American companies that also supply OpenAI, Anthropic & Co. likely also sell their training data to Chinese AI companies. This often involves Reinforcement Learning Environments, which play a major role in determining how well an AI performs a task.
Reinforcement Learning (RL) occurs after the pre-training of the model and ensures that an AI model becomes really good at a task – for example, "operating Slack" – through repeated practice in a test environment. The result is evaluated based on a set of predefined criteria, and the AI is rewarded or punished with points. Over the course of many iterations, the model becomes better and better.
Post-training is often the decisive factor that makes a difference between model versions. And these sets of different environments, criteria, and training data, these US companies apparently sell to Chinese AI operators.
Videos by heise
