Harshita Nimbal explains: The Evolution of AI UX

Introduction: How AI Redefined the Interface

[0:00]Hi, I'm Harshita Nimble, an AI product designer. This video is about how artificial intelligence actually changed the way we use software, not the hype, the interface itself. Here's the shape of it. For 40 years, UX meant one screen, built once, shown to everyone. Then the screen started guessing. Three

Act 1: The Fixed Interface Era (1984–2000s)

[0:20]Acts, the era before AI, when every interface was fixed. The era of prediction, when software began guessing what you do next. And the era now arriving, when the interface disappears into conversation. Act one. Before prediction, before personalization, decades where the interface was the same for absolutely everyone. 1984 to 2010. The graphical interface matures, icons, windows, menus, skeuomorphic buttons that looked like real switches.

Act 2: Prediction & Generative Interfaces (v0, Netflix)

[0:51]Every design decision baked in at ship time, tested once, shipped to millions unchanged. And here's the part worth sitting with. This exact screen, these exact menus, shipped to a teenager and a surgeon alike. The software had no idea who was using it, and it never asked. Picture the funnel. Every user enters the same front door, sees the same layout, follows the same three clicks to check out. Personalization existed, but a human had to notice the pattern and hard code the exception. Design at this scale meant Nielsen's heuristics, style guides, and a slash B tests read by hand.

Act 3: Disappearing Screens & Agentic Workspaces

[1:30]Every rule the interface followed, a person had written down first. The system could not learn, it could only be told. Act two. The moment the interface stopped waiting to be told and started predicting. One app, one code base, and now a thousand different front doors. A recommendation model watches what you click, and the layout itself starts to fork. Your home page is no longer identical to your neighbor's. Prediction crept in one keystroke at a time, autocomplete, then predictive text, then a search box finishing your sentence before you did. Small, cheap guesses running underneath every screen you already knew. By the 2010s, the guess

The Cockpit Exception & Summary

[2:10]Became the whole homepage. Netflix doesn't have one front page, it has hundreds of millions, each one assembled live from what you specifically tend to watch. Generative interfaces go one step further. The layout itself is no longer pre-built. A model reads the request and assembles the screen components, copy, even the layout tree, generated new every time for this one ask. Compare the two build processes side by side. A designer's fixed component tree versus a model choosing components from a library on the fly. Same building blocks, but nobody wrote the blueprint in advance. Tools built entirely on this idea now exist. V0 from Vercel is one of them. Type a sentence and a working interface appears, never designed by a human hand at all. What's behind me is its direct output for one prompt, and it's a loop, not a snapshot. Your clicks retrain the model,

The model reshapes the interface, and your next click retrains it again. The designer's job quietly shifts from drawing screens to shaping what the loop is allowed to learn. Act three, the logical end point of a system that can guess and generate, it stops needing a screen at all. Watch the toolbar itself start to vanish. Menus collapse into a single search field, then the search field collapses into a chat box. Every button that used to sit on the screen now just a sentence you type. One text box replacing an entire menu tree. Instead of navigating to the setting, you describe the outcome you want, and an agent finds or builds the path there for you. This is what that looks like today, a real screenshot of Claude, not a mock-up. The sidebar still holds your history and your files, but the work itself happens in one

Box. Whatever you want, you just type it. We asked Claude to sketch what comes after the chat box, an interface for an agent that acts, not just answers. An agentic interface isn't a chat log, it's a workspace where you watch the agent plan, call tools, and check its own work, one visible step at a time. The interface's job becomes narrating the machine's reasoning, not hiding it. Here's what would make this whole argument wrong, a domain where a fixed, hand-built interface is still the better choice. And one exists, an airplane cockpit. Every switch is identical every time for every pilot on purpose. Predictability is the feature, not the limitation. Push it to its end point, and the interface stops being a place you visit. It becomes a layer that meets you inside whatever you're already doing. 40 years of UX bends toward disappearing. Everywhere flexibility beats

Certainty as the thing worth optimizing for. Let's recap with Claude. For decades, one interface served everyone, hand-built and hand-tuned. Then prediction let the interface fork per person. Generative models let it assemble itself on request. And now agents are collapsing the whole thing into a conversation that acts on your behalf. The through line, control keeps moving from the designer's hand to the model's judgment. Your turn. Take one interface you use every day and ask Claude, which of its screens are still hand-built for everyone, which ones already predict or personalize for you specifically, and what would it look like as a single conversational agent instead? Ask Claude to sketch all three versions of that one screen, the static one, the predictive one, and the agentic one, and tell you what your product would gain and lose at each step. How AI changed user experience. I'm Harshita Nimbolkar.