What's this about?
Ethan Mollick generated a video — with sound — of an otter using a laptop on an airplane, using MiniMax H3, an open-weights video generator running entirely on his own computer. It took about three minutes. He posted it beside his own attempt at the identical prompt from under two years earlier, which he described at the time as looking "like something from 2022."
What exactly does "open weights" mean for video?
It means the model files are published and you can run them on your own hardware, with no API and no per-second billing. For video this is a bigger deal than for text, because video generation is enormously more expensive per output — which has kept it locked behind paid services. Running locally means no content policy other than your own, no upload of whatever you're generating, and no marginal cost beyond electricity. The otter-on-a-plane prompt has become an informal longitudinal benchmark precisely because it's absurd enough that nobody optimizes for it.
Why is this worth noticing?
- The controlled comparison is what makes it persuasive — same person, same prompt, same constraint of running locally, under two years apart.
- Audio generated together with video is the newer capability; until recently local video generation meant silent clips, which is why today's Pika Audio-style releases matter.
- It relocates the frontier question. The interesting number is no longer what the best hosted model can do, but how short the lag is before that capability runs on a laptop — and here it's measured in months.