Toss a set of keys to a friend and the hand comes up to catch them. Nobody measures the distance first or works out the speed, yet the hand knows where to go. People have been doing this since they learned to walk: working out where a thing is, whether it will collide, whether an arm can reach it. That untaught instinct is exactly the part today's most fluent models handle worst.
Change the scene and much the same happens. Ask a phone how to arrange a sofa so the living room feels roomier and a model will produce advice that sounds plausible, something about the window, the center of the room, keeping a walking lane, even a design principle or two. Push one question further, asking how far the first chair on the left is from the TV cabinet, and it will mostly start making things up.
According to a media report, she once used an example to make the point: asking a model to count the chairs in a video of an office is still not easy. The example has been cited repeatedly. This article treats it only as a lead, uses no quotation marks, supplies no scene from an interview, and does not present the original wording as verified when its source could not be pinned down.
When a Key Comes Flying
Pouring a cup of coffee, slipping sideways through a crowd, picking a key off the corner of a table take almost no thought. Behind those actions sits a whole set of estimates about distance, orientation, occlusion and collision. A machine can write a polished article and still have no living room in its head that could be measured with a tape.
This is a different matter from speaking well. Language models have read enough text to answer questions smoothly; ask how far the door is from the sofa or whether turning around will knock the coffee table, and they can mostly only keep inventing along the line of the words. Words that follow one another neatly do not mean the space has been laid out correctly.
Judging the gap between a bumper and a curb while reversing, pouring coffee half asleep without spilling it, all rely on the same ability. It is not written in textbooks and is hard to turn into a question one can memorize, yet everybody carries it, while a machine has to build it from scratch.
The Chairs It Cannot Count
The chair example is easy to grasp because it puts an abstract weakness onto a physical object. Counting does not come from reciting an encyclopedia; it comes from separating the things in a frame one by one, recognizing them, holding them in mind, and not counting the same chair twice from two angles. People learn this in minutes with eyes and a body.
As the camera sweeps several rooms, some chairs are half hidden by a table corner, some are only an outline in the distance, and some show up again from a changed angle. A person glances once and has the number; a model easily treats an occlusion as two objects, a repeat as a new one, or simply guesses the total wrong.
According to a media report, she described this gap as a shortfall outside language. This article does not dress that judgment up in finer words, nor present it as a verbatim line from any one interview. What can be confirmed is only this: the example is used again and again to show that answering questions and seeing the world are not the same thing.
First See, Then Step In
There is a clear line through her career. Fei-Fei Li is a professor at Stanford University and a co-founder of World Labs; earlier, she led the creation of a large labeled-image dataset called ImageNet, which gave machines enough samples to learn to recognize objects.
A labeled-image dataset is, plainly, a large pile of photographs each paired with a note telling a machine that this one holds a cat and that one a cup. With enough samples piled up, a model slowly learns to recognize things on its own. The idea sounds ordinary now, but at the time it was a key step that pushed a whole field forward.
ImageNet solved the problem of seeing clearly: given a picture, say whether it holds a cat or a cup. Seeing clearly is only the first step. After that comes knowing where things sit, how they sit against one another, what changes when one of them moves, and only then reaching out to pick something up. According to a media report, she once said that the meeting of two kinds of technology was the starting point for founding the company.
A World Beyond Words
The name she gives this latter ability is spatial intelligence, or Spatial Intelligence in English. Put simply, it means letting a machine, like a person, have a sense of the things in a three-dimensional world: how big, how far, how arranged, where one would go if pushed, and the ability to plan its own actions from that.
This is not as simple as adding one more subject to a language model's timetable. Text is a compressed, second-hand description, while the world still holds gravity, occlusion, friction and distance. A model raised only on words can use the two characters for living room smoothly and still have no room one could actually walk into.
For the same reason, making a model better at answering questions will not fill this gap. However hard the question, the answer still lands on paper; work in the real world, done wrong, bumps, spills or falls short of reach. To cross over, a model needs an updatable state of the world in hand, rather than a smoother string of words.
The World Model Hurdle
Follow spatial intelligence downward and another term appears: the world model, or World Model in English. According to a media report, she once wrote an essay calling spatial intelligence the next frontier of artificial intelligence; this article gives neither the title of that essay nor its publication date, and keeps only the judgment about direction.
The World Labs website describes what it wants to do quite plainly: build large world models that can perceive, generate and interact with the three-dimensional world. Perceiving is seeing clearly, generating is producing a scene that holds up, interacting is working out what happens next when given an action. Only with all three does one move from images to space.
This is not the same as generating a lively-looking stretch of video. Video passes as long as the frames hang together for a few seconds; when the camera turns and an object quietly changes shape, that is often let go. A world model has to survive going back for another look: from a changed angle, or after a while, that table still has to be where it was.
According to a media report, she also described a world model as a place that can be set up from a prompt, walked into, looked around in, where objects can be picked up and arrangements changed. It sounds like a game, but the point is that a machine must stay consistent about the same room, rather than redrawing a pretty picture each time.
Leaving Some Room
Two boundaries need stating. First, she has not said language models are finished; the point is only that stopping at the layer of language cannot deliver promises that touch the real world, such as robots or scientific discovery. Second, this article does not predict how many years it will take before robots enter the home, since no such timetable is confirmed in public information.
A further list of things this article leaves out: it attaches no label of a judgmental kind to her, does not retell her remarks about other companies or people, and does not invent a source for the chair remark. A profile risks most when it turns hearsay into a quote and inference into fact; better to lose a lively line than to add one with no root.
From getting a machine to see a picture clearly, to getting one to stand steady inside a room, the distance is not just more data. Where the hand goes when a key comes flying, that small thing everybody has, turns out to be the part of this research least open to a shortcut.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.