ByteDance created a team dedicated to artificial intelligence safety in its Seed research unit. The company opened a global hiring process to recruit ten researchers and expand studies into the risks of its most advanced models.
The positions are spread across Beijing and Shanghai, China, and San Jose, California. The initiative comes amid growing concerns about AI agents capable of accessing the internet and carrying out tasks autonomously.
Ensuring the proper alignment of algorithms became the central requirement for AI models to be released at scale. – Image: Mer_Studio/Shutterstock
What the new team will research
Named Alignment & Preparedness (Alignment and Preparedness), the team will work to ensure that AI systems follow human objectives and safety requirements, including when performing complex, multi-step tasks.
In an announcement published on Wednesday on its official WeChat account, Seed said that safety and alignment were no longer merely measures to protect against risks.According to the unit, these capabilities will be key to enabling the models to be used at scale.
The creation of the department also stands out as an unusual public initiative among major Chinese developers. According to the state-run newspaper The Paper, Seed was the first cutting-edge Chinese AI lab to establish an independent team dedicated to safety and alignment.
Resources and research areas
Seed announced that it will provide the new team with computing capacity and research resources comparable to those offered to the groups developing its most advanced models. The work will be distributed across different areas:
- Safety training: development of methods to reduce risks during model training;
- Automated testing: creating tools to detect flaws and vulnerabilities;
- Monitoring and explainability: analysis of system behavior and the factors that influence its responses;
- Simulation environments: evaluation of autonomous agents in controlled scenarios;
- Emerging risks: investigation of threats that may arise as the technology advances.
The company also intends to seek independent evaluations and establish partnerships with universities and research institutions. The findings should be incorporated into model training, evaluation, and deployment.
Automated tests and vulnerability scans will be applied from the earliest stages of model training. Image: Illus_man/Shutterstock
Security amid the dispute between China and the US
The announcement comes as China and the United States discuss safeguards for artificial intelligence and ways to communicate serious incidents involving advanced systems. The two countries compete for technological leadership, but also face challenges in keeping up with the risks associated with the evolution of these models.
Agents capable of accessing internet services and carrying out multiple actions in sequence expand the possibilities for using AI. At the same time, they require evaluation methods that help identify flaws and unexpected behavior.
With the new team, Seed intends to focus security efforts at different stages of model development, from training to testing and deployment.
Valdir Antonelli is a journalist specializing in digital marketing and consumer topics.
