An investigation found nine American data companies, including Mercor, Surge AI and Turing, selling reinforcement-learning environments to Chinese labs such as Alibaba, Tencent and ByteDance for post-training.

What separates a decent model from a frontier one is post-training, and an RL environment is the simulated world it trains inside. Chinese labs buy the same environments American labs used and save serious compute.
What surprised me most is the markup: products have been sold for 1,000 times what they cost to make. Vendors outsource the work to contractors in countries like India, and those contractors post client names on LinkedIn; that is how the trail was found.
Should it be stopped? The authors themselves say it is not apocalyptic for now. But nine companies is only what they could find, and that is the part that really unsettles me.

Curated from high-quality sources, with concise summaries and key takeaways.