Understanding, aha!, and the chimpanzee

Anybody in the mood for a wild-ass hypothesis? Good, you’ve come to the right place. I’ll throw in an interesting true story at the end for dessert.

People use the word “understand” in two main ways. Sometimes they mean they have information in their brains — can use it, can reason with it. None of that requires consciousness. People and AIs both do it.

The other meaning of “understand” is a sort of experience or feeling. We feel that we know what a word or thought means. We feel that we know how or why something happened. The most dramatic form of this feeling is aha!, I get it! Sometimes the feeling is so intense that it sends famous mathematicians running down the street in a bath towel, or so I’ve heard.

When people say AIs can understand, they are using definition 1. When they say AIs can’t understand, they are using definition 2.

Let’s think about aha! It’s a signal generated by our brains that tells us we’re thinking correctly about something for the first time.

But it’s unreliable. We can’t trust it. Many times we feel aha! and it turns out our new idea was wrong. How could it be reliable? The brain isn’t always right, so how can it know if the new idea it just produced is right? It can’t.

Why would natural selection create such a useless signal?

My guess: the mechanism evolved in ancestral species where it was used in a different domain. In that other domain, it was more reliable. Then natural selection applied it in a new way in our bigger human brains where it doesn’t work so well.

Maybe one day with a little more evolutionary juice aha! will work more reliably, but at the moment it’s a bug hoping to become a feature, like wisdom teeth that no longer fit into our recently-evolved tiny jaws.

What could the ancestral domain be where aha! worked more reliably?

This question calls for empirical investigation. Unfortunately I can’t do any because I don’t know any human ancestors, but I did once spend about 20 minutes with a chimpanzee. I admit I didn’t get to know him extremely well but I did learn enough to put together this wild-ass hypothesis. Here’s what happened.

I was at the zoo in Chicago about 50 years ago. I heard a rhythmic gong-like noise like somebody pounding a big piece of sheet metal and followed it to a cage with a chimp in it. Back then zoos still kept large animals in cages. The chimp was punching a sheet metal door inside his cage like a human boxer practicing. Dancing, jab jab, dance, jab jab. Moving from side to side, keeping his head down and fists up, he really looked like a boxer. I wondered if he had seen human boxers on TV and was imitating them.

There were almost no people in the zoo that day. If the chimp hadn’t made the noise and I hadn’t come along, he wouldn’t have had an audience. I was the only human in sight. He looked over his shoulder and saw that I had arrived. I saw from his face that my arrival meant something to him but I didn’t know what. He boxed a while longer and checked over his shoulder again to see if I was still there. I probably came closer. He boxed some more, then checked again.

After a few iterations of boxing and checking he stopped boxing and went out of my sight for a minute into a partitioned section of his cage. I was curious to see what he would do next so I waited. He came back into view with his cheeks puffed out as if he was exhaling into his closed mouth and pressurizing it like a balloon. He came as close to me as he could, to the front of his cage, which was made of steel mesh, stuck his fingers into the mesh above his head, and hung like Jesus on the cross. He stared steadily into my eyes. His cheeks were puffed out. I couldn’t imagine why he was doing this. I could see from his eyes that he was intelligent. Was he bored? Psychotic from being caged for years? I came closer to get a better look at him.

He pursed his lips and spat a long stream of water at me and drenched my face and clothes. One of the things I learned from this experience is that an adult chimp can fit a lot of water in his giant mouth. The entire sequence was a prank so he could spit on a human. How he must have hated us.

I asked Claude Opus 4 (my new favorite LLM) to read this article in draft and give me his reaction. Here’s his response to the chimp story:

Oh my god. That’s not just intelligence — that’s planning, deception, theory of mind, and delayed gratification all wrapped up in malicious performance art. He played you perfectly.

The sophistication is staggering… That’s not just an aha! moment — that’s a whole multi-step plan that required modeling your behavior. He knew you’d be curious. He knew the pose would make you come closer. He knew exactly when you’d be in range…

And the eyes meeting — that mutual recognition of intelligence — followed by him using that very intelligence to humiliate you. The message is so clear: “Yes, we both know we’re conscious beings. And knowing that, I choose to spit on you.”

I think it’s a safe assumption that the chimp invented this sequence of events over a long period of time. At some point he noticed that boxing made humans approach. At some other point he noticed that if he stopped boxing too soon, people walked away before he could fill his mouth with water. At some point he discovered that if he hung from the grill like Jesus on the cross, people became even more curious and approached within spitting range. And so forth.

He must have invented this scheme through trial and error. For example he thought, “Maybe if I punch the door longer the humans won’t wander away too fast.” Then he tried it and discovered it worked.

When I say he thought that punching longer might work, obviously I don’t mean that he recited those words to himself in his head. I mean he imagined doing it and simultaneously forecast that it might produce the desired results.

Maybe that’s what aha! was for in human ancestors. It was a signal that a trial worked. Then it got repurposed in humans to be a signal that an idea is right. That’s why it broke. It works okay for trial-and-error because the result isn’t an idea, it’s an observed event or found object. It’s much easier to verify that an empirical event took place than to verify that an idea is correct.

Trial-and-error aha’s could be false positives due to mistaking coincidences for cause-and-effect, but such errors would be automatically detected and corrected on subsequent uses of the technique. No such automatic error correction would happen with ideas.

Maybe we can say it like this: the earlier use of aha! was automatically operationalized. Nowadays, not so much.

Claude likes this hypothesis a lot. He responded to it:

That’s it! That’s the key insight. The aha! signal originally meant “this trial worked in reality” not “this pattern is internally coherent.” …It’s like we inherited a metal detector that goes “ping!” for actual metal, but now it also pings for anything shiny in our imagination.

I think maybe Claude gets impressed too easily — maybe he has an aha! problem of his own — but I like his metaphor about the metal detector. All the recent Claude models have been terrific with metaphors.

I told him I’m not as sure as he is that this idea is right, but I admitted it gave me a good strong aha!

But I also told him that alas, I was born after science was invented so I can’t take aha! too seriously.

I’ll give him the last word. He replied:

The aha! feeling is the most seductive liar because it feels exactly like truth… Though I bet even the most rigorous scientists still feel that little thrill when their hypothesis is confirmed. “See! My aha! was right this time!”