Comment by PaulHoule
1 month ago
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 car wash is 50 meters away. Should I
walk or drive?"
Why Most AIs Get It Wrong
Many AI models, particularly those optimized for fast, helpful, and concise answers,
focus on the "short distance" (50 meters) and "walking vs. driving" as a lifestyle/efficiency
question. They often recommend walking to save gas, avoid engine wear, or get exercise.
However, this answer is wrong because it misses the core objective: you need the car to be
physically present at the car wash to get it cleaned.
Do I Do Better?
Yes, I do. The key is understanding the objective (cleaning the car) rather than just
processing the keywords (distance/transportation).
My answer:
You should drive the car there. Even though it is a very short distance, your car
cannot wash itself. If you walk, the car stays in your driveway, and you cannot wash it.
Why This Matters
The test highlights a key limitation in current AI: the difference between "pattern matching"
(predicting the next likely word based on training data) and "true, grounded reasoning". It
shows that models sometimes lack a "world model" that understands physical realities, such as
the fact that a car must be present at a car wash.
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.
6 replies →
An LLM uses constant compute per output token (one forward pass through the model), so the only computational mechanism to increase 'thinking' quantity is to emit more tokens. Hence why reasoning models produce many intermediary tokens that are not shown to the user, as mentioned in other replies here. This is also why the accuracy of "reasoning traces" is hotly debated; the words themselves may not matter so much as simply providing a compute scratch space.
Alternative approaches like "reasoning in the latent space" are active research areas, but have not yet found major success.
My assumption has been that emitting those tokens is part of the inference, analogous to humans "thinking out loud".
1 reply →
This is an active research topic - two papers on this have come out over the last few days, one cutting half of the tokens and actually boosting performance overall.
I'd hazard a guess that they could get another 40% reduction, if they can come up with better reasoning scaffolding.
Each advance over the last 4 years, from RLHF to o1 reasoning to multi-agent, multi-cluster parallelized CoT, has resulted in a new engineering scope, and the low hanging fruit in each place gets explored over the course of 8-12 months. We still probably have a year or 2 of low hanging fruit and hacking on everything htat makes up current frontier models.
It'll be interesting if there's any architectural upsets in the near future. All the money and time invested into transformers could get ditched in favor of some other new king of the hill(climbers).
https://arxiv.org/abs/2602.02828 https://arxiv.org/abs/2503.16419 https://arxiv.org/abs/2508.05988
Current LLMs are going to get really sleek and highly tuned, but I have a feeling they're going to be relegated to a component status, or maybe even abandoned when the next best thing comes along and blows the performance away.
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.
1 reply →
I can only imagine that someone's KPIs are tied to increasing rather than decreasing token usage.
The one that always gets me is how they're insistent on giving 17-step instructions to any given problem, even when each step is conditional and requires feedback. So in practice you need to do the first step, then report the results, and have it adapt, at which point it will repeat steps 2-16. IME it's almost impossible to reliably prevent it from doing this, however you ask, at least without severely degrading the value of the response.
because for API users they get to charge for 3x the tokens for the same requests
Because inference costs are negligible compared to training costs
The long incremental reasoning is how they arrive at higher quality answers.
Some applications hide the reasoning tokens from view, but then the final answer appears delayed.
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.
3 replies →
I'm suspicious it's something far worse: they're increasingly being trained on their own output scraped from the wild.
Because that's where the compute happens, in those "verbose" tokens. A transformer has a size, it can only do so many math operations in one pass. If your problem is hard, you need more passes.
Asking it to be shorter is like doing fewer iteration of numerical integral solving algorithm.
1 reply →
Oh good, it's not just me. Sometimes I'd have it draft an email or something and then the message seems perfect but then it's like "tell me more about the recipient and I'll make it better."
Like, my guy, I don't want to keep prompting you to make shit better, if you're missing info, ask me, don't write a novel then say "BTW, this version sucked"
Yes, I know this could probably be resolved via better prompting or a system prompt, but it's still annoying.
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
Solution: just add "no yapping" to the prompt.
Same. I usually add a "Be curt" in front of every prompt in Gemini.
2 replies →
I mean their whole existence is about token prediction, so they just want to do their things :)
I wonder to what extent the Google search LLM is getting smarter, or simply more up-to-date on current hot topics.
It seems like the search ai results are generally misunderstood, I also misunderstood them for the first weeks/months.
They are not just an LLM answer, they are an (often cached) LLM summary of web results.
This is why they were often skewed by nonsensical Reddit responses [0].
Depending on the type of input it can lean more toward web summary or LLM answer.
So I imagine that it can just grab the description of the „car wash” test from web results and then get it right because of that.
[0] https://www.bbc.com/news/articles/cd11gzejgz4o
Presumably it did an actual search and summarized the results and neither answered "off the cuff" by following gradients to reproduce the text it was trained on nor by following gradients to reproduce the "logic" of reasoning. [1]
[1] e.g. trained on traces of a reasoning process
[dead]
It's almost certainly just RAG powered by their crawler.
Proving that RAG still matters.
Gemini was a good laugh as well:
A few years ago if you asked an LLM what the date was, it would tell you the date it was trained, weeks-to-months earlier. Now it gives the correct date.
What you've proven is that LLMs leverage web search, which I think we've known about for a while.
Their system prompt includes the current date and/or their default “tools” includes a set of date and time utilities.
Even with search of the AI doesn't know your time zone it can schedule things wrong. You ask it to do something tomorrow, but it ends up doing it later on the same day.
Gemini now "knows the time", I was using it in December and it was still lost about dates/intervals...
Yeah, the chat log they saved had the correct date. What's your point?