Runway has introduced Solaris, a research system that generates a program's interface in real time instead of running one. A language model decides how the screen should change; a world model draws each new frame. Underneath there is no application — no layout, no components, no code. Solaris is built on Gen-4.5, Runway's video model, and continues the direction the company opened with GWM-1. It is not a product. Runway is looking for launch partners and handing out early access through a form.
The company lists three places it thinks this fits. In online commerce, the store becomes the interface itself: a shopper drags a garment onto their own photo to try it on, and the storefront restyles itself for each person. In product visualization, saying "move the table" or "change the color of the sofa" makes the scene respond immediately, which Runway ties to furniture sales and interior planning. In learning, instead of the text answer a chatbot would return, Solaris shows the next step in context; the example Runway gives is visually demonstrating how different materials burn.
Source: the-decoder.com
The split between the two models is the design. Intent goes to the language model, pixels to the world model, and nothing in between is compiled or laid out in advance. Runway also frames Solaris as a training environment for AI agents. Agents today fail at tasks like booking a hotel, the company says, because they are trained on fixed layouts and break on sites built differently. A system whose output is a different interface every time produces that variation by construction, which is at least a coherent answer to the problem.
Source: the-decoder.com
The larger claim is that the app, as a fixed unit, eventually goes away. Today every action means opening a separate finished program — one for shopping, another for messages, a third for bookings. If the operating system can assemble a fitting interface on demand, the need to keep programs filed in a permanent catalog shrinks. An online store in that model stops being the same page for everybody and becomes a generated environment that reshapes itself per user. Runway leaves open whether people actually want interfaces that keep changing.
Source: the-decoder.com
That open question is not a footnote, and I would put it harder than Runway does: an interface that is different every time is one you cannot learn, and learnability is most of what makes software usable. Muscle memory, the position of the button, the fact that the checkout flow was the same yesterday — those are features, not limitations waiting to be generated away. A storefront that restyles itself per person is also a storefront no two customers can talk about, and no regulator can screenshot.
Runway's own list of limits is unusually blunt for this kind of announcement. Stable, legible text is still hard, and interfaces are mostly text. Long sessions and reliability remain open. The sharpest line in the whole thing is the company's admission that a convincing but wrong result can be worse than no result at all — which, in a system that renders a checkout page rather than executing one, is the entire risk surface in a single sentence. Runway also notes the generated interface has to work with assistive technology, including screen readers. That one is harder than it sounds in a system whose output is frames rather than structure: a screen reader needs something to read, and a video model does not emit it as a byproduct.
So Runway is asking for two futures at once. In one, the app dissolves into an interface the operating system generates fresh for every person. In the other, Solaris trains the agents whose job is navigating apps built by somebody else. Whichever of those arrives first makes less sense of the other.