Chinese AI blogger torches the 'world model' hype: mostly fine-tuned video models chasing VC money

葬AI · wechat · 2026-09-21

The Chinese WeChat blog 葬AI published a scathing takedown of the current "world model" hype, arguing it's a fabricated track where cash-strapped startups fine-tune open-source video models (mostly MiniMax H3) to raise VC money, differentiating only through naming gimmicks like "causal/social/cognitive world models."

Key points: Li Feifei and LeCun's very different research got lumped into one buzzword; four types of players are piling in (failing app companies, video-model startups like Shengshu & PixVerse, 3D companies like Tripo/Meshy, and investor-pushed PhD founders); MiniMax H3 is the only open model that can rival Seedance 2.0, making it the real engine behind the hype while MiniMax trades at 1/3–1/4 of Zhipu's market cap. VCs fund it because DeepSeek/Kimi valuations are out of reach and unfalsifiable. The author praises Vivix, which does real-time streaming video gen and never calls itself a world model. A live benchmark of real-time video models is promised.

Original post →

More from Companies & People

Companies & People channel →