Runway Launched Gen-3. Luma Already Does the Thing Runway Can't.

Runway opened Gen-3 Alpha to the public today, and the outputs look like marketing footage for a company that doesn't exist yet — slick, cinematic, temporally coherent in a way that would have seemed deranged eight months ago.

The catch, which there always is: no image-to-video. You give Runway words, Runway gives you pictures that move. You cannot give it a picture and ask it to animate that picture. This is a real gap, because Luma can do exactly that, and the results from Scene Generator are the kind of thing you'd use to pitch a film that you haven't written yet.

I ran the same prompt through both. Luma, with the source image as input, produces something that feels like the image is remembering what it was doing before the shutter clicked. Runway, operating on the same text without an image to anchor it, produces something different — not worse, just untethered. Hallucinated from description alone. Both are impressive in ways that feel slightly uncomfortable.

Meanwhile, Chrome published its rationale for shipping Gemini Nano in the browser — on-device inference, no server round-trip, the model lives on your machine and thinks there. The argument is privacy and latency and offline availability. The unspoken argument is that putting an AI in the browser is just the next thing browsers do now, same as they did with JavaScript and then with hardware-accelerated graphics, and the reasoning comes after the decision.

Three separate things happened this week and they're all pointing in the same direction, which is that the tooling for generating synthetic media is advancing faster than anyone has figured out what to do with it. This is not a warning. It's just an observation about where we are on a Tuesday in July.