Sora
OpenAI showed a model that generates minute-long video from a text description, with objects that persist and camera motion that stays consistent.
Why it matters
Video generation crossed the line beyond which the result stopped announcing itself as machine-made.
The model works as a diffusion transformer: video is cut into space-time patches and the transformer gradually removes noise from them. A length of up to a minute was an order of magnitude beyond earlier systems. OpenAI did not open the model at once, giving access only to artists and safety specialists, and named the risk of plausible fakes directly. Public release came only at the end of the year.