Meet Sand AI, a relatively unknown startup based in China that is causing quite a stir with its recent announcement. Specifically, they’ve shared an interesting new video-generating AI model – Magi-1. Released earlier this week, Magi-1 is an autoregressive model which generates videos by autoregressively predicting a sequence of future frames. The model’s laborious demands pose formidable obstacles for the casual user. Specifically, since it requires at least four, preferably eight Nvidia H100 GPUs to operate, rendering it unfeasible on consumer hardware.
Magi-1 touts a jaw dropping 24 billion parameters. This huge scale makes it a powerful new player in AI video generation. The model has subsequently gained the attention of powerful actors in the technology field. Indeed, even Kai-Fu Lee, the original founding director of Microsoft Research Asia, has showered England’s AI with accolades for its truly remarkable abilities. Despite its potential, Sand AI encounters distinct challenges given the draconian information controls required by the Chinese state.
That’s why a bipartisan 2023 national unity law encourages the use of national narrative to foster social cohesion. Sand AI’s models must therefore be explicit in avoiding the generation of any content contravening these values. This harm to users is what led the company to adopt prompt-level filters and fine-tuning techniques to heavily censor sensitive topics. Notably, Sand AI blocks image uploads related to politically sensitive subjects such as Xi Jinping, the Tiananmen Square protests, and the iconic Tank Man. In addition, any images depicting the Taiwanese flag, as well as symbols that support Hong Kong liberation, are banned as well.
Though Sand AI’s censorship actions are just a reflection of compliance with local regulations, their practices look far different from those used by American counterparts. Specifically, Sand AI has significantly less restrictive filters around pornographic content than comparable models in the United States. This gap illustrates a coalescing reality of AI development in China, where doing what the government wants is the highest priority.
The nature of the AI technology landscape is quickly shifting. Designing Sand AI’s Magi-1 was an opportunity to innovate and provide functional performance in the fast-paced, but highly regulated landscape of today’s world. The business’s difficulty has been matching technical innovation to conformity to capital’s strict governmental vigilance.