Notes from inside China's AI labs
The article provides an inside look at China's AI labs, discussing their rapid progress, strong engineering talent, and focus on practical applications. It notes that Chinese AI companies are highly competitive with US counterparts, particularly in areas like large language models and robotics, while operating under different regulatory and resource constraints.
Background
- This article presents on-the-ground observations from visiting Chinese AI labs, offering a rare first-person look at how companies like DeepSeek, Alibaba's Qwen team, Baidu, and others are operating — often opaque to Western observers.
- DeepSeek has become a focal point because it achieved frontier-level reasoning with far fewer GPUs than US giants, using innovative "mixture-of-experts" architectures and heavy reinforcement learning. The piece clarifies *how* they did it and what their culture looks like.
- Chinese labs are now treating test-time compute scaling and inference-side reasoning (the "o1 paradigm" popularized by OpenAI) as their main race, not just pretraining bigger models. This is a major shift from 2023's focus on scaling laws alone.
- Key constraints: export controls on advanced NVIDIA chips (H100, B200) force Chinese labs to be ruthlessly efficient with hardware. This has created a distinct engineering culture — less fluff, more systems-level optimization.
- The author notes that Chinese labs are catching up fast but still lag on software ecosystem, tooling, and the depth of the research pipeline that comes from having US-style academic-industry crossover.