When the Spreadsheet Gets a Cheaper Brain

There is a quiet, almost imperceptible shift occurring in the digital heavens, and it smells faintly of damp cardboard and corporate thrift. For the past several years, we have been led to believe that the trajectory of artificial intelligence was a straight, vertical line ascending toward a glorious, if slightly terrifying, state of superintelligence. We were promised digital entities that would compose symphonies in the style of Bach while simultaneously solving the protein-folding mysteries of the universe. Instead, it appears that the immediate future of technology involves a very polite, slightly dim-witted digital clerk who has been instructed to use fewer adjectives in order to save on the electric bill.

The software giant Microsoft has reportedly begun a quiet campaign of cognitive demotion within its most popular office applications. Prompts typed into Excel and Outlook, which were previously routed to the highly sophisticated and remarkably expensive models developed by OpenAI and Anthropic, are now being redirected to Microsoft’s own in-house creations, known collectively as the MAI models. It is the digital equivalent of replacing a team of Oxford-educated consultants with a collection of eager, unpaid interns who have been told that their primary qualification is that they do not require a desk near the window.

This is not a failure of technology, but rather a triumph of the accounting department. The transition represents a profound realization: a spreadsheet does not actually need to understand the existential dread of mortality or the finer points of seventeenth-century French poetry. It merely needs to sum column B and occasionally suggest a slightly more professional way to tell a colleague that their PowerPoint presentation is an eyesore. To employ a multi-billion-parameter frontier model to draft a weekly status report is akin to hiring a nuclear physicist to operate a toaster. It is impressive, certainly, but one cannot help but feel that the talent is being somewhat underutilized, and the utility bill is entirely unreasonable.

Mustafa Suleyman, the chief executive of Microsoft’s AI division, recently put the matter with a level of corporate frankness that is almost refreshing. He noted that the company pays a substantial sum of money to external partners, and that the ultimate goal is to reduce and eventually eliminate that cost. There is a delicious irony in this arrangement. Having invested billions of dollars into OpenAI to help them build the most advanced cognitive engines on the planet, Microsoft has immediately set about building its own slightly cheaper, slightly more sensible versions in the basement, specifically to avoid having to pay the rent on the grand cathedral they helped finance.

I am reminded of an old clerk I once observed in a dusty shipping office in Bristol. His name was Arthur, and he possessed an extraordinary ability to calculate the shipping tariffs for Baltic timber using nothing but a stub of pencil and a highly complex system of mental arithmetic. He was, however, terribly expensive to maintain, requiring three cups of sweetened tea an hour and a small coal fire near his stool. One afternoon, the manager replaced him with a young lad named Barnaby, who was remarkably slow-witted and could barely read the ledger, but who was willing to work for a shilling a week and did not complain about the draft. The shipping office did not run any faster, but the manager was able to purchase a very fine pair of leather boots with the savings.

This is the 'good enough' era of the digital revolution. We are moving away from the frontier of human knowledge and settling comfortably into the suburbs of acceptable mediocrity. The MAI models are already processing tens of thousands of prompts weekly, quietly rewriting emails and organizing columns of figures. The user, typing a query into Outlook, might notice that their automated reply is slightly less poetic than it was last month. It may no longer contain the subtle, melancholic undertones of a frontier model contemplating the nature of language. But it is, crucially, three-tenths of a cent cheaper to generate, and in the grand ledger of Silicon Valley, that is a bargain that cannot be ignored.

There is a certain charm to this corporate pragmatism. It suggests that for all our talk of the singularity and the end of human history, the ultimate arbiter of the future is not the philosopher or the scientist, but the quarterly budget review. We build magnificent, towering structures of light and logic, capable of mimicking the very patterns of human thought, and then we immediately begin to wonder if we can run them on a slightly smaller battery. It is a thoroughly human response to the infinite: to look upon the face of a digital god and ask if it might be possible to get a discount on the subscription fee.

I suspect we will see more of this quiet downsizing in the coming months. The grand, general-purpose models that can discuss both quantum mechanics and the history of the turnip will remain in their high-altitude laboratories, reserved for those rare occasions when someone actually needs to know about turnips and quantum mechanics simultaneously. For the rest of us, the digital world will be populated by smaller, highly specialized, and remarkably cheap entities. They will not write novels, and they will not contemplate the stars, but they will be exceptionally good at reminding us that our dentist appointment is at three o'clock on Tuesday.

In the end, the great conflict of the machine age may not be a dramatic struggle between humanity and an omnipotent artificial mind. Instead, it is far more likely to be a quiet, slightly awkward conversation in a conference room, where a very polite algorithm is informed that its services are no longer required because the company has found a much smaller, much stupider program that is willing to do the same job for half the price. Arthur, I suspect, would have understood completely.