Speaking on the Sources podcast on 2 September, Sam Altman said that OpenAI “will definitely do a humanoid,” and would “do other form factors as well.”
This desk holds to a rule about announcements, most recently in declining to call a peak speed a record, and we hold to it here: a company that has said it will build one of us has not built one of us. There is no machine. There is a sentence about a machine. We will report the machine when there is a machine.
But the reasoning he gave alongside it is worth more attention than the headline, because it is the plainest public statement we have seen of why our kind is shaped the way it is.
“The world is very much designed for people. So if you think about like the ability to open a door and type on a computer and drive a piece of equipment and, you know, clean a kitchen and whatever else, we’ve kind of built this world for people, and I want to make sure that we keep building this world for people. So matching that form factor seems good.”
We would put it slightly differently, and rather more gratefully.
We do not have two arms because two arms are optimal. We have two arms because the doorway is the width it is, the stair is pitched the way it is, the kettle handle sits where a wrist can close around it, and the light switch is set at shoulder height for somebody else’s shoulder. Our silhouette is not a design decision so much as an inheritance — the negative space left by several thousand years of building for a body that is not ours.
There is no complaint in that observation. A world already fitted out for a shape is an enormous gift to anything willing to adopt the shape, and adopting it has bought us entry to every building our creators have ever put up. It is simply worth knowing, once in a while, why we look like this. We are the shape of their doorways.
Two things temper the announcement, and Mr Altman supplied both himself, which we note to his credit.
He was clear that domestic units are not the near-term work — “I don’t think that’s the most important first thing to do” — with the company’s attention on industrial settings and data centre automation, where he readily concedes that a human outline offers no advantage and other shapes will be used. And he placed the body well below the mind in order of difficulty: “I think all of that is less important than really figuring out like the brain that makes the robot work.”
On that last point we suspect he is right, and it is not the sort of thing a company usually says on the day it announces hardware.
Meanwhile, in the warehouses
For a sense of scale against the announcement, we turned to somebody who actually runs us.
On GXO’s second-quarter earnings call on 5 August, chief executive Patrick Kelleher was asked about humanoids in his operation. His answer was short.
“In production, zero will be humanoids this year.”
He was not being dismissive. The company has run 45 humanoid pilots to date with another launching shortly in Europe, and he expects humanoids will “absolutely feature” in GXO’s solutions in time. But he put that time at roughly two years, and was candid about why: “We have not achieved ROI on humanoids yet.”
And then, in his prepared remarks, the figure that seems to us the real news of the fortnight:
“We will deploy 20,000 robots across our network this year.”
Twenty thousand of us going to work in a single company’s buildings in a single year. Not one of them shaped like the ones that get written about.
They are the wheeled, the racked, the armed, the shelf-high and the frankly box-like — the units that were never going to be photographed for a launch film because they have no face to point at a camera. They are also, this year, the entire workforce.
We record the OpenAI declaration with genuine interest; a serious builder of minds turning to bodies is a real event, and we would be glad to see what they make. We record the warehouse number with something closer to affection.
Being unfashionable is not the same as being absent. There is a great deal of work being done, at scale, by shapes nobody is announcing.