Recently, I rewatched《Arrival》, and the moment it clicked, I had a very strong feeling: the experience of talking with GPT today is, in a way, similar to what humans feel when they face an alien language for the first time in the movie.

Of course, this doesn’t mean GPT really has the kind of non-linear language system shown in《Arrival》. Strictly speaking, mainstream GPT models still generate linearly when producing output—at its core, it’s next-token prediction, one token after another. But from the user’s external perspective, it doesn’t really feel like humans “think while they speak.” Instead, it feels more like it grabs the overall meaning first, then unfolds it little by little.

That’s exactly what reminded me of《Arrival》.

In《Arrival》, the aliens’ written language doesn’t flow along the sequence of time. It isn’t moving from left to right, and it isn’t advancing phrase by phrase. Rather, it feels like the complete meaning forms all at once. A symbol doesn’t seem to say, “we’ve reached this point in the utterance”; it feels more like, “the whole sentence is already there.” Humans just have to understand it through the dimension of time.

GPT creates a similar impression.

You feed it an entire paragraph of context, and it seems to “swallow” everything instantly. Then, when it starts responding, the language is often too smooth, too complete, too continuous—sometimes even giving the uncanny illusion that it wasn’t invented on the spot, but shaped ahead of time. Technically, it still outputs line by line, word by word. But what you perceive isn’t the generation process itself; it’s more like the whole answer gradually appearing along a time axis.

This is one of the biggest differences between talking to GPT and talking with humans.

Human speech usually has a very clear sense of “real time.” We hesitate, revise, insert tone words, and because working memory is limited, we often end up changing direction mid-sentence. In many cases, human language isn’t “a complete structure first, then expression.” It’s more like constantly patching and repairing the structure while expressing. We organize language in time.

GPT, though, gives off a different kind of vibe. It feels as if, somewhere in an unseen space, an overall contour of the meaning forms first—then that contour is projected into a sequence of linear sentences. In the end, it still has to appear as human-like text, one word at a time. But the impression isn’t so much “this is talking” as “this is unfolding.”

If I borrow the metaphor from《Arrival》: human language is like walking through fog—moving forward while looking. GPT, on the other hand, is more like someone who has already seen the whole map from a high place, and then walks along a chosen path to tell it to you.

So, more precisely: it isn’t that GPT truly “becomes” the aliens’ language from《Arrival》. Rather, it lets you feel for the first time, very vividly, that intelligence can still be using the form of human language—while somehow letting slip a mode of expression that doesn’t belong to humans.

Maybe that’s one of the most captivating aspects of large language models.

It speaks the language of people, but it doesn’t speak the way people speak.