--- title: "This AI Does Not Write Code. It Draws the Interface, Frame by Frame" date: 2026-09-06 category: AI in Action site: NeuroAI canonical: https://neuroai.site/a/na-ai-draws-interfaces-not-code language: en --- # This AI Does Not Write Code. It Draws the Interface, Frame by Frame > Video generation company Runway introduced Solaris, which predicts and renders the next screen directly instead of generating code first. In testing with 250 evaluators, 61% preferred its interfaces on instruction-following and 71% on natural feel. ## Key takeaways - **Solaris**, from video generation company Runway, **predicts and renders the next screen directly** rather than writing code that a browser then interprets. - In a comparison test with **250 evaluators**, **61% preferred Solaris-generated interfaces** on "does it understand the instruction," and **71%** on "does it feel natural to use." - The chain it compresses: *idea → get a quote → write a requirements document → develop → revise three times* becomes *idea → describe it → have something clickable*. - What it cannot yet do: **a button must still be the same button on the fifth click.** Saved state, accessibility and error recovery are the hard parts of software, and they are not solved. ## The wall you hit at RMB 8,000 You have had this thought. You want a small tool to manage members at your shop. A sign-up page for a club. A medication reminder for an elderly relative. You ask for a quote. Eight thousand yuan, minimum, two weeks' wait. So you give up and keep struggling along with Excel and a WeChat group. That wall may genuinely be coming down. ## It works backwards from every other AI coding tool Here is how other AI coding tools operate: write a large amount of code, then let the browser translate that code into the picture you see. Solaris inverts it. **It does not write code. It draws the interface itself, frame by frame.** You click, you drag, and it predicts what the next screen should look like and generates that screen directly. Runway ran a comparison with 250 evaluators. On "did it understand the instruction," 61% preferred the Solaris result. On "does it feel natural to use," 71% did. What that means in practice: the long chain from an idea to something usable — quotation, requirements doc, development, three rounds of revision — is being compressed into *describe it, and have something you can click*. ## Before anyone announces the death of programming Some cold water, because that announcement is premature. Solaris's weakness is visible. **A button has to still be that same button after five clicks.** Saved state, accessibility support, recovery after an error — the things that look trivial are exactly what makes software hard. Generating one attractive screen is not difficult. Keeping a hundred screens mutually consistent, without losing data or crashing, is. It is also still an application-based early access programme, not something anyone can use. A more accurate read: - **What will be replaced:** simple interfaces that only need to look right and work — event sign-up pages, internal utilities, dashboard prototypes, in-store ordering screens. - **What will not be replaced soon:** systems involving money, multi-user collaboration, long-lived data, or high cost of failure. Put differently, AI eats first at the small requirements that were *too expensive to be reasonable but had to be done anyway*. That layer used to sustain a great many small outsourcing teams. ## Where the opportunity actually is The real change is not "fewer programmers." It is that **who gets to make a product has been rewritten.** Previously, building a tool required understanding technology or affording someone who did. Afterwards, what determines whether you can build it is how well you understand the scenario. You understand the pain of running a restaurant, so you can build yourself a shift scheduler. You understand the chaos of a parent group chat, so you can build a sign-up counter. Three directions you can move on immediately: - **Start with your own small annoyance.** Do not begin by trying to build a platform. Build one page that solves one problem of yours, and think about others only once you have used it. - **Convert "I understand the domain" into a moat.** Once the technical barrier collapses, what remains is who better understands what the user is struggling with. This is a dividend period for industry veterans. - **Do not bet on a new tool before checking whether it is stable.** Early products change fast. Prototypes and validation are fine; do not move critical operations onto one yet. Worth remembering: **when the cost of making something approaches zero, what is valuable is no longer being able to make it, but knowing what should be made.** Over the next two years the scarcest person may not be the one who can write code. It may be the one who knows clearly who a thing is for and what pain it removes. ## Honest limitations The preference percentages come from a vendor-run evaluation with 250 participants and have not been independently replicated. Solaris was described as early-access at the time of writing, so availability and capabilities will have changed. The "button must remain the same button" limitation is an interpretation of the state-consistency problem rather than a quoted statement. *Sources: Runway product material and evaluation results; contemporaneous reporting (September 2026). Information only.* --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-ai-draws-interfaces-not-code Free to quote with attribution and a link to the original.