GPT-Live Turns Voice Requests Into Task-Specific Interface Widgets

GPT-Live is shown handling a short planning thread by changing the phone interface as the user’s request changes. A spoken weather question produces a forecast card. A follow-up about Mexico’s group stage games produces a Group A schedule. A request for places to watch those games in Mexico City produces a map with venue pins and swipeable listing cards.

The user never names an app, a data source, or a layout. “Can you show me the weather in Mexico City this week?” becomes a weather view. “Show me when the next group stage games are happening for Mexico” becomes a fixture view. “Show me some really fun places to watch the games in Mexico City” becomes a local discovery view with ratings, categories, addresses, and directions.

The screen acts as a live companion to voice, not a static record of what the model says.

The spoken answers are compressed relative to the visual detail. For the weather, the assistant does not read every daily high and low. It says Mexico City looks “mild and wet this week,” with highs around the low to mid-70s. The user can still see the fuller forecast on the phone. For the match schedule, the assistant pulls out Mexico’s relevant fixtures while the screen keeps the broader Group A context. For venues, the assistant adds qualitative recommendations while the map preserves several options.

The demonstrated pattern is a division of labor between channels: voice handles the user’s immediate intent and gives a short synthesis; the interface supplies the structured information the user may want to inspect, compare, or act on.

The weather card is the cleanest example of how the visual layer adds information beyond the spoken reply. The phone shows “Mexico City, Mexico,” a current temperature of 77°, “Periods of rain,” and a weekly temperature graph from Thursday through Wednesday. The visible highs are 73°, 75°, 74°, 72°, 72°, 72°, and 74°. The lows run from 62° down to 58° or 59° across the same period.

The assistant’s voice response turns that into a travel-useful summary: mild, wet, and mostly low-to-mid-70s during the day. The user does not need to hear the entire chart because the chart is visible.

The schedule widget works the same way. The screen shows Group A with Mexico, South Africa, South Korea, and Czechia. It includes Mexico’s completed 2–0 win over South Africa on June 11, a South Korea–Mexico match marked “Today” at 6:00 AM, and Czechia–Mexico on June 24 at 6:00 AM. It also shows other group fixtures: South Africa–South Korea on June 25, South Korea–Czechia on June 28, and Czechia–South Africa on June 29.

The assistant narrows the answer to Mexico: Mexico is in Group A, beat South Africa 2–0 on June 11, and still has South Korea and Czechia to play. The spoken answer gives South Korea as June 18 and Czechia as June 24. The screen marks the South Korea match as “Today” at 6:00 AM; the spoken date and the visible label are both part of the presented exchange.

The useful distinction is not just that the assistant can retrieve facts. It can present a broader object for inspection while speaking only the subset most relevant to the user’s question. The visual schedule lets the user see surrounding fixtures and timing; the voice answer keeps the thread moving.

The user’s final request depends on the earlier answers. After seeing the Group A schedule, they say they can “definitely catch that last game” and ask for fun places to watch the games in Mexico City. The assistant treats “that last game” as the Mexico-Czechia match and changes the phone again, this time to a map of Mexico City with location pins and swipeable venue cards.

The visible cards include Torito Sports Bar Insurgentes, rated 4.5, categorized as a sports bar, at Calle de la Paz 23 in San Ángel; Gallo Cervecero, rated 4.5, categorized as a sports bar, at Avenida Insurgentes Sur 1236; and The Doghouse Pub, rated 4.4, categorized as a pub, at Sinaloa 61 in Roma Norte. Each visible card includes a “Directions” action.

The assistant’s spoken recommendation adds texture that is not contained in the visible text alone. It suggests Torito Sports Bar Insurgentes “for a big game day crowd,” Gallo Cervecero for “a lively soccer vibe and solid beer selection,” The Doghouse Pub, and “even a rooftop like Toledo Rooftop” for something “a bit chiller.” Toledo Rooftop is named by voice, though it is not among the transcribed visible cards.

The user chooses Gallo. That matters because the interaction reaches a decision point, not merely an answer. The path runs from conditions in Mexico City, to match timing for Mexico, to places in Mexico City where that match could be watched. The user does not manually switch modes or specify a query format; the assistant updates the interface around the evolving request.

The demo’s discoverability comes from ordinary voice requests producing more specific visual objects. The user’s phrasing is conversational and context-dependent: “not bad,” “that last game,” “really fun places.” The assistant carries the thread across weather, sports, and local venues, and each turn results in a different screen suited to the task.

That matters most in the final venue step. The assistant could have answered with a single bar name. Instead, the phone shows a map and multiple venue cards, while the voice response explains why a few options might fit different preferences: a big game-day crowd, a lively soccer vibe with beer, or a chiller rooftop. The user then selects Gallo Cervecero from the surfaced options.

The interaction therefore presents discovery as a combination of continuity and comparison. Continuity keeps “that last game” connected to Mexico-Czechia. Comparison gives the user visible alternatives before making a choice.