NOTES ·
I Looked at a Window
That night, I was lying in bed looking at the window. The room was not brightly lit, and outside it was already dark. Only a few scattered lights from distant buildings fell upon the glass. I looked at it for a while, and suddenly a very simple question came to me: why did I know that what I was looking at was glass? Could it have been made of some other material? Or could it even be possible that it did not exist at all?
The thought came for no particular reason. Of course I knew it was a window, and I knew there was glass in it. It hardly seemed worth thinking about. But when I actually looked at it, I realized that my eyes had never really told me that it was glass. What I saw was only the light outside passing through something transparent, while at certain angles the shadows of the room appeared on its surface. As for “glass,” that was only the conclusion I had given to all of this.
I reached out and touched it. It was a little cold. That seemed to make it more likely to be glass, but coldness itself is not glass. Metal is cold in winter too, and some plastics can feel cold as well. I tapped it lightly with my finger. The sound was clear and sharp, and that gave me a little more confidence, but sound could not prove anything either. When I changed the angle of my head, the view outside faded slightly and my own face appeared on the surface, faint and blurred.
I suddenly found it a little strange.
Perhaps human beings have always lived like this. We never receive the world itself directly. We only keep receiving the traces it leaves behind: a ray of light, a sound, a temperature, a touch, and memories from the past. Then we put these scattered pieces together and naturally say: this is a piece of glass, this is a table, this person is smiling, someone is walking over there. We take the physical phenomena that our eyes see and assume that they are simply things that ought to exist as we perceive them, but we rarely stop to ask whether what we see is actually real.
Perhaps all of this happens too quickly for us ever to notice.
Then I thought of a cup.
Suppose an agent is looking at a table. A cup suddenly falls from its edge. In the next moment, the camera sees fragments scattered across the floor, while the microphone hears a sharp cracking sound. If the agent has received enough training, it will probably arrive at an answer very quickly: the glass cup fell to the floor and shattered.
There is, of course, nothing unreasonable about that answer.
But then I wondered: does it really know what just happened?
Perhaps the cup was not made of glass at all, but ceramic. Perhaps there had already been a tiny crack in it before it fell, only no one had noticed. Perhaps the fragments had already been lying on the floor and had simply been hidden behind the leg of the table. The sound might even have come from somewhere else in the room. To take it a little further, perhaps someone was standing outside the camera’s field of view and shattered another object at exactly the same moment.
Human beings are actually the same.
We see dark clouds and then it rains, so we know that there is some relationship between dark clouds and rain. We see someone frown, and then hear their voice become colder, so we guess that they are angry. Most of the time, these guesses are correct, and so gradually we forget that they were guesses in the first place.
But how many people in the world are like Newton? An apple falls from a tree and hits him, and instead of simply accepting that it fell, he asks why the apple fell toward the ground.
Has anyone ever thought about the fact that our senses can deceive us?
Our eyes can.
Our ears can.
Memory can even more easily.
Even memory itself may not be what we imagine it to be—not something that happens once and is then quietly stored away. Researchers at Northwestern University found that when we remember something again, the brain does not simply retrieve the original memory and replay it unchanged. The act of remembering can make that memory susceptible to change again, and distortions introduced during retrieval may become part of what is remembered later.1 The next time we return to the same event, what comes back to us may no longer be exactly what happened at the beginning, but what our previous self remembered it to be.
Seen this way, even the past is not as stable as we imagine.
If our five senses cannot deliver the world to us without loss, and if even the memories we use to understand the world can quietly change as new experiences arrive, then giving artificial intelligence more cameras, more microphones, and more sensors will not automatically give it access to the truth of the world. It will simply receive more observations.
More observations are useful, of course.
But observation is still not state.
The more I think about it, the more I feel that the difference between these two things may be more important than it first appears. An image is only an image. A sharp sound is only a sound. If one object disappears at one moment and another thing appears in the next, that alone cannot immediately become a causal relationship. The true state of the world lies somewhere behind these phenomena, and an observer can only move toward it little by little.
If that is the case, then an artificial intelligence that truly understands the world perhaps should not immediately output, after seeing the cup hit the floor:
The glass cup has shattered.
Perhaps it should preserve something else.
The cup may have shattered.
Very likely.
But not absolutely.
If someone later picks the cup up from the floor completely intact, then the previous explanation should be overturned. If another camera reveals that what actually shattered was a plate beside it, then the meaning of that earlier sound should change as well. If the machine had hands, perhaps it could even pick the cup up, look at it, touch it, and tap it.
A new action would bring a new observation.
And a new observation would, in turn, change the world that it previously believed in.
I suddenly felt that there was something very strange and fascinating in this.
Perhaps true intelligence is not simply knowing a great deal. Perhaps it is knowing that what one knows can still change. It can believe something without completely killing off every other possibility. It can make a judgment while still leaving a small space for future facts to enter and overturn what it currently believes.
This is not a delay in performance, nor is it hesitation.
On the contrary, perhaps this is also how human beings manage to live.
Every day, we make countless judgments about the world, yet none of these judgments are taken directly from the world itself. We simply see some things, hear some things, and then maintain within ourselves a certain picture of what the world is like. When something new happens, that picture quietly changes a little.
Sometimes the change is very small.
Sometimes the entire world becomes different because of it.
I looked at the window again.
It was still standing there quietly, reflecting the faint light of the room. I still believed that it was glass. Those thoughts had not made me truly doubt it. And yet that belief somehow felt different from the one I had held only a few minutes earlier.
I knew that what I was seeing was not glass itself.
It was only a collection of things sufficient to make me believe that it was glass.
Then, if an artificial intelligence can likewise never directly see the true state of the world, and can only rely on light, sound, touch, time, and observation after observation—none of them completely reliable—to guess what is really happening in front of it,
what, then, should it truly preserve inside itself?