The Future of UI in an AI-first World

[Grant Hulbert]

As the time—and cost—of producing new user interfaces trends toward zero,

what will future user experiences be like?

--The Innovation Office

"Vibe coding" has become the new norm: non-technical people can suddenly create and publish apps with simple user interfaces quickly, with very little experience. What used to require an engineering team and months of effort has shrunk to a few days. Let's project into the future of this trend, and imagine what will be possible in a few years:

While impressive, the current state of vibe coding still requires specialized tools: in most cases, the same toolset we've always used is still required to create and publish web apps. You still need an IDE, choice of language, libraries, build tools, a hosting environment, and a devops pipeline. It's great because it's fast, and reduces your team size drastically, but it's not fundamentally different: it's just a faster way to do what we've done for decades.

But speed isn't the real revolution. The real change is that the UI itself disappears and reappears as needed. Instead of being something we 'build' ahead of time, it becomes something we 'summon' in the moment.

So what if we could bypass all that predetermined design? What if UI could be created on the fly so quickly and easily that it's considered disposable? What if you didn't have to define buttons or widgets or charts, and AI just 'knew' what made most sense? What if the exact UI you need right now, in the moment, just appeared as soon as you need it, and disappeared and was replaced when your next need arose? Imagine if every website and every software product worked this way: would you need as much documentation or training to use them? Would you have to beg the developers for new features?

In a way, Google has already done this for the search use-case for decades: a simple page with a single text box that lets you type what you want to find, and results appear immediately. But it only finds things—it doesn't create them. If you type "Build me a spreadsheet app" into the search box, it's not going to create a whole new user interface for you. Search replaced navigating a directory of websites. Adaptive UIs will replace navigating menus and buttons. You’ll just state your intent, and the right interface will materialize.

Now, let's project GPU speed trends: as AI gets faster and more capable, the time to generate new UI drops from months, to days, to minutes, and finally one second or less. One day we'll see "App Autocomplete", as my teammate Sasson Jamshidi says!

Why is speed so important? It's all about 'flow': as you work with this new instantaneous UI, your mind stays in the moment, and you can explore and try different angles and viewpoints as you work towards your goal. Fresh ideas surface, and you can experiment freely without obstacles. The UI becomes your partner, helping you achieve goals in ways you may not have even thought of. You're not "building an app"; you're just getting work done, and the AI is deciding what UI would be most helpful in the moment.

What matters isn’t just that the UI is faster to produce — it’s that it no longer needs to be produced in the traditional sense at all.

With one notable exception, all of today's AI offerings that I've experimented with aren't quite fast enough to maintain a sense of flow when generating code in frameworks such as React or plain HTML with Javascript. Top reasoning models still take about 30 seconds to finish writing the code. Imagine if every Google search took 30 seconds, or a simple spreadsheet took forever to calculate a simple column total, or saving a file to disk took more than a moment! (I'm old enough to remember when I had to take a coffee break every time I compiled my small program...there was no 'flow' back then!)

Recently I was fortunate to be given generous online access to Cerebras' wafer-scale inference chips, which are 15X - 100X faster than other common GPU offerings today. Wow. I never want to go back. Using prompts I developed, large models like Qwen 235b were able to generate bespoke HTML+CSS and React code in about 1 second, and I can say from personal experience that I stayed in flow while experimenting with this new way of thinking about UI.

We're not quite there yet, but my experience with Cerebras proves that it's possible in the not-too-distant future. Someday, this will be the norm for all of us. What will your app completion do for you?