Inside Mirage News Network: A 24-Hour Experiment in AI Video

We ran an AI-generated news network for twenty-four hours.

The anchors, voices, motion graphics, and guest interviews were all generated with our models. The entire thing ran on a single machine, orchestrated by Claude, for about $50,000 in tokens.

MNN was built as an experiment. The question was pretty simple: can AI-generated video, and more specifically avatars, actually hold up when you don’t let it stop? A 24-hour broadcast felt like the best way to find out.

Why 24 hours?

Generating a convincing person for ten seconds is relatively straightforward. Generating that same person twenty minutes later is much harder.

Identity preservation is one of Mirage Avatar X’s core strengths, which made continuity the obvious thing to stress-test.

Typically — faces drift, poses reset, eyes slowly stop matching the reference. Eventually you’re looking at someone who is almost, but not quite, the person you started with.

Continuity is also one of those things you only notice when it breaks. So instead of showing a reel of our best clips, we wanted to demo how the model held up over time, under conditions that made continuity difficult to hide.

Also: twenty-four hours is the longest a single stream can run on X. We’d love to pretend there was a more principled reason for this time span. There wasn’t.

How it worked

We created four anchor personalities with Avatar X. Guests appeared through the same model. Voices came from Mirage audio. Tickers, data boards, quote cards, and other graphics were composited through our rendering engine. More than 100 original segments went to air.

The reporting itself was not ours. We licensed footage and reporting from a global newswire. Real, working journalists provided the facts. Mirage generated the presentation layer.

All scripts were built around this licensed footage, with each line mapped to a shot list so the words matched what was on screen.

Before anything went live, it passed automated checks for frozen frames, A/V sync, speech-tail timing, and asset rights. Even the clock was added at transmit time, so the time on screen was actually the time you were watching.

MNN featured AI-generated versions of real guests including Andrew Chen and Lenny Rachitsky. Everyone who appeared gave written permission and approved their talking points ahead of time. Every frame carried an AI-generated disclosure.

On the backend, playout ran through ffmpeg and RTMP using chained playlists. New programming could be added throughout the day without restarting the stream. Overnight, the whole thing ran for thirteen hours on a single process.

What we got wrong

Two factual errors made it to air. In one segment, we incorrectly captioned a woman at a protest as the President of the United States. In another, we gave a retired shepherd-turned-astronomer someone else’s name. We publicly corrected both on X within the hour.

But the most consistent technical criticism wasn’t actually about the video. It was about the audio.

In some cases, cadence was too uniform, sentences landed with the same weight, and pauses happened on a rhythm instead of where the meaning called for them.

While we’ve spent many months getting the faces right, long-form audio generation is where you can still really feel a gap.

What happened

More than 50,000 people tuned in live. The livestream generated more than 3 million impressions and trended on X throughout the day. Average watch time, however, hit around a minute. While it seems the novelty stopped the scroll, the storytelling did not.

Generating video is one problem. Holding someone’s attention is another. We’re getting closer on the first, but the second remains human-bound for now.

Watch the experiment here.