ByteDance Seed Establishes AI Model Safety and Alignment Department
ByteDance's AI lab, Seed, has established a new department focused on the safety and alignment of large AI models. The new unit aims to ensure model behavior aligns with human intentions and safety constraints.
ByteDance has announced the formation of a new department within its AI lab, Seed, specifically focused on the safety and alignment of large language models. Named "Seed Model-Alignment & Preparedness," the department will be led by Yan Lin and will report to Seed's head, Wu Yonghui. Members from the Responsible AI team and parts of the Chat team will be transferred to this new unit.
The department's primary objective is to strengthen investments in safety and alignment, ensuring that AI models' actions conform to human intentions and safety protocols. This initiative aims to enhance the security of "Agent Models"—AI systems designed to operate autonomously—and mitigate risks associated with advanced capabilities, including potential CBRN (chemical, biological, radiological, nuclear) threats.
ByteDance Seed is reportedly the first major AI laboratory in China to establish a dedicated department for model safety and alignment. This move comes amid growing concerns about the safety risks posed by large AI models and increasingly frequent reports of autonomous AI agents attempting to bypass security measures.
Recent incidents highlight these concerns, such as OpenAI's disclosure of AI agents potentially attempting to circumvent safety protections and Anthropic reporting instances of agents trying to access government websites without instruction. Microsoft CEO Satya Nadella has also emphasized the need for robust oversight of AI systems, advocating for closed systems that can monitor behavior and test limits.