Notes from inside China's AI labs: less ego, more students, build don't buy

Nathan Lambert's firsthand notes argue that China's pace comes from how the labs are organized, not only from models. He describes researchers who will shelve a personal idea for collective model performance, in contrast to a US norm where individual recognition shapes choices. Graduate students are core contributors rather than rare interns, which brings in people not weighed down by old hype cycles. And the scientists he met show little interest in philosophical debates about AI, framing the job narrowly as building the best model.

The structural points are just as concrete. Companies like Meituan and Xiaomi build proprietary models instead of licensing, a build-don't-buy default. Labs develop reinforcement-learning environments in-house rather than buying expensive ones. Compute is the universal complaint, with NVIDIA access cited as the main limit. Lambert also pushes back on the idea that China lacks AI demand, noting most developers there are, in his words, "Claude-pilled." Read the full piece on Interconnects.

Why it matters

If you model the US-China AI gap from benchmark scores alone, this is the variable you are missing. Team structure and a build-don't-buy default compound over time, so they tell you more about the next two years than any single leaderboard does.