Today, I had a meal with a classmate who had returned from overseas after many years, and we talked about the globally popular generative AI.

Regarding Sora released by OpenAI this year, the most obvious difference compared to domestic "text-to-video," "image-to-video," and "video-to-video" technologies is that domestic AI video-generated content lasts only a few seconds, while Sora can generate up to 1-minute videos based solely on text descriptions. However, my classmate believes that Sora's real advantage lies in generating multi-shot, coherent, and smooth videos with intricate scenes and vivid character expressions, especially in facial details and scene transitions. He thinks Sora's fundamental competitiveness comes from its ability to learn complex image rendering, intuitive physical behavior, long-term planning, and semantic understanding through denoising and gradient calculations, thereby simulating real or fictional worlds, predicting upcoming changes, and formulating strategies accordingly.

This reminds me of $Baidu(BIDU.US)'s "ERNIE Bot" and Moonshot AI's Kimi, two large language models. Kimi's standout feature at launch was its ability to process 200,000-character texts, surpassing last year's "ERNIE Bot" and other large language models. However, this isn't a technological breakthrough—it's just a refinement of existing methods, unlike Sora's clear technical lead over domestic generative AI video. This year, Baidu quickly expanded its long-text processing capability to millions of characters, proving that text length isn't a significant technical barrier.

Even when I create content, I often use Baijiahao, which helps filter out typos and punctuation errors—Kimi isn't noticeably better than Baidu. From the perspectives of short-term computing power competition and long-term commercialization models, Kimi doesn't seem to have any groundbreaking tech that puts it generations ahead of "ERNIE Bot." In fact, Baidu's commercialization scenarios are more advanced!