TLDR
China Talk host Irene interviews Xian Yu, author of the satirical AI blog 'Funeral AI' (障碍). They discuss how ordinary users' experience with AI matters more than lab benchmarks, the brutal manufacturing-like competition among Chinese model makers (Kimi, Zhipu, Qwen) who have caught up to Opus 4 level, a meme-based video benchmark where Minimax's open-source H3 beat closed models, China's massive AI adoption (Doubao's 200-300M users), and why the industry's 3-month hype cycles and concentration of 'abnormal' personalities make it endlessly fascinating to cover.
For you
PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.
Sign in to unlockSurprising Insights
Minimax's open-source H3 model beat closed-source Sora 2.5 and CogVideoX 2.0 in a user-voted meme video benchmark, despite being technically inferior on comprehensive metrics.
Doubao (ByteDance's chatbot) has 200-300 million daily active users in China — a scale rarely acknowledged in Western coverage.
Chinese model companies released three Opus 4-level models (GLM 4.5, Qwen 3.8 Max, Kimi K3) within a single month, matching the US pace of GPT-4.1, Opus 4, and Grok 3.6 releases.
Minimax HR tried to hire Xian Yu's partner the same day the company formally complained about their benchmark article — revealing a split between technical staff who value honest evaluation and corporate PR.
Agent token consumption is 10-50x higher than chatbot usage (hundreds of millions vs millions of tokens per hour), creating a compute demand surge that few anticipated.
Xian Yu believes the Chinese AI industry's mental health crisis (NPD, ADHD rates) correlates positively with hype intensity — the more money flows in, the more 'abnormal' the participants.
Key takeaways
Ordinary people's lived experience using AI products is more valuable and original than technical benchmarks, because frontier models are accessible at consumer prices ($20-100) and anyone can evaluate them.
Chinese domestic models (Kimi K3, Zhipu GLM 4.5, Qwen 3.8 Max) have reached parity with Anthropic's Opus 4 level within the past six months, making them fully viable replacements for most professional work.
Large model development is manufacturing, not research: there are no architectural secrets, talent circulates freely, and any lead lasts less than two months — competition is purely on data, compute, and execution.
China's AI adoption is far higher than commonly assumed: Doubao has 200-300 million daily users, and tens of millions already use desktop agents (Codex equivalents) from ByteDance, Alibaba, and Baidu.
The Chinese AI startup ecosystem runs on 3-month hype cycles that blindly follow Silicon Valley narratives (AI apps → world models → personal agents), driven by dollar-fund VCs who need familiar stories to sell to LPs.
User-centric benchmarking (like Xian Yu's 'meme benchmark' on B站) reveals different winners than technical benchmarks: Minimax's open-source H3 beat closed models like Sora 2.5 for entertainment because it's productized for that use case.
Open source is the 'united front' of AI: Minimax H3's release let everyone run a near-Sora-quality model locally, breaking ByteDance's API monopoly and forcing incumbents to respond.
Xian Yu writes AI satire not from tech optimism but because the industry concentrates extraordinary characters — fraudsters, NPD cases, people who've never done anything interesting building 'AI applications' — making it a uniquely rich subject for observation.
Notable quotes
“任何一个普通人都有权利来讲述自己使用AI时的感受”
▸ Notable quotes · 0:44“一句话生成AI没有解决信息输入的问题。伟大的向农说,传播信息的关键是降低噪音,没有原创信息输入,AI只是重复互联网已有信息,制造噪音”
▸ Notable quotes · 3:24“大模型是制造业。本质是算力和数据,你有足够的数据,你有足够的算力,那你的模型就没有道理很差。”
▸ Notable quotes · 50:37“你会发现你从结果上来看,这些模型厂就没有谁能够领先超过一个月两个月”
▸ Notable quotes · 53:43“中国有两三亿的人每天用豆包,以及中国还有更多的AI应用,所以AI在中国的普及率是相当高的”
▸ Notable quotes · 56:16“这个行业充斥着如此之多的骗子,NPD和你不知道他在干嘛的人,但这些人有的人居然还能拿到钱,有的人居然还能成为行业明星”
▸ Notable quotes · 61:37Chapters
0:00 — Intro & Guest Background: Irene introduces Xian Yu, author of the satirical AI blog 'Funeral AI' (障碍), which uses subculture style to write about AI from an ordinary user's perspective.
3:12 — Origin of 'Hype Dog' Critique: Xian Yu reads from his first article attacking one-sentence AI generation products as noise without original information input, and reflects on how the critique holds up a year later.
6:06 — AI Creativity & Benchmark Philosophy: Discussion of why AI writing quality has declined as coding ability improved (data trade-off), and why engineering benchmarks from user experience matter more than lab metrics.
14:22 — The Meme Video Benchmark Story: Xian Yu details their '梗 benchmark' testing video models on B站 memes: Minimax H3 beat closed models (Sora 2.5, CogVideoX 2.0) in user voting, causing despair, Minimax complaints, and eventual reconciliation.
28:45 — Chinese AI Startup Hype Cycles: Analysis of 3-month narrative cycles (AI apps → world models → personal agents) driven by dollar-fund VCs copying Silicon Valley stories to satisfy LPs, leaving most startups without real user insight.
38:40 — US vs China Hype & Term Creation: Contrast between US term-creation dominance (Benchmark, RSI, AGI) and Chinese reluctance to define new concepts, rooted in confidence gaps and funding structures.
49:20 — Large Models as Manufacturing: Core thesis: model development is manufacturing — no secrets, talent circulates, leads last <2 months, competition is on data/compute/execution, analogous to China's EV industry.
55:30 — Mass Adoption & Compute Surge: Doubao's 200-300M users, tens of millions on desktop agents (Codex equivalents), and agent token consumption driving unprecedented compute demand — with no bubble burst in sight.
Speakers
Irene — Host
Xian Yu (贤瑜) — Guest