Putting a Price Tag on Common Sense

There is something deeply comforting about the fact that it costs eleven billion dollars to teach a machine how to pick up a sock. In a world where we are constantly told that artificial intelligence is on the verge of transcending human limitations, it is heartening to know that the simple act of navigating a kitchen without causing a minor domestic catastrophe remains a multi-billion-dollar engineering challenge. Physical Intelligence, a startup with a name that sounds like a particularly earnest gym franchise, is reportedly in talks to raise a billion dollars at a valuation that would make most small nations blush. This would bring their total worth to over eleven billion dollars, a figure that suggests they aren't just building robots; they are building robots that can presumably find the TV remote on the first try.

The goal, we are told, is 'general-purpose' AI for the physical world. For those of us who struggle with the specific-purpose task of remembering where we put our spectacles, the idea of a general-purpose entity is quite appealing. However, the 'physical world' is a notoriously difficult place to do business. Unlike the digital realm, where an error results in a polite dialogue box, an error in the physical world usually results in a broken vase or a robot accidentally attempting to use the cat as a doorstop. It is a place of friction, gravity, and the inexplicable tendency of toast to land butter-side down. To master this, Physical Intelligence is attempting to create a 'foundation model' for robotics, which is essentially a very large digital brain that has spent a great deal of time watching videos of people doing things and thinking, 'Yes, I could probably do that, provided I had a billion dollars and a team of Stanford PhDs.'

One cannot help but admire the sheer velocity of the valuation. Doubling one's worth in four months is a feat usually reserved for rare tulip bulbs or particularly aggressive strains of sourdough starter. It suggests a level of investor enthusiasm that borders on the religious. There is a sense that if we just throw enough capital at the problem of 'stuff,' the stuff will eventually start behaving itself. We have spent the last decade teaching AI to write poetry and generate images of astronauts riding horses, only to realize that what we actually need is something that can empty the dishwasher without a philosophical crisis. It turns out that writing a sonnet is significantly easier than understanding the structural integrity of a damp paper towel.

I find myself reflecting on the nature of this 'intelligence.' We call it 'physical,' as if to distinguish it from the sort of intelligence that lives in a box and tells us the weather. But intelligence in the physical world is mostly a series of humiliations. It is the realization that the door opens 'pull' not 'push,' and the subsequent attempt to look like you were just testing the hinges. If Physical Intelligence succeeds, we will have robots that can navigate these humiliations with grace. They will move through our homes with the quiet confidence of a butler who knows exactly where the sherry is hidden. But for eleven billion dollars, one does wonder if it might be cheaper to simply buy more socks and accept that the ones under the sofa are lost to the ages.

There is a certain irony in the fact that we are using the most advanced technology ever conceived to replicate the basic motor skills of a toddler. We are building supercomputers to understand the concept of 'up' and 'down,' and spending the GDP of a small island to ensure a mechanical arm doesn't squeeze a tomato too hard. It is a testament to the complexity of the mundane. We take our bodies for granted until we try to build one out of aluminium and code, at which point we realize that being a human is actually quite a sophisticated operation, even on the days when we walk into a room and immediately forget why we are there.

The investors, of course, are not looking at the socks. They are looking at the 'foundation.' They believe that once you have a model that understands the physical world, you can put it into anything—a delivery drone, a factory arm, or perhaps a very expensive toaster that knows when you're feeling sad. It is the dream of the universal tool, the digital Swiss Army knife that can turn its hand to anything. But for now, the hand is still learning how to hold the knife. And as long as it costs a billion dollars to teach it, I suspect my dishwasher is safe from automation for a few more years.

It is a strange era we live in, where the most valuable thing a machine can do is act like a person who isn't particularly busy. We are subsidizing the mechanical equivalent of 'pottering about,' and we are doing so with a fervor that suggests we have run out of other things to buy. Perhaps, in the end, the robots will learn to value themselves as much as the venture capitalists do. One can only hope they don't start asking for a cut of the valuation; an eleven-billion-dollar robot would likely have very expensive tastes in lubricants.