attention you get worse results with way more parameters.

If I had to guess it seems to be the difference between memory (params) and intelligence (attention density). I think you need both.

After MoE entered the mix, raw parameter count is less useful a measure.
Or, GLM 5.2 simply had more time in the RL oven.

Deepseek V4 Flash, the 284B model, is roughly equivalent to launch GLM 5, the 744B [sic] model.

It's almost like they priced models based on their performance or something...
You have to look at the size of each expert; Kimi's has about 50G parameters while GLM's has 40G. The number of the experts tells you about the diversity of its skills.
It will be valuable to have two types of benchmarks: ones that evolve alongside the models and ones that never change. You probably can't get historical stability and resistance to flooding and training on at least some parts of it from the same test
Kimi is right out since they use classical music branding to sell their slop. At least McDonalds does not sell Verdi or Allegro burgers.

Why does Kimi not use a "Double Cheese Whammy" branding for "their" butchered and stolen IP?

One thing i keep thinking: you only run the pelican once per model. Run the same model a few times and you get some different pelicans, so some of "this one is better" might just be which run you picked for it. Would love to see 8 runs per model side by side. I bet for two close models, the gap between runs is about as big as the gap between the models.
Old and busted: benchmaxxing

New hotness: pelicanmaxxing