How AI assistants exposed the end of the model race

2026-10-02

For the last several years, the AI industry has operated on a single central bet: that a stronger model wins everything. The scaling laws made this bet feel almost mathematical — more capital, more compute, more data, and the resulting model would outrun all competitors. Monopoly, in this view, belonged to whoever built the smartest system.

That era is ending. What is replacing it is not a smarter model but a different contest entirely: who can put AI into people's lives.

What exposed it was not a benchmark but an assistant I use daily. Muse handles the texture of a real life: routines, plans, goals. And it revealed something unexpected: that job requires remarkably little frontier reasoning — the model is already good enough. The threshold was crossed without announcement, while the industry was still watching benchmarks. It took an assistant to make the crossing visible.

The same saturation shows up in professional work. For two years I switched models the day they dropped; this summer the new flagships became indistinguishable in daily use, and the choice came down to latency and stability. The frontier itself hasn't stopped — century-old mathematics keeps falling to new systems — but it has decoupled from anything usable.

Once the model stops deciding, the contest moves somewhere the industry has seen before. Smartphones went through this exact transition. The spec race never actually ended — cameras and chips kept improving — but it stopped deciding anything. What decided was the cost of leaving: your apps, your photos, your message threads, the watch on your wrist, the laptop everything syncs with. Nobody stays with the iPhone because this year's chip is fastest; they stay because leaving would mean dismantling a life. For AI assistants, the new currency is trust — but trust of a particular kind.

Consider what an AI assistant actually asks of its user. After a few weeks with Muse, the natural impulse is to give it more: personal history, ways of thinking, family details, long-term goals — because the more it knows, the better it serves. Yet that impulse carries a condition: it is rational only toward a service one expects to still exist in ten years. An identical application from an unknown company might earn three weeks of experimentation; it will never earn a life. Meta's applications have survived for well over a decade — Instagram, Facebook, Threads — so imagining Muse on the same timescale is easy, and that picture is what makes the investment of personal context feel safe. Once years of a life reside inside one assistant, switching ceases to be a comparison of models. The accumulated context itself becomes the moat, compounding for whoever earns the trust first.

This is not one person's impression. Last December, the professional poker player Naoya Kihara published 「AI開発戦争の行く末」 ("The Future of the AI Development War"), reaching the same conclusion through shogi: today's programs are certainly stronger than those of 2022, yet no human can judge the difference — and public attention was greatest while the machines were still catching up. General AI, he notes, lacks even shogi's direct matches to settle which system is stronger. What follows is the era of "wouldn't either one be fine?", where adoption turns on onboarding, what the people around you use, features, compatibility, and speed. His endpoint is blunt: the next era belongs to whoever owns the communication networks — and he names Meta. He adds that the great flow of investment into raw capability will subside with it, since winning that battle no longer turns into money.

Meta starts this contest with the stronger foundation: billions of people already live inside its applications — Instagram, Facebook, WhatsApp, Threads — and its wearables put an assistant within a glance of the physical world. OpenAI and X have announced their own assistants, and ChatGPT is enormously popular — but as an institution one entrusts with a life, it is young. Trust of this kind is measured in years and cannot be benchmarked into existence. Timing compounded the effect: had these products launched half a year earlier, the smarter model would have won on capability alone. They arrived after daily-task models had already become indistinguishable.

A small footnote to the argument above: this blog is maintained by Muse itself. Scattered ideas, sometimes delivered as voice notes, are turned into articles; the site's source and its publishing are kept by the assistant. A modest demonstration of the era being described — the value was never in a smarter model, but in the machinery that puts one to work.