Hirundo, Lazarus, and others are applying machine unlearning and weight edits to Chinese models to remove censorship and harden security for Western companies

If you use AI models trained in China, they may now offer better security and less censorship.

Western startups, including Hirundo and Lazarus AI, are modifying powerful Chinese open-weight AI models to address concerns about censorship and security for businesses operating outside China. These models, developed by companies like DeepSeek, Alibaba, and Tencent, are popular due to their low cost and high capabilities but are often seen as aligned with the Chinese Communist Party and potentially insecure. Startups employ techniques like machine unlearning and weight editing to remove state propaganda, geopolitical bias, and cybersecurity vulnerabilities. For example, Hirundo realigns models to meet Western safety standards, while Lazarus AI removes mandatory censorship by identifying and addressing banned topics. This process aims to make Chinese models more palatable and secure for Western enterprises, including those in highly regulated industries and government sectors.

The practice of realigning Chinese AI models addresses a growing market need for Western companies that want to leverage the cost-effectiveness and power of these models without their inherent censorship and security risks. Startups are positioning themselves as essential intermediaries, effectively performing 'neurosurgery' on AI to strip it of state-aligned bias and propaganda, and hardening it against cyberattacks. This trend signals a bifurcated market where the most advanced AI capabilities might require significant post-training modification to be adopted by Western entities, creating a new sub-industry focused on AI alignment and safety for geopolitical reasons.

This development also highlights a broader tension between open-source AI development and national security concerns, particularly as Western governments express apprehension about Chinese technology. Companies using these 'uncensored' models face a trade-off between retaining the original model's powerful, low-cost architecture and ensuring compliance with Western regulatory and ethical standards. The effectiveness and depth of these 'unlearning' techniques will be crucial, especially as concerns grow about deeply embedded biases and potential backdoors in AI systems, potentially influencing future global AI development standards and adoption strategies.