The sentence they left out

A frontier AI lab built the machine Turing described in 1950. They put his words at the top of the paper. They stopped four lines short

Wired ran a piece yesterday with the headline “AI Isn’t Smarter Than a Baby, Yet.” It is a good piece. The argument is that babies learn from almost nothing, that AI models need an ocean of data and a small country’s worth of electricity, and that the difference might be where the next real advance comes from.

That is all true. But the thing that stopped me is not in the Wired piece. It is in the paper the Wired piece links to, on page one, above the title of the first section.

What they built

The paper is called EgoBabyVLM. It came out of Meta, Stanford, the École Normale Supérieure and the University of Tokyo in May, with twenty-two authors on it. What they did is roughly this: strap cameras to the heads of infants and toddlers, record about eight hundred and sixty hours of what those children see and hear, and then train an AI on that footage and nothing else. No internet, no curated text. Just a child’s experience, and the question of whether that is enough.

It is careful work by serious people, and I think it is the most interesting thing happening in machine learning right now.

At the top of the paper, before anything else, they put an epigraph. It reads:

Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child’s? If this were then subjected to an appropriate course of education one would obtain the adult brain.

Alan Turing, 1950

I have been living inside that sentence for about a year, so seeing it on the front of a Meta paper was a strange morning.

The part people tend to skip

Turing’s 1950 paper is the one with the Turing test in it. Almost everyone reads the first six sections, which is where the famous stuff lives. The seventh section is where he stops asking how to test a thinking machine and starts asking how to build one, and it is mostly treated as engineering speculation. Interesting if you like the history. Skippable if you are doing philosophy.

The epigraph above is from that seventh section. So the Meta paper is not decorating itself with a famous name. They are doing exactly what the passage says. This is Turing’s proposal, executed, seventy-six years later, with head-mounted cameras.

Which is why what happens four lines further down matters.

Turing writes:

The example of Miss Helen Keller shows that education can take place provided that communication in both directions between teacher and pupil can take place by some means or other.

Sit with why he picks her. In the whole of that section, in the middle of an engineering discussion about building a machine, Turing names exactly one human being. Out of the entire history of teaching anyone anything, he chooses Helen Keller. Deaf and blind from nineteen months old. Every ordinary way into a child’s mind, looking at things, hearing people talk, closed. And she was educated anyway, brilliantly, by Anne Sullivan.

If you want to prove that the channel does not matter, that sight and sound are not the essential ingredient, Keller is the strongest case there is. Everything was taken away except one thing, and she was still formed. What was left was the relationship. A teacher and a pupil, going back and forth.

Turing says this and moves straight on. He does not appear to notice that he has just said something about what intelligence requires.

The Helen Keller sentence is not in the EgoBabyVLM paper.

What happened when they ran it

Here is the part that caught my attention the most.

Think about what a recording actually is. Everything a child took in through her eyes and ears is on that tape, timestamped, in order, at the moment it reached her. Nobody has ever held data like this before, and I do not want to undersell how remarkable it is that it exists.

What a recording cannot do is answer you. The parent’s face is in the video. The parent’s voice is in the audio. But none of it is listening. The tape will not notice the model is confused. It will not slow down, try again, or point at something twice because the first time did not take.

So the setup, without anyone intending it, is this: content kept, relationship removed. Two different teams have now run it on the same corpus of children’s footage, and the results come apart along a line I have not been able to look away from.

The physical world comes through. A group at Stanford, publishing separately in April, trained a different kind of model on this data and got depth, motion, object segmentation and basic physical intuition out of it. Not roughly. Well enough to compete with systems trained on the entire internet, from a hundred and thirty-two hours of one single child. Their model’s internal workings even line up with human brain scans. That is an astonishing result and it deserves more attention than it is getting.

Language and everything social does not come through. That is the EgoBabyVLM half. Grammar sits at roughly chance. No theory of mind. No sense of what people are doing or why.

Same footage. Opposite outcomes, split by what was being learned.

And there is one number I cannot stop thinking about. The researchers measured how tightly a child’s view lines up with the words spoken over it, moment by moment. On the curated picture-and-caption data that trains ordinary AI systems, that score comes out at 0.916. On the real thing, a real child’s real day, it is 0.012. Effectively zero. Nothing in a toddler’s life is captioned. People talk about what happened yesterday, or what is behind you, or nothing much at all. The models cannot get off the ground without the tidy number. Every one of us managed it on the other one.

What I think, and what I don’t

Their reading is that this is an engineering problem. The signal is in there, weakly, and we do not yet have architectures that can pull it out. They may well be right, and I want to be honest that I cannot rule it out from where I am sitting.

There is a second thing I cannot rule out, and it cuts against me, so I should say it plainly. The two halves I just put next to each other come from two different teams using two different kinds of model. So the split might be about the architectures rather than about what was being learned. I do not think that is the whole story, because the failures cluster so tidily at the social end, but I notice that I want it not to be the whole story, which is exactly when a person should slow down.

Still, there is another reading available, and it is Turing’s own. What the recording removed was not signal. It was the second direction. And you cannot have a two-way relationship with a tape, no matter how good your architecture is.

The thing that makes me suspect this is not just my own hobbyhorse is what they propose to fix it. Near the end of the paper they list what they think is needed, and the list includes gaze, pointing, and joint attention. Joint attention is two people attending to the same thing and each knowing the other is doing it. It is a two-way structure. They are trying to reconstruct the second direction from a recording of it, which is a bit like trying to have a conversation with a photograph of someone talking.

Anne Sullivan answered back. The tape does not.

I do not know whether this is the explanation for their results. I am a philosopher reading a machine learning paper, and I am aware of what that is worth. What I am fairly confident about is narrower: the sentence that would tell you where to look is in the same section of the same paper they took their epigraph from, four lines down, and nobody has picked it up.

It is still there. It has been there since 1950.

I think this has consequences for how we evaluate these systems, and for who is left able to evaluate them, and that part is a longer argument than a post can carry. I am presenting it at a conference in Vilnius in September, and there is a paper coming. For now I just wanted to put the epigraph and the missing sentence next to each other, because as far as I can tell nobody has, and it seems worth someone doing.

July 2026