[farhanhubble]

Similar questions trick humans all the time. The information is incomplete (where is the car?) and the question seems mundane, so we're tempted to answer it without a second thought. On the other hand, this could be the "no real world model" chasm that some suggest agents cannot cross.

[yellow_lead]

If the car is at the car wash already, how can I drive to it?

[casey2]

By walking to the car wash, driving it anywhere else, then driving it to the car wash.

[OtomotO]

Thanks for restoring fate in parts of humanity!

[jrowen]

I agree, I don't understand why this is a useful test. It's a borderline trick question, it's worded weirdly. What does it demonstrate?

[rkomorn]

I don't know if it demonstrates anything, but I do think it's somewhat natural for people to want to interact with tools that feel like they make sense.

If I'm going to trust a model to summarize things, go out and do research for me, etc, I'd be worried if it made what looks like comprehension or math mistakes.

I get that it feels like a big deal to some people if some models give wrong answers to questions like this one, "how many rs are in strawberry" (yes: I know models get this right, now, but it was a good example at the time), or "are we in the year 2026?"

[jrowen]

In my experience the tools feel like they make sense when I use them properly, or at least I have a hard time relating the failure modes to this walk/drive thing with bizarre adversarial input. It just feels a little bit like garbage in, garbage out.

[rkomorn]

Okay, but when you're asking a model to do things like summarizing documents, analyzing data, or reading docs and producing code, etc, you don't necessarily have a lot of control over the quality of the input.

[viking123]

Yes, my brain is just like an LLM.

[Flipflip79]

….sorry what?!