As models become increasingly capable, safety and alignment are essential to prevent risks while supporting the deployment of models that can benefit people at scale.

Seed recently established an Alignment & Preparedness team to further strengthen its models’ safety and alignment capabilities.

We are committed to investing in this field for the long term and providing the team with resources and support commensurate with frontier model research.

We plan to initially recruit 10 researchers worldwide. Together, we will advance model capabilities and safety in tandem, and continue to push the boundaries of intelligence.

What you will work on

Strengthen safety foundations and processes

Develop model specifications and safety policies, build rigorous safety evaluation systems, and improve models' ability to identify and defend against risks.

Continuously improve model safety

Explore more effective safety and alignment training methods across pre-training, mid-training, SFT, and RL, and use safety adversary training through attacker–defender self-play to strengthen safety over time.

Build a comprehensive safety system for emerging and unknown risks

Develop more generalizable safety training methods, improve models' judgment, and strengthen monitorability and guardrail systems while improving attribution for emerging and unknown risks.

Work with external partners to validate and improve model safety

Advance independent third-party evaluations, collaborate with universities and research institutions on frontier safety research, and continue to learn from issues users encounter in real-world use.

What you can expect

Substantial research and development resources

The team will provide the compute needed to sustain research on frontier models.

The opportunity to tackle the hardest problems in frontier AI safety

Study new safety challenges that emerge as model capabilities advance, and explore technical directions for which the field does not yet have mature solutions.

The opportunity to help advanced intelligence better serve people

As models evolve, we will continue to integrate safety research into model training, evaluation, and deployment so that model capabilities can be used safely and reliably in real-world products and services.

Who we are looking for

We welcome researchers from a wide range of backgrounds. You may be a strong fit if you:

Have a strong sense of purpose about advancing AI and believe in the long-term value of AI safety and alignment.
Bring deep curiosity and a passion for AI safety and alignment, and are eager to tackle difficult research problems.
Think from first principles, learn quickly, and can rapidly understand new problems, develop potential solutions, and test them in practice.
Have relevant research experience in areas such as foundation model safety, alignment, evaluations, red teaming, reinforcement learning, agents, or systems security.