In a recent conversation published by ChinaTalk, independent writer Xianyu, creator of the subcultural satirical essay series Funeral AI, discusses the shifting landscape of China’s artificial intelligence industry, the decline of creative writing quality in mainstream models, and the creation of alternative benchmarking tools like MemeBench.
Beyond the Tech Hype and Corporate Grand Narratives
Xianyu started publishing Funeral AI to approach artificial intelligence through a subcultural lens rather than a purely technical one. While frontier labs pour tens of billions of dollars into developing advanced models and APIs, consumer access remains relatively affordable—ranging from roughly $10 to $100. This low barrier to entry means ordinary people retain the right to describe how these technologies actually feel to use, offering original observations distinct from internet noise.
Reflecting on an earlier stint in institutional tech media outlets like 36Kr and Huxiu, Xianyu notes that traditional publishing proved too buttoned-up, requiring tedious networking with industry elites. Opting for independence instead, the writer argues that humor and true creativity require conviction and non-consensus beliefs. While many consumers in China lean toward techno-optimism and view AI as a primary productivity multiplier, Xianyu maintains that ordinary human experiences matter far more than the trillion-dollar IPO feasts pursued by major model companies.
The Engineering Trade-Off: Programming Gains Versus Writing Decay
As models evolve, their technical capabilities have grown increasingly uneven. In late 2025, Xianyu stopped using artificial intelligence to draft articles because mainstream models have experienced a distinct degradation in creative writing performance. Here we are in September 2026: models are more advanced, but their originality and ability to express human thought through creative writing, rather than an engineering task, have declined substantially.
| Project / Model | Developer / Origin | Primary Function / Domain |
|---|---|---|
| Seedance 2.0 to 2.5 | ByteDance | Video generation and agent platforms |
| H3 | MiniMax | Open-source video generation model |
| MemeBench | Xianyu & Muqiu | Meme-ability ranking benchmark for video models |
Benchmarking Video Generation and Meme-Ability
To evaluate capabilities beyond standard technical metrics, Xianyu and partner Muqiu developed MemeBench, focusing heavily on the explosive growth of Chinese video-generation models. Over recent months, ByteDance’s Seedance—ranging from version 2.0 to 2.5—has largely dominated the international market, while platforms like LibTV in China and Higgsfield in the United States rely heavily on its underlying architecture. Meanwhile, MiniMax open-sourced its H3 model, performing at roughly 80 percent of Seedance 2.0’s comprehensive capabilities while allowing local deployment.
MemeBench tests these video models using unconventional material gathered from Bilibili and Douyin comments—such as bizarre pop-culture references and viral imagery—paired with wild prompts. When launching the project’s evaluation arena, Muqiu recruited university students via QQ groups to label outputs and rank initial results, offering a generous ten-yuan payment per task. This unexpected bounty led many students to question whether the job was a scam, highlighting the vast demographic realities of China’s digital economy.
