Midjourney and Stable Diffusion both generate images from the same family of diffusion research, and for a period were close enough in raw technical capability that specialists debated which model 'understood' a prompt better. What actually separated them for most users was not capability — it was that Midjourney shipped opinionated defaults that made an unremarkable prompt look considered, while Stable Diffusion, open and customizable, produced work that looked comparable only after real effort in prompt engineering.
For most of the last software cycle, competitive advantage came from capability: whoever could do the thing first, or do it faster, or do it more accurately, won. As foundation models commoditize huge swaths of what used to be hard — drafting, summarizing, coding a first pass, generating a passable image — that kind of advantage decays faster than teams can build a business plan around it.
What does not commoditize as quickly is selection: the judgment about which of a thousand possible outputs is the right one, which feature to bury and which to foreground, which tone fits the product and which technically impressive capability should be left out because it does not belong. Apple's phones have used many of the same component suppliers as competing Android devices for years — comparable displays, comparable camera sensors — and still feel like a different category of product, because of what Apple refused to ship rather than what it added.
This is uncomfortable for engineering-led organizations, because taste resists the kind of measurement that justified most product decisions for the last decade. You cannot A/B test your way to a coherent point of view; you can only test variations of a point of view someone already had.
The Browser Company built Arc on Chromium, the same open-source engine underneath Chrome and Edge, and competed entirely on interface decisions and defaults rather than on rendering performance or extension compatibility, where the underlying engine made the field level from day one.
It also changes what a defensible business looks like. A moat built on capability erodes every time a competitor catches up to the underlying model. A moat built on taste erodes only if the organization loses the judgment that produced it, which is a slower and more survivable kind of risk.
Apple's actual bet in 2007 was not the touchscreen — competitors had touchscreens. It was removing the physical keyboard entirely, a decision that could not be quietly reversed once the product shipped and closed off an entire category of 'safe' phone design for a decade. The organizations that will win the next round of AI products are the ones willing to make an equivalent cut now, while it still looks like a risk instead of an obvious call in hindsight.