one year on
OpenAI unveils Sora, a text-to-video model generating realistic scenes from text prompts
The new model can produce video from simple written prompts, but commenters quickly note artifacts and limitations.
OpenAI today released Sora, a text-to-video generation model that can produce video from simple written prompts. The model demonstrates a level of motion coherence and object permanence not seen in previous video-generation systems.
The demos draw praise for motion and object permanence, even as commenters point to artifacts in the legs and movement. Commentators on Hacker News, where the announcement is the top story with over 3,600 points and more than 2,200 comments, are split between awe at the realism and identification of glitches. Some note that while motion now appears natural at first glance, close inspection reveals artifacts in legs and consistent pacing issues that may become obvious with exposure.
The announcement reignites debates about synthetic media and truth. Several posters argue that video has never been a reliable source of truth, while others counter that the ease of generation at scale fundamentally changes the disinformation landscape. The thread also discusses potential military and political uses of the technology, from resurrecting deceased leaders to creating plausible deniability for real events.
Impressed by motion quality and object permanence, calls it a breakthrough in animation.
Points out artifacts like limping legs and uncanny motion, comparing to earlier issues with AI-generated hands.
Notes that while tells exist, the model represents a breakthrough that will be judged differently over time.
Argues that movements have a 2-keyframe-per-second feel, making them uncanny once noticed.
Warns that lower resolution and compression could make clips convincing enough to cause real-world harm.
One year later — open only if you can handle spoilers
Sora was not released publicly in the following year; OpenAI instead offered limited access via red teaming and artists. Competitors including Runway and Google followed with their own text-to-video models, but none achieved the same viral impact.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy