A China-based startup, Sand Eyelaunched an overtly licensed video era AI mannequin that has attracted reward from entrepreneurs like Kai-Fu Lee, founding director of Microsoft Analysis Asia. Nevertheless, in accordance with TechCrunch assessments, Sand AI seems to be censoring pictures that would increase the outrage of Chinese language regulators from the host model of the mannequin.
Earlier this week, Sand Ai introduced MAGI-1A mannequin that generates movies by predicting a sequence of frames “autoreagressive”. The corporate claims that the fashions can produce top quality, controllable footage that captures physics extra precisely than rival open fashions.
👀I am unable to consider that is generated by AI
I am unable to consider it is an open components
🎬I am unable to consider it is freeA humiliating business video device with simply the MAGI-1 video mannequin
Particulars and examples under: 👇 pic.twitter.com/zlxrecweqh
– Fahan (@mhdfaran) April 21, 2025
MAGI-1 is so unrealistic that it can’t run on most client {hardware}. It has a dimension of 24 billion parameters and requires 4-8 Nvidia H100 GPUs to run. (The parameters are inner variables that use the mannequin mannequin to create predictions.) Many customers (which incorporates this reporter) are the one place the place Sand AI’s platform can drive MAGI-1.
The platform requires a “immediate” picture to start out video era. Not all prompts are acceptable, and TechCrunch was shortly found. Sand AI blocks picture uploads of Xi Jinping, Tiananmen Sq., Tank Man, Taiwan’s flag and insignia supporting Hong Kong’s liberation. Filtering seems to be taking place on the picture stage. I renamed the picture file and didn’t skirt blocking.
Sand AI isn’t just a Chinese language startup that stops importing politically delicate pictures to video era instruments. Shanghai-based Minimax’s era media platform Excessive-Ore AI can also be blocking images of Xi Jinping. Nevertheless, Sand Ai’s filtering seems to be significantly aggressive. hauiluo permits pictures of Tiananmen Sq..
As defined in wired within the January work, the Chinese language mannequin should comply with strict data administration. The 2023 regulation prohibits the mannequin from producing content material that “undermines the unity and social concord of the nation.” In different phrases, it contradicts the historic and political narratives of presidency. To conform, Chinese language startups typically censor fashions through prompt-level filters or fine-tuning.
Curiously, Chinese language fashions have a tendency to dam political speeches, however typically have fewer filters than American counterparts in pornographic content material. 404 was recently reported Many video mills launched by Chinese language firms don’t have fundamental guardrails that stop folks from producing unconsensual nudes.
