Startups recalibrate Chinese AI for the West

Startup Hirundo has developed technology to recalibrate Chinese artificial intelligence models to align with Western values and reduce geopolitical biases. This comes as Chinese open-weight large language models like DeepSeek's R1 have gained popularity among American startups and businesses due to their lower costs and customization options. However, concerns remain about the potential security risks and political biases present in these models.

Hirundo's CEO, Ben Luria, explains that the company's intervention, akin to a neurosurgery procedure, is designed to mitigate these issues by editing the weights of open models, enhancing their security, fairness, and reducing harmful knowledge. The company analyzes open models to identify and remove potentially dangerous information, such as bio weapons knowledge, and ensures the models can generate accurate responses on politically sensitive topics without engaging in censorship.

Some models, like Alibaba's Qwen, already demonstrate a lack of engagement on politically charged subjects related to China's history and sensitive topics. However, the alignment process is not straightforward, as Oriane Peter, an AI researcher, points out that it is about steering the model's output within a specific cultural context rather than eliminating bias entirely.

While Hirundo and other companies like Lazarus AI are successfully aligning Chinese models to Western standards, the process is complex and imperfect. Critics argue that achieving a universally safe model is an impossible standard, as any model must be adapted to the specific social context in which it will operate. As more countries develop their own sovereign AI models, the pushback against Western bias in AI systems intensifies, prompting the need for tools that uncensor and remove prohibitive elements from these models.