Runway's new AI renders apps as real-time video with zero code

Imagine opening a website where not a single line of code is running under the hood. No HTML, no CSS, no React. Every button you see, every shadow that falls, and every layout shift is being dreamt up as live video, rendered frame by frame the exact millisecond you click.

That is the premise behind Solaris, a new system Runway is calling the first "Interface World Model." The company just rolled it out into early access, betting that the future of computing might actually be one continuous hallucination.

Ditching code for pure video

Every interface we use today relies on an intermediate layer: developers take visual ideas and translate them into deterministic code. Runway argues that this translation is fundamentally lossy. It freezes software into predefined paths and forces designers to sacrifice rich visual details just so pages can load predictably.

Solaris tosses that intermediate code out the window. Built on Runway's Gen-4.5 video foundation model, it treats your mouse clicks and drags directly as conditioning prompts for the next video frame.

To make this feel like real software instead of waiting on a slow video render, Runway distilled the denoising pipeline into an autoregressive real-time engine. It runs within a half-second latency window at 720p, fast enough to mimic actual responsiveness.

A dual-engine setup powers the experience. A frontier language model handles the reasoning by interpreting what the user wants and choreographing the scene logic, while the world model takes care of the rendering. In Runway demos, users can drag a shirt off a rack directly onto an uploaded photo to try it on, or toss ingredients into a bowl to watch a salad assemble itself naturally.

Runway also claims this solves a major hurdle for AI computer-use agents. LLMs usually break when web layouts shift because they memorize rigid code trees. Solaris offers a sandbox of infinite, unpredictable, dynamically rendered environments for agents to train on.

In Runway's internal evaluations, testers picked Solaris interactions over pages coded by Claude Opus 5 in 71% of in-scene behavior matchups and 61% on instruction following. Still, the company admits serious friction remains, including fuzzy text legibility, long-session drift, and "convincing-but-wrong" interface hallucinations.

Why it matters

Runway is making a wild conceptual leap here: instead of teaching AI how to write code for human interfaces, why not just let AI render the interface directly out of thin air?

For everyday users and developers, this is not something that will replace your browser tomorrow. The early access build is essentially a tech preview. Serving real-time video frames is still vastly more expensive and computationally heavy than pushing static text files across a CDN.

Where this gets genuinely exciting is in dynamic commerce and agent training. If software no longer needs every single interaction scripted beforehand, dynamic mockups and adaptive shopping environments could become fluid and conversational in ways traditional app development cannot match.

We think the sheer ingenuity of bypassing code is brilliant, but the fundamental flaw of world models remains their lack of determinism. Software has relied on strict code for decades because clicking "pay now" cannot afford a hallucination. Until Runway proves Solaris can guarantee absolute precision alongside visual fidelity, it remains a mind-bending peek into the future rather than an imminent replacement for the web.