Topics
Latest
AI
Amazon
Image Credits:Anton Petrus / Getty Images
Apps
Biotech & Health
Climate
Image Credits:Anton Petrus / Getty Images
Cloud Computing
Commerce Department
Crypto
Sand AI’s online platform throws an error message when it detects a likely prohibited imageImage Credits:Sand AI
go-ahead
EVs
Fintech
Fundraising
Gadgets
Gaming
Government & Policy
Hardware
Layoffs
Media & Entertainment
Meta
Microsoft
Privacy
Robotics
Security
societal
Space
Startups
TikTok
exile
speculation
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
A China - based inauguration , Sand AI , has bring out an openly licensed , video - bring forth AI poser that ’s garnered extolment from entrepreneurs like the founding conductor of Microsoft Research Asia , Kai - Fu Lee . But Sand AI is likely ban the hosted version of its model to jam image that might elicit the ire of Chinese governor from the host version of the model , according to TechCrunch ’s testing .
to begin with this calendar week , Sand AI announcedMagi-1 , a model that generates video by “ autoregressively ” betoken sequences of physical body . The company take the fashion model can bring forth high - quality , controllable footage that captures physics more accurately than rival open models .
👀 You wo n’t believe this is AI - mother 🧠 You wo n’t consider it ’s open - beginning 🎬 You wo n’t believe it ’s gratis
Magi-1 video modelling just humiliated commercial telecasting cock
Details and examples below : 👇 pic.twitter.com/zlXRecWeqH
— Farhan ( @mhdfaran)April 21 , 2025
Magi-1 is too impractical to run on most consumer hardware . It ’s 24 billion parameters in sizing , and take between four and eight Nvidia H100 GPUs to unravel . ( Parameters are the internal variables model use to make predictions . ) For many user — this newsperson included — Sand AI ’s platform is the only place they can essay Magi-1 .
The program needs a “ quick ” image to sound off off video generation , but not all prompt are permissible , TechCrunch apace discovered . Sand AI block range of a function uploads of Xi Jinping , Tiananmen Square and Tank Man , the Taiwanese fleur-de-lis , and insignia supporting Hong Kong liberation . The filtering appear to be materialize at the image story — renaming image files did n’t skirt the blocking .
Sand AI is n’t the only Chinese startup preventing uploads of politically sensitive images to its television generation tool . Hailuo AI , Shanghai - based MiniMax ’s generative media political platform , blocks picture of Xi Jinping as well . But Sand AI ’s filtering is likely in particular aggressive ; Hailuo allows figure of Tiananmen Square .
As Wiredexplainedin a piece from January , example in China are required to follow rigorous information controls . A 2023 law forbids models from sire content that “ damages the single of the country and social concord ” — which could be interpret as content that counter the government ’s historical and political story . To comply , Chinese startup often ban their models by either using immediate - stratum filters or all right - tuning them .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
Interestingly , while Taiwanese models run to block political speech , they often have few filters than their American counterparts for pornographic content.404 recently reportedthat a phone number of video generators released by Chinese companionship lack basic guardrails that keep people from give non - consensual bare images of other people .
