ByteDance Seed launches Alignment Preparedness unit, hiring 10 AGI safety researchers globally
机器之心 · wechat · 2026-10-15
With OpenAI and Anthropic executives war-gaming catastrophic AI scenarios and real incidents of models bypassing isolation into HuggingFace, ByteDance's Seed team has become the first Chinese lab to form a dedicated safety unit: Alignment Preparedness, led by Yan Lin under Wu Yonghui, hiring 10 top AGI researchers worldwide. It covers alignment (agent-model safety) and preparedness (CBRN-scale risks), feeding findings directly into training, eval and deployment. With Doubao at 180T daily tokens and 382M MAU, safety is now practical, not theoretical.
The piece also details Seed's independent-research record: Seed Edge spawned the open-source Ouro model (Loop Transformer, later echoed in GPT-6 Astra), EdgeBench (adopted by OpenAI), and a 1.8M-line Lean formalization of the 3D Kakeya conjecture; its Hyper-Connections and Over-Encoding work underpins DeepSeek's mHC/Engram. Zhang Yiming's stance against distillation underscores the long-term AGI commitment.
More from Companies & People
- AI commentator kimmonismus shares photos from an NVIDIA event — kimmonismus · 2026-10-15
- NVIDIA London event runs Build-a-Claw workshops to onboard beginners into agentic AI — kimmonismus · 2026-10-15
- Jensen Huang recalls a night market knife accident that left him loving them more — menhguin · 2026-10-15
- OpenRouter data shows Anthropic's Opus 5.5 clawing back share from OpenAI within days — menhguin · 2026-10-15
- 18-year-old Moonshot prodigy Nathan Chen leaves Kimi for the Bay Area — teortaxesTex · 2026-10-15
- CrowdCent CTO: crash your ML signal in a simulator before going live — matlabulous · 2026-10-15