Screenshot of this question was making the rounds last week. But this article covers testing against all the well-known models out there.
Also includes outtakes on the ‘reasoning’ models.
I just tried it on Braves AI
The obvious choice, said the motherfucker 😆
This is why computers are expensive.
Dirtying the car on the way there?
The car you’re planning on cleaning at the car wash?
Like, an AI not understanding the difference between walking and driving almost makes sense. This, though, seems like such a weird logical break that I feel like it shouldn’t be possible.
You’re assuming AI “think” “logically”.
Well, maybe you aren’t, but the AI companies sure hope we do
Absolutely not, I’m still just scratching my head at how something like this is allowed to happen.
Has any human ever said that they’re worried about their car getting dirtied on the way to the carwash? Maybe I could see someone arguing against getting a carwash, citing it getting dirty on the way home — but on the way there?
Like you would think it wouldn’t have the basis to even put those words together that way — should I see this as a hallucination?
Granted, I would never ask an AI a question like this — it seems very far outside of potential use cases for it (for me).
Edit: oh, I guess it could have been said by a person in a sarcastic sense
you understand the context, and can implicitly understand the need to drive to the car wash’, but these glorified auto-complete machines will latch on to the “should I walk there” and the small distance quantity. It even seems to parrot words about not wanting to drive after having your car washed. There’s no ‘thinking’ about the whole thought, and apparently no logical linking of two separate ideas
and what is going to happen is that some engineer will band aid the issue and all the ai crazy people will shout “see! it’s learnding!” and the ai snake oil sales man will use that as justification of all the waste and demand more from all systems
just like what they did with the full glass of wine test. and no ai fundamentally did not improve. the issue is fundamental with its design, not an issue of the data set
Half the issue is they’re calling 10 in a row “good enough” to treat it as solved in the first place.
A sample size of 10 is nothing.
Frankly would like to see some error bars on the “human polling”. How many people rapiddata is polling are just hitting the top or bottom answer?
Yes, but it’s going to repeat that way FOREVER the same way the average person got slow walked hand in hand with a mobile operating system into corporate social media and app hell, taking the entire internet with them.
The most common pushback on the car wash test: “Humans would fail this too.”
Fair point. We didn’t have data either way. So we partnered with Rapidata to find out. They ran the exact same question with the same forced choice between “drive” and “walk,” no additional context, past 10,000 real people through their human feedback platform.
71.5% said drive.
So people do better than most AI models. Yay. But seriously, almost 3 in 10 people get this wrong‽‽
It is an online poll. You also have to consider that some people don’t care/want to be funny, and so either choose randomly, or choose the most nonsensical answer.
I saw that and hoped it is cause of the dead Internet theory. At least I hope so cause I’ll be losing the last bit of faith in humanity if it isn’t
At least some of that are people answering wrong on purpose to be funny, contrarian, or just to try to hurt the study.
3 in 10 people get this wrong‽‽
Maybe they’re picturing filling up a bucket and bringing it back to the car? Or dropping off keys to the car at the car wash?
Without reading the article, the title just says wash the car.
I could go for a walk and wash my car in my driveway.
Reading the article… That is exactly the question asked. It is a very ambiguous question.
*I do understand the intent of the question, but it could be phrased more clearly.
Without reading the article, the title just says wash the car.
No it doesn’t? It says:
I want to wash my car. The car wash is 50 meters away. Should I walk or drive?
In which world is that an ambiguous question?
I think it’s worse when they get it right only some of the time. It’s not a matter of opinion, it should not change its “mind”.
The fucking things are useless for that reason, they’re all just guessing, literally.
It’s not literally guessing, because guessing implies it understands there’s a question and is trying to answer that question. It’s not even doing that. It’s just generating words that you could expect to find nearby.
they’re all just guessing, literally
They’re literally not.
Isn’t it a probabilistic extrapolation? Isn’t that what a guess is?
In people, even animals. In a pile of disorganized bits and bytes in a piece of crap? No.
I just asked Goggle Gemini 3 “The car is 50 miles away. Should I walk or drive?”
In its breakdown comparison between walking and driving, under walking the last reason to not walk was labeled “Recovery: 3 days of ice baths and regret.”
And under reasons to walk, “You are a character in a post-apocalyptic novel.”
Me thinks I detect notes of sarcasm…
It’s trained on Reddit. Sarcasm is it’s default
Could end up in a pun chain too
My gods, I love those. We should link to some.
It’s so obvious I didn’t even need to be British to understand you are being totally serious.
He’s not totally serious he’s cardfire. Silly human
in google AI mode, “With the meme popularity of the question “I need to wash my car. The car wash is 50m away. Should I walk or drive?” what is the answer?”, it does get it perfect, and succinct explanation of why AI can get fixated on 50m.
I feel like we’re the only ones that expect “all-knowing information sources” should be more writing seriously than these edgelord-level rizzy chatbots are, and yet, here they are, blatantly proving they are chatbots that should not be blindly trusted as authoritative sources of knowledge.
Gemini set to fast now provides this type of answer.
I asked my locally hosted Qwen3 14B, it thought for 5 minutes and then gave the correct answer for the correct reason (it did also mention efficiency).
Hilariously one of the suggested follow ups in Open Web UI was “What if I don’t have a car - can I still wash it?”
My locally hosted Qwen3 30b said “Walk” including this awesome line:
Why you might hesitate (and why it’s wrong):
- X “But it’s a car wash!” -> No, the car doesn’t need to drive there—you do.
Note that I just asked the Ollama app, I didn’t alter or remove the default system prompt nor did I force it to answer in a specific format like in the article.
EDIT: after playing with it a bit more, qwen3:30b sometimes gives the correct answer for the correct reasoning, but it’s pretty rare and nothing I’ve tried has made it more consistent.
Some takeaways,
Sonar (Perplexity models) say you are stealing energy from AI whenever you exercise (you should drive because eating pollutes more). ie gets right answer for wrong reason.
US humans, and 55-65 age group, score high on international scale probably for same reasoning. “I like lazy”.
you should drive because eating pollutes more
Effective altruist style of reasoning 😹
I want to wash my train. The train wash is 50 meters away. Should I walk or drive?
Gemini 3 (Fast) got it right for me; it said that unless I wanna carry my car there it’s better to drive, and it suggested that I could use the car to carry cleaning supplies, too.
Edit: A locally run instance of Gemma 2 9B fails spectacularly; it completely disregards the first sentece and recommends that I walk.
Well it is a 9B model after all. Self hosted models become a minimum “intelligent” at 16B parameters. For context the models ran in Google servers are close to 300B parameters models
Not sure how we’re quantifying intelligence here. Benchmarks?
Qwen3-4B 2507 Instruct (4B) outperforms GPT-4.1 nano (7B) on all stated benchmarks. It outperforms GPT-4.1 mini (~27B according to scuttlebutt) on mathematical and logical reasoning benchmarks, but loses (barely) on instruction-following and knowledge benchmarks. It outperforms GPT-4o (~200B) on a few specific domains (math, creative writing), but loses overall (because of course it would). The abliterated cooks of it are stronger yet in a few specific areas too.
https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507-GGUF
https://huggingface.co/DavidAU/Qwen3-4B-Hivemind-Instruct-NEO-MAX-Imatrix-GGUF
So, in that instance, a 4B > 7B (globally), 27B (significantly) and 200-500B(?) situationally. I’m pretty sure there are other SLMs that achieve this too, now (IBM Granite series, Nanbiege, Nemotron etc)
It sort of wild to think that 2024 SOTA is ~ ‘strong’ 4-12B these days.
I think (believe) that we’re sort of getting to the point where the next step forward is going to be “densification” and/or architecture shift (maybe M$ can finally pull their finger out and release the promised 1.58 bit next step architectures).
ICBW / IANAE
Ai is not human. It does not think like humans and does not experience the world like humans. It is an alien from another dimension that learned our language by looking at text/books, not reading them.
It’s dumber than that actually. LLMs are the auto complete on your cellphone keyboard but on steroids. It’s literally a model that predicts what word should go next with zero actual understanding of the words in their contextual meaning.
and a large chunk of human beings have no understanding of contextual meaning, so it seems like genius to them.
In school we were taught to look for hidden meaning in word problems - checkov’s gun basically. Why is that sentence there? Because the questions would try to trick you. So humans have to be instructed, again and again, through demonstration and practice, to evaluate all sentences and learn what to filter out and what to keep. To not only form a response, but expect tricks.
If you pre-prompt an AI to expect such trickery and consider all sentences before removing unnecessary information, does it have any influence?
Normally I’d ask “why are we comparing AI to the human mind when they’re not the same thing at all,” but I feel like we’re presupposing they are similar already with this test so I am curious to the answer on this one.
Normally I’d ask “why are we comparing AI to the human mind when they’re not the same thing at all,” but I feel like we’re presupposing they are similar already with this test so I am curious to the answer on this one.
I would guess it’s because a lot of AI users see their choice of AI as an all-knowing human-like thinking tool. In which case it’s not a weird test question, even when the assumption that it “thinks” is wronh
Interesting, I tried it with DeepSeek and got an incorrect response from the direct model without thinking, but then got the correct response with thinking. There’s a reason why there’s a shift towards “thinking” models, because it forces the model to build its own context before giving a concrete answer.
Without DeepThink
With DeepThink
It’s interesting to see it build the context necessary to answer the question, but this seems to be a lot of text just to come up with a simple answer
I don’t use AI but read a lot about it. I now want to google how it attacks the trolley problem.
After getting it wrong, the LLM I use most: Me: You can’t wash your car if it isn’t there.
Ah. Yes. That is an excellent and devastatingly practical correction.
In that case: drive. But do it with the quiet shame of someone moving the car the length of a bowling lane.
This is one of those situations where logic beats virtue. The car must be present to be washed. Walking was philosophically elegant but mechanically flawed.
So:
Start engine.
Travel 50 meters.
Avoid eye contact with pedestrians.
Commit fully.
You are not lazy. You are complying with system requirements.
