Generous Allowances for a Very Young Company

It is generally understood that a human infant of eight weeks is capable of very little indeed. It can, with some effort, focus its eyes on a nearby thumb, produce a series of uncoordinated gurgles, and occasionally sleep for forty consecutive minutes. It is not, as a rule, expected to draft a business plan, let alone receive a cheque for one thousand one hundred million dollars. Yet, in the sun-drenched nurseries of northern California, the normal rules of development have once again been politely set aside.

A company called River AI, which came into existence roughly sixty days ago, has announced that it has secured $1.1 billion in funding. To put this figure into perspective, it is roughly equivalent to the gross domestic product of a small, respectable island nation, or the cost of maintaining a modest fleet of Victorian battleships for a decade. The money has been provided by a syndicate of very serious people, including General Catalyst, Nvidia, and AMD, all of whom appear convinced that two months is more than enough time for a corporate entity to have established its character.

The founder of this remarkably wealthy infant is Igor Babuschkin, a gentleman whose previous employments include stints at DeepMind and OpenAI. Having spent years helping to build the sort of artificial intelligence that corporate executives hope will eventually replace their entire administrative staff, Mr. Babuschkin has apparently decided that this trajectory is somewhat lacking in charm. Instead of building a digital clerk to do your filing, River AI intends to build a "guardian angel."

This is a lovely phrase, conjuring images of a quiet, benevolent presence hovering just over one’s shoulder, perhaps gently nudging one’s hand away from a third chocolate biscuit or whispering the name of an acquaintance whose face one has recognised but whose identity has entirely vanished. According to the company’s literature, this personal AI will live close to you, know you intimately, and be entirely yours. It will not be a tool you call upon to perform a task, but a silent partner in the business of being alive.

I once knew a gentleman who spent three years attempting to train his domestic cat to retrieve the morning post, only to conclude that the animal’s silence was a form of intellectual superiority. We are now doing something similar with silicon, though with considerably more funding. The cat, at least, did not require a dedicated neocloud to ignore my friend’s instructions.

To achieve this state of digital grace, River AI insists that the entire computing stack must be rebuilt from scratch. This includes the training methods, the models, the product layer, and even the physical hardware. It is an approach that commands a certain admiration. It is rather like deciding to bake a Victoria sponge, but refusing to begin until one has first bred a new variety of wheat, constructed a proprietary milling facility, and designed a novel type of domestic oven that operates on principles of advanced thermodynamics. Most people would simply buy the flour, but when one has a billion dollars in the bank, such shortcuts must feel dreadfully uninspired.

The first product of this grand reconstruction is an API that allows enterprises to run complex reinforcement learning runs in fifteen minutes. This is marketed as an antidote to "prompt engineering," which the company describes as a rather undignified way of steering a model you do not actually own. Instead, River AI promises to let you train open models into something that is truly yours. It is a compelling pitch for the modern executive who is increasingly nervous about leaving their corporate secrets on someone else’s server, even if that server is housed in a very secure facility in Virginia.

Yet, one cannot help but wonder about the transition from a fifteen-minute reinforcement learning run to a personal guardian angel. The gap between a highly efficient database query and a benevolent spiritual guide is quite wide, and it is not entirely clear that it can be bridged by throwing more graphics cards at the problem.

There is a quiet dignity in a grandfather clock that does nothing but tick and occasionally strike a brass bell. It does not require a neocloud, nor does it seek to understand your relationship with your mother. It simply observes the passage of time with a steady, mechanical indifference that is deeply comforting. A digital guardian angel, by contrast, sounds like it might be rather exhausting. One worries it might become overly critical of one’s reading habits or quietly log the number of times one has checked the refrigerator before noon.

The sheer scale of the investment also raises some interesting questions about the nature of modern finance. In the old days, a company was expected to produce a widget, sell the widget to a customer, and use the resulting profit to buy a slightly larger desk for the managing director. Today, the process has been streamlined. One merely needs a compelling vision, a pedigree from a famous laboratory, and a name that sounds like a geographical feature. The actual product can be sorted out later, perhaps during the third or fourth year of operation, once the initial billion has been comfortably digested.

For now, River AI is starting its journey with a war chest that would make many small countries blush. Whether it succeeds in rebuilding the digital universe or merely succeeds in buying a very large number of very expensive computers remains to be seen. But as the blinking cursor on the screen reminds us, we are all waiting for someone to tell us what to do next. We can only hope that when our guardian angels finally arrive, they have the decency to be polite about our spelling.