Intelligent UI: When the Question Shapes the Interface

Original OpenAI demonstration frame. Publisher mark does not imply ownership.

Two Implementations, One Direction

On October 7, 2026, OpenAI described GPT-6 Intelligent UI in its GPT-6 announcement: answers can include visuals, buttons, forms, charts and interactive experiences, not only text. In the official video, a question about the North Star becomes an interactive guide you can step through. The page also notes that its Work and Codex models did not change in this release.

At Houston Hackathon 2026, I had been experimenting with the same direction in Tiersel: let the question shape the interface. A question about Houston's 311 data could surface a map, a trend, a ranking or an evidence table. You could click a chart, ask a follow-up and keep the context.

These are separate implementations. Nothing here claims that Tiersel uses GPT-6 or OpenAI's SDK, and nothing implies a city endorsement or a deployment with customers. I am comparing design ideas, not products.

A real frame from OpenAI's official clip. Source: OpenAI.

Where Tiersel Draws the Line

The Tiersel design splits responsibilities. A model, which I call Jev, helps choose the views that fit a question. Deterministic code calculates the numbers. Users can inspect the underlying records behind any figure.

That split is the point. If a model both chooses a chart and invents its values, a polished interface can make a wrong answer look authoritative. If code calculates and the model only selects and arranges, errors become inspectable: you can click through to the rows. In my view an interface should help you explore an answer and check it.

Design Criteria for Question-Shaped Interfaces

Whether you adopt a vendor's native approach or build your own, these criteria help:

1. Separate selection from calculation. Let the model decide which view fits. Let tested code compute aggregates, counts and rankings.
2. Make evidence one click away. Every number should link to the records or sources behind it.
3. Keep context across interactions. A click on a chart should become part of the conversation, so follow-ups refer to what the user is looking at.
4. Limit the component vocabulary. A fixed set of views is easier to test than arbitrary generated interfaces.
5. Define failure states. When data is missing or a question is out of scope, the interface should say so instead of producing a plausible chart.
6. Test with real tasks. Judge an interface by whether people finish their next step faster and with fewer errors, not by how impressive the demo looks.

What the Demos Do and Do Not Show

The video I shared on LinkedIn places OpenAI's complete original clip first, followed by my complete Houston hackathon demo of the real Tiersel UI. Counts and dates inside the Tiersel film are its original September 2026 snapshot. The demo music, "Digital Lemonade" by Kevin MacLeod, is licensed under CC BY 4.0.

Watch them for different reasons. OpenAI's clip shows what a first-party native implementation looks like. The Tiersel film shows one team's approach to inspectability. Neither shows measured user outcomes, and I make no claim about speed, accuracy or adoption.

You can watch the original OpenAI clip and the Tiersel Houston demo directly.

Why Agent Builders Should Care

For anyone building agents, the interface is part of the control surface. A text-only answer asks users to trust. An interactive answer can let them verify, compare and steer. That ties into the questions about permissions and evidence in the Frontier Academy article, and into the idea of small, checkable decisions inside a workflow covered in the Liquid AI d1 piece. The week's other releases are collected in the October 2–8 recap.

Read Next

The original post is on LinkedIn. Primary source: OpenAI's announcement. The Agent companion, Interactive answers inside an agent loop, looks at how this interface layer connects to agent operation.

If you are planning an interface where AI chooses views over your own data, our AI agent development service covers that kind of build. Or get in touch with the data and the question you want users to answer.