I don’t even
I was surprised that both Grok and Gemini 2.5 got it right once, only to fuck it up on the refresh
Omg
Some of these are absolutely hilarious
This is my favorite
gaslight clock
So the prompt is intentionally breaking things…?
Edit: I guess this was somehow the chatty output doing who know wtf
Really cool idea, but the site seems a bit biased for the chinese models, or is otherwise set up weird. I’m not able to reproduce how consistently bad the others are in web dev arena, which generally accepted as the gold standard for testing AI web dev ability.
Each model is allowed 2000 tokens to generate its clock. Here is its prompt: Create HTML/CSS of an analog clock showing ${time}. Include numbers (or numerals) if you wish, and have a CSS animated second hand. Make it responsive and use a white background. Return ONLY the HTML/CSS code with no markdown formatting.
are you using the same prompt?
There’s a couple differences. It’s giving it the current time as part of the prompt, which is interesting. The other difference is that it’s asking it to make it responsive. But even when I use that exact prompt (inserting the time obv), it works fine on claude, openai, and gemini.
So there’s definitely an issue specific to this page somewhere. Maybe it’s not iframing them? I’m on mobile so I can’t check.
Given that the AI models are basically constructing these “blindly”- using the language model to string together html and javascript without really being able to check how it looks- some of these are actually pretty impressive. But also making the AI do things it’s bad at is funny. Reminds me of all the AI ASCII art fails…
deleted by creator
Well, KIMI K2 seems to have created the working one. Others failed. I suppose that this model was optimized for this while others not.
The clocks change every minute. I’ve seen some from deepseek and qwen that looked ok. But kimi seems to be the most consistent
Removed by mod
Thanks for sharing that mindfuck. I honestly would’ve thought something was wrong with my cognition if you hadn’t mentioned it was a test beforehand.
that’s scary how dementia works :'(
This kinda freaked me out: AI models fed their own outputs as training data will quickly start making distorted images that look spookily like human painting made under the progression of mental illness or drugs.
well that was terrifying
Thanks for sharing this! I really think that when people see LLM failures and say that such failures demonstrate how fundamentally different LLMs are from human cognition, they tend to overlook how humans actually do exhibit remarkably similar failures modes. Obviously dementia isn’t really analogous to generating text while lacking the ability to “see” a rendering based on that text. But it’s still pretty interesting that whatever feedback loops did get corrupted in these patients led to such a variety of failure modes.
As an example of what I’m talking about, I appreciated and generally agreed with this recent Octomind post, but I disagree with the list of problems that “wouldn’t trip up a human dev”; these are all things I’ve seen real humans do, or could imagine a human doing.
such a variety of failure modes
What i find interesting is that in both cases there is a certain consistency in the mistakes too - basically every dementia patient still understands the clock is something with a circle and numbers and not a square with letters for example. LLMs can tell you cokplete bullshit, but still understands it has to be done with perfect grammar in a consistant language. So much so it struggles to respond outside of this box - ask it to insert spelling errors to look human for example.
the ability to “see”
This might be the true problem in both cases, both the patient and the model can not comprehend the bigger picture (a circle is divided into 12 segments, because that is how we deconstructed the time it takes for the earth to spin around it’s axis). Things that seem logical to use, are logical because of these kind of connections with other things we know and comprehend.
… what
Basically this: https://www.psychdb.com/cognitive-testing/clock-drawing-test
Thanks.
What I still didn’t figure out about the comment I replied to is:
- What is each row? They’re labeled I, II, III, IV. What’s being counted?
- Why did they link to a home interior design website under “via”?
Get educated
Is it just me or the clocks frequently break or change appearance without the page being refreshed?
Edit: nevermind, I skipped past the sentence explaining that every minute, the site prompts LLMs for a new solution. This is hilariously sad how LLMs aren’t able to be consistent from one prompt to another.
It’s the expected result if your big ol’ artificial intelligence wannabe is ultimately just a stochastic word combinator.
if every single token is, at the end, chosen by random dice roll (and they are) then this is exactly what you’d expect.
that’s a massive oversimplification
This is hilariously sad how LLMs aren’t able to be consistent from one prompt to another.
Typically that’s configurable. Like for a chatbot, you’d want it to give the same/similar results for a given question, where with a character creator, you might want the results to vary so you can re-run until you get something you like.
Of course that wouldn’t be as funny here.
You can click on the button in the top right corner (with a question mark) to have explanations. The clocks are refreshed every minute
It’s funny how GPT-5 is consistently the worst one, and it’s not even close.
qwen 2.5 is absolutely pants on head ridiculous compared to gpt5 when I’m looking at it right now.
The last one, Kimi K2, has been consistently good as long as I’ve been looking at it. That’s pretty impressive.
The rest are hilarious!
By far the best, but still off. These three were loaded in the same order as i post them:
Haha, I found myself thinking the same thing, and then caught myself, realizing all the other LLMs on this page had lowered the bar immensely for what I’m considering impressive.
Deepseek has a recognisable clock now and then, too. They both mix up the current time, though.
I thought the same and then Kimi K2 came up with a clock that has two 12 and no 11…
qwen is trying her best 😭😔
So far, I’d give qwen the prize for most artistic impression of a clock.
Kimi K2 appears to consistently get it right.
And just as I typed that, Kimi made one where 9 and 10, and 11 and 12 overlapped.
What is this obsession with clocks recently?
I don’t know if it’s actually related, but I’ve read that asking people to draw a clock face is a simple way to identify some brain problems
Quick screening for dementia, according to this
Edit: I guess this means most of the AI has ‘Conceptual Deficits’, pretty accurate lol
Would be funny if AI models are generating such wildly useless “clocks” because they ingested too many dementia screening tests in their training data
There is someone training the biggest, bestest model to draw clock faces to pass that test as we speak.
I’m guessing it’s an easy metric to compare benchmarks. “Write a clock”.
Gpt 5 👍
Ummm…
You know, I don’t, and what the fuck?
