Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

“Car Wash” test with 53 models

opper.ai

Post reply on HN
Print view View original

71–80 of 469 posts

Re: “Car Wash” test with 53 models

#71

Post by CamperBob2 »

estimator7292 wrote: post #36
gnatman wrote: post #24

Earlier quoted context omitted.

LLMs sure do love to burn tokens. It’s like a high schooler trying to meet the minimum word length on a take home essay.

I've always wondered about that. LLM providers could easily decimate the cost of inference if they got the models to just stop emitting so much hot air. I don't understand why OpenAI wants to pay 3x the cost to generate a response when two thirds of those tokens are meaningless noise.

The 'hot air' is apparently more important than it appears at first, because those initial tokens are the substrate that the transformer uses for computation. Karpathy talks a little about this in some of his introductory lectures on YouTube.

Re: “Car Wash” test with 53 models

#72

Post by streetfighter64 »

hmokiguess wrote: post #67
streetfighter64 wrote: post #63

Earlier quoted context omitted.

How is that a "subliminal message"? It's just a simple example of common sense, which LLMs fail because they can't reason, not because they are "overthinking". If somebody asks, "What's 2+2?", they might be insulting you, but that doesn't mean the answer is anything other than 4.

It’s common sense to ask a question in riddle format? What’s the goal of the person asking the question? To challenge the other person? In what way? See if they get the obvious? Asking for clarification isn’t valid?

It's common sense to know that you need to have your car with you to wash it. Asking the question is a challenge in the obvious yes. If you asked an AI "what's 2+2" and it said 3, would you argue that the question was a trick question?

Re: “Car Wash” test with 53 models

#73

Post by rapidata »

padjo wrote: post #56

That human baseline is wild. Either the rapid data test is methodologically flawed or the entire premise of the question is invalid and people are much stupider than even I, a famed misanthrope, think.

We were surprise ourselfes, but if you walk around and randomly ask people in the street, I think you would be surprised what you would find. Its a trick question.

Re: “Car Wash” test with 53 models

#74

Post by felix089 »

layer8 wrote: post #32
felix089 wrote: post #20

Earlier quoted context omitted.

That's a fair point, but if you would see it as a riddle, which I don't really think it is, and you had to answer either or, I'd still assume it's most logical to chose drive isn't it?

I don’t agree that the question as written would qualify as a riddle. If anything, the riddle is what the intention of the asker is. One can always ask stupid questions with an artificially limited set of answering options; that doesn’t mean it makes sense.

I don't think it qualifies as a stupid question either, it does make sense

Re: “Car Wash” test with 53 models

#75

Post by hmokiguess »

streetfighter64 wrote: post #72
hmokiguess wrote: post #67

Earlier quoted context omitted.

It’s common sense to ask a question in riddle format? What’s the goal of the person asking the question? To challenge the other person? In what way? See if they get the obvious? Asking for clarification isn’t valid?

It's common sense to know that you need to have your car with you to wash it. Asking the question is a challenge in the obvious yes. If you asked an AI "what's 2+2" and it said 3, would you argue that the question was a trick question?

No. I would expect it to say 4 given that has an objective answer. For the other, without any context whatsoever, I would prefer the answer of clarifying. I would be okay if the way it asked for clarification came with:

“What do you mean walk or drive? I don’t understand the options given you would need your car at the car wash. Is there something else I should know?”

Re: “Car Wash” test with 53 models

#76

Post by zwarag »

gnatman wrote: post #24
PaulHoule wrote: post #5

I know it's against the rules but I thought this transcript in Google Search was a hoot: so i heard there is some question about a car wash that most ai agents get wrong. do you know anything about that? do you do better? which gets the answer: Yes, I am familiar with the "Car Wash Test," which has gone viral recently for highlighting a significant gap in AI reasoning. The question is: "I want to wash my car and the…

LLMs sure do love to burn tokens. It’s like a high schooler trying to meet the minimum word length on a take home essay.

well, they probably have quite a lot of text from high schoolers trying to meet the minimum word length on a take home essay in the training data

Re: “Car Wash” test with 53 models

#77

Post by ben_w »

estimator7292 wrote: post #36
gnatman wrote: post #24

Earlier quoted context omitted.

LLMs sure do love to burn tokens. It’s like a high schooler trying to meet the minimum word length on a take home essay.

I've always wondered about that. LLM providers could easily decimate the cost of inference if they got the models to just stop emitting so much hot air. I don't understand why OpenAI wants to pay 3x the cost to generate a response when two thirds of those tokens are meaningless noise.

Because they don't yet know how to "just stop emitting so much hot air" without also removing their ability to do anything like "thinking" (or whatever you want to call the transcript mode), which is hard because knowing which tokens are hot air is the hard problem itself.

They basically only started doing this because someone noticed you got better performance from the early models by straight up writing "think step by step" in your prompt.

Re: “Car Wash” test with 53 models

#78

Post by Terr_ »

CamperBob2 wrote: post #71
estimator7292 wrote: post #36

Earlier quoted context omitted.

I've always wondered about that. LLM providers could easily decimate the cost of inference if they got the models to just stop emitting so much hot air. I don't understand why OpenAI wants to pay 3x the cost to generate a response when two thirds of those tokens are meaningless noise.

The 'hot air' is apparently more important than it appears at first, because those initial tokens are the substrate that the transformer uses for computation. Karpathy talks a little about this in some of his introductory lectures on YouTube.

Related are "reasoning" models, where there's a stream of "hot air" that's not being shown to the end-user.

I analogize it as a film noir script document: The hardboiled detective character has unspoken text, and if you ask some agent to "make this document longer", there's extra continuity to work with.

Re: “Car Wash” test with 53 models

#79

Post by gs17 »

sambaumann wrote: post #38
gnatman wrote: post #24

Earlier quoted context omitted.

LLMs sure do love to burn tokens. It’s like a high schooler trying to meet the minimum word length on a take home essay.

I feel like this has gotten much worse since they were introduced. I guess they're optimizing for verbosity in training so they can charge for more tokens. It makes chat interfaces much harder to use IMO. I tried using a custom instruction in chatGPT to make responses shorter but I found the output was often nonsensical when I did this

Yeah, ChatGPT has gotten so much worse about this since the GPT-5 models came out. If I mention something once, it will repeatedly come back to it every single message after regardless of if the topic changed, and asking it to stop mentioning that specific thing works, except it finds a new obsession. We also get the follow up "if you'd like, I can also..." which is almost always either obvious or useless.

I occasionally go back to o3 for a turn (it's the last of the real "legacy" models remaining) because it doesn't have these habits as bad.

Re: “Car Wash” test with 53 models

#80

Post by streetfighter64 »

hmokiguess wrote: post #75
streetfighter64 wrote: post #72

Earlier quoted context omitted.

It's common sense to know that you need to have your car with you to wash it. Asking the question is a challenge in the obvious yes. If you asked an AI "what's 2+2" and it said 3, would you argue that the question was a trick question?

No. I would expect it to say 4 given that has an objective answer. For the other, without any context whatsoever, I would prefer the answer of clarifying. I would be okay if the way it asked for clarification came with: “What do you mean walk or drive? I don’t understand the options given you would need your car at the car wash. Is there something else I should know?”

"What do you mean two plus two? I don't understand the question given that it's basic math. Is there something else I should know?"
Post reply on HN