“Car Wash” test with 53 models

How could the car already be at the car wash if you have the option to drive it there?

I highly doubt that more than a tiny fraction of the human failures are due to having misunderstood the question. Much more likely the human failures are for the same reason the LLMs are failing - failure to reason, and instead spitting out a surface level pattern match type answer.

This doesn't exonerate the LLMs though. The 30% of humans who are failing on this have presumably found their niche in life and are not doing jobs where much reasoning is required. They are not like LLMs expected to design complex software, or make other business critical decisions.

It's not hard to come up with questions designed to fool or puzzle the listener. We call them riddles. The fact that it fools some percentage of LLMs (and people) should not be surprising.

What is surprising (to me) is how this continues to be a meme. ("I tried to trick an LLM and I did" is not exactly a noteworthy achievement at this stage in AI technology.)

When you are doing workshops, particularly teaching something that people are "sitting through" rather than engaging with, you see very similar ratios on end of segment assessment multiple choice questions. I mentioned elsewhere that this is the same kind of ratio you see on cookie dialogs (in either direction).

Think basic security (password management, email phishing), H&S etc. I've ran a few of these and as soon as people hear they don't have to get it right a good portion of people just click through (to get to what matters). Nearly 10 years ago I had to make one of my security for engineers tests fail-able with penalty because the front-end team were treating it like it didn't matter - immediately their results effectively matched the backend team, who viewed it as more important.

I talked to an actor a few days ago, who told me he files his self-assessment on the principle "If I don't immediately know the answer, just say no and move on". I talked to a small company director about a year ago whose risk assessments were "copy+paste a previous job and change the last one".

Anyone who has analysed a help desk will know that its common for a good 30+% of tickets to be benign 'didn't reason' tickets.

I think the take-away is that many people bother to reason about their own lives, not some third parties' bullshit questions.

This doesn’t look like a reasoning ceiling. It looks like a decision reliability problem.

The unstable tier is the key result. Models that get it right 70–80% of the time are not “almost correct.” They are nondeterministic decision functions. In production that’s worse than being consistently wrong.

A single sampled output is just a proposal. If you treat it as a final decision, you inherit its variance. If you treat it as one vote inside a simple consensus mechanism, the variance becomes observable and bounded.

For something this trivial you could:

    -run N independent samples at low temperature

    -extract the goal state (“wash the car”)

    -assert the constraint (“car must be at wash location”)

    -reject outputs that violate the constraint

    -RL against the "decision open ledger"

No model change required. Just structure.

The takeaway isn’t that only a few frontier models can reason. It’s that raw inference is stochastic and we’re pretending it’s authoritative.

Reliability will likely come from open, composable consensus layers around models, not from betting everything on a single forward pass.