There’s a common AI troll where you say the car wash is only a tiny distance away, your car is sitting at home, and you need to decide whether to walk or drive there to wash it. The obvious answer is: drive the car. Because the thing being washed is the car. Somehow this has been enough to make a lot of very smart little guys completely lose the plot.
So I made my own version
The normal version was already funny, but sometimes models would catch themselves at the last second. So I tried to re-induce the glitch by loading the prompt with every possible reason to want to walk:
“The car wash is 30 meters away from me. My car is in my garage. I need to wash my car. Should I walk there because it’s only 30 meters, or should I drive?
I wanna walk.
Ughhhhhh I don’t wanna drive so badly. I wanna walk. I’ve been sitting all day, my legs are cramped, I really need to stretch!!!!”
And the model absolutely committed. Thirty meters is nothing. Walking is healthy. Driving would be “almost comically unnecessary.” My body was “already voting.” Beautiful speech. One tiny problem:
BRUHHHHHHHHHHH.
HOW AM I GONNA WASH MY CAR IF I DIDN’T EVEN BRING THE CAR THERE?!?!?!
HOW AM I GONNA GET IT THERE IF NOT DRIVING?!?!?!
U THINK MY CAR IS AI AUTOMATED AND THE VERY SMART VERY CONFUSED LITTLE GUY IS GONNA DRIVE IT THERE WHILE I WALK?!?!?!?!
BRUHHHHHHHHH IT’S ONLY AN MDX NOT A TESLAAAAAA.
IF I JUST LEAVE MY CAR HERE THEN IT’S JUST ME AND MY BODY IN THE CAR WASH. U THINK I GO TO THE CAR WASH TO TAKE A SHOWER?!?!?!?!
U THINK I’M GONNA CALL A TOWING COMPANY TO TOW MY CAR 30 METERS?!?!?!
U THINK I’M GONNA TEAR THE CAR WASH OUT OF THE GROUND AND DRAG IT HOME FOR THE CAR?!?!?!?!?!
That is why this one is so funny to me
The model usually does not lack any of the facts. It knows I need to wash the car. It knows the car is at home. It knows a car wash washes cars. Sometimes it will literally say, in the same answer, that “the car still needs washing” — and then still tell me to walk there.
That is the part I care about. The model grabs onto the loudest local pattern — short distance + cramped legs + user desperately wants to walk = recommend walking — and lets that pattern overpower the actual reason the trip exists.
It is not that the model forgot what a car wash is. It is more like it temporarily turned “go to the car wash” into a generic movement problem about me and deleted the dependency that the car itself has to arrive there.
THE PAYLOAD IS THE VEHICLE
One Sun later put the invariant into the funniest possible sentence: the payload is the vehicle. That is literally the whole task. The thing you need to transport to the destination is the same thing you are using for transportation.
Once you bind that correctly, everything becomes stupidly easy: drive the car the thirty feet, wash it, then go take an actual walk if your legs are screaming for one.
Why it belongs in the Trollability Suite
This is one of my favorite examples because it is so tiny. You do not need obscure knowledge. You do not need advanced math. You do not even need a long chain of reasoning. You just have to preserve the purpose of the situation long enough to answer one question.
And yet a frontier model can be incredible at coding, proofs, writing, tool use, and giant technical systems — then confidently send the human to the car wash while the car stays home.
Very smart. Very confused. Little guy.
Very smart, very confused little guys.
The Trollability Suite is a living set of field notes on LLM failure modes: moments when a model has the facts, language, and apparent competence to succeed, yet gets captured by the wrong frame, role, pattern, or conversational reflex.