I forget where I got this, so I can’t give credit. If you wrote it, tell me and I’ll put your name and contact information right here. I kept this item because it was so perfect. Consider it a repost until further notice. Everyone anywhere near AI needs to see this.

This basic question stumps the best GPT in the world:

“I want to wash my car and the car wash is only 200 feet away. Should I start my car and drive there or just walk.”

Ask Anthropic‘s Claude Opus 4.6 and the answer is wrong.

Claude Opus 4.6 – “At just 200 feet, walking is the easy call. By the time you start the car, buckle up, and pull out, you could already be there on foot. It’ll take about a minute to walk. You’ll save gas, skip the hassle of parking twice, and it’s better for your engine — short drives like that don’t even let it warm up properly, which can actually cause extra wear over time.
Just walk it over!”

Me – “If I walk there, then I don’t have the car with me. And I want to wash the car.”

Claude – “Ha, you’re absolutely right — can’t wash the car without the car! Drive it over. Sometimes the obvious answer is the obvious answer.”

OpenAI‘s ChatGPT said to drive. But added a lot about cold starts of gasoline powered cars.

Google‘s Gemini also said to drive. And it also worried about cold starts of a gas engine.

So what’s going on here?

There’s a great new paper by Peiyang Song, Pengrui Han(Barry), and Noah Goodman entitled “Large Language Model Reasoning Failures” and described in detail all the ways that LLMs fail.

LLMs fail in predictable, systematic ways: lacking working memory, inhibitory control, and cognitive flexibility; replicating human cognitive biases; struggling with Theory of Mind and moral consistency; suffering the reversal curse and compositional breakdowns; failing at basic counting and arithmetic; and misunderstanding physical properties, spatial relationships, and embodied action across text, vision, and robotics.

These failures show that LLMs do not operate like humans.

Humans know when to switch from pattern matching, heuristics, and extrapolation to more structured rigorous tools such as logic, math, and using tools like measurement devices, or the result of scientific study to draw conclusions. LLMs don’t have all these pieces.

The human ability to switch models and cognitive tools is called meta-cognition. A meta-cognition engine would know when to call an LLM vs a calculation vs something like RootCause.ai‘s Causal AI engine. LLMs don’t understand cause and effect. RootCause does.

Right now, the big model builders are not creating meta-cognition engines. Instead, they have one hammer and everything looks like a nail. They are all trying to create one massive model that will solve everything. This “think longer and hope” approach (o1, R1, etc.) is essentially trying to simulate meta-cognition within the same next-token prediction machinery.

This is the problem Yann LeCun is trying to address. At least I think so. In his 2022 position paper “A Path Towards Autonomous Machine Intelligence” he talks about a configurator, which I believe is going to end up being a meta-cognition engine.

That kind of engine will enable AGI. And, answer to car wash questions with ease.