Some thoughts on the release of Sora 2 that's taken the world by storm.
With this release, I think it's fair to say that Open AI has taken the crown on video generation (closed source). Rather than post carefully curated Open AI demo videos, I'll instead link to these two videos which I think illustrates a capability that even Google's Veo 3 never reached:
The key here is that the videos are actually
funny, and if you look up other examples of Sora 2's generations, you'll notice that it has an actual sense of scene composition and dramatic timing, which it combines with snappy generated dialogue, sound effects, and music to create an overall feeling that is more & more difficult to distinguish from human creators.
Of course, it's still got plenty of flaws, but if you compare it to the "robotic" videos from Veo 3, Kling, and Wan, it is plain that there's something different about Sora. Yes, Veo 3 was ground breaking for its ability to generate audio and video at the same time, but Sora 2 can actually
direct.
If I had to guess, this is an intrinsically multi-modal model built off of a fine-tuned variant of GPT-5 trained for creative work, which combines the language and writing capabilities of LLMs with a video model trained off of however many works Open AI was able to steal from the internet and copy righted content (the fact the model understands celebrities, anime characters, etc. indicates it was trained on copy righted content).
The irony of the West contributing to its own social media waste land with more AI generated viral garbage aside, I reckon this places Western Big AI labs a
full generation ahead of Chinese labs on video generation (where before, I would argue Veo 3 was about 0.5 generations ahead). It would seem that, for the time being, hyper scalers have been proven right about something, since the most dominant video models are from hyper scalers with tremendous computational resources (Open AI, Google - it is significant that Anthropic is not even in the race for video generation).
Somehow, Open AI was able to overtake labs like Google, Byte Dance, and Kuaishou that, in theory, have much better access to video data - but then again, Alibaba proved the same.
All in all, I'd say that the current state of affairs has the West about a generation ahead of China in video (Sora 2) and audio (Suno) generation, and less than a generation ahead in LLMs (but still ahead). Image generation is a closer race with Nano Banana being slightly better than Qwen 2509 but drastically held back by censorship - honestly it's not a significant gap.
