| Episode | Status |
|---|---|
This special ChinaTalk cross-post features Zixuan Li of Z.ai (Zhipu AI), exploring the culture, incentives, and constraints shaping Chinese AI development. PSA for AI builders: Interested in alignment...
Zixuan Li of Z.ai (Zhipu AI) discusses how Chinese AI labs build and ship models at breakneck speed, often releasing within hours of training completion. The conversation reveals that open-sourcing models is primarily a marketing strategy to gain Western mindshare since enterprises won't use Chinese APIs, while domestic revenue comes from subscriptions and coding tools. Key insights include China's distinct AI use cases like role-playing, the extreme talent competition, and how Chinese companies still view themselves as upstarts chasing Western leaders rather than peers.
Introduction to Z.ai (Zhipu AI) and their GLM 4.6 model, which ranks #19 on LM Arena and is competitive with top Chinese models. Discussion of the company's founding in 2019, early work on graph networks, and shift to LLMs in 2020, culminating in their breakthrough with GLM 4.5 and 4.6 models that gained international recognition.
Candid discussion of why Chinese AI companies open source their models - not ideological commitment but practical necessity. Western enterprises won't use Chinese APIs due to data sovereignty concerns, so open weights enable adoption through third-party platforms like Fireworks. DeepSeek's success demonstrated you can be famous globally while still generating business returns.
Deep dive into 'role-playing' - a culturally important Chinese use case where models follow very long system prompts to embody characters with specific emotions and behaviors. Z.ai specifically trained on role-playing data to serve platforms like Character.ai and Janitor AI, representing different fine-tuning priorities than Western chatbots.
Insights into China's competitive AI talent market where large companies like ByteDance and Alibaba have first pick due to higher pay, while startups like Z.ai attract people who want to 'fight together' and build something from nothing. Surprisingly, few colleagues even know Zixuan studied at MIT - credentials matter less than execution.
Z.ai's remarkable practice of releasing models within hours of completing training, with no pre-arranged marketing or benchmark submissions. This creates intense pressure on partnerships team to negotiate with US firms at 2-3 AM, but reflects company philosophy that 'open source itself is the biggest event.'
Discussion of hitting architectural limits with current transformer approaches. Z.ai believes there is 'a wall' that data alone cannot cross, requiring new architectures for breakthrough improvements. They experiment on smaller 9B-30B models before scaling to production 350B parameter models.
Z.ai's revenue strategy focuses on subscription plans (GLM Coding Plan) rather than API sales, as users prefer not worrying about token consumption. ByteDance dominates Chinese API market, while many enterprises require on-premise deployment. International revenue is 50% from US despite India having most users.
Z.ai's translation capabilities rival Gemini 2.5 Pro, with special strength in Chinese-English translation including memes, emojis, and cryptic internet slang. They actively collect meme data and train vision models to understand image-based memes, critical for platforms like Xiaohongshu and TikTok.
Chinese society views AI as valuable rather than scary, with developers being the most fearful group due to direct experience with tools like Cursor and Claude Code. General public sees AI as just another helper like SaaS tools. Long-term AI safety concerns are discussed internally but not prioritized over performance improvements.
Despite technical achievements, Chinese AI companies still view themselves as chasing Western leaders rather than competing as peers. DeepSeek's fame dominates even domestically - many Tsinghua students haven't tried GLM. Companies are happy to secure meaningful niches rather than aiming to beat OpenAI/Anthropic directly.
China's AI Upstarts: How Z.ai Builds, Benchmarks & Ships in Hours, from ChinaTalk
Ask me anything about this podcast episode...
Try asking: