Inside Alibaba's AI Model Spec: Defining AI Values, Safety & Behavior | Jiayu Shen (Alibaba)

Jiayu Shen (Alibaba) on the company's model spec, a layered framework that turns high-level values into concrete AI behavior and safety boundaries. Shen, from Alibaba's security team, presents a spec built from six core principles (aligned with China's ethical norms for AI), 43 safety and helpfulness guidelines, and a four-tier permission hierarchy, root, system, developer, and user, that decides what a model does when rules conflict. He works through each principle, from human welfare and fairness to controllability in the agentic era and avoiding both sycophancy and preachy refusal. He then describes a three-phase method for checking whether a model follows the spec, introduces Oyster-2, a model built on Qwen3 14B using "constructive safety alignment" that resists jailbreaks, and outlines a co-construction process combining expert review with public feedback. The spec is open-sourced. Chapters 0:00 What is a model spec? 0:44 Six core principles and 43 guidelines 1:27 The four-tier permission hierarchy 1:46 Principle 1: advancing human welfare 2:35 Principle 2: fairness and justice 3:23 Principle 3: privacy and security 4:04 Principle 4: controllability in the agentic era 4:36 Principle 5: accountability 5:14 Principle 6: ethics and avoiding sycophancy 5:47 Resolving conflicts: root, system, developer, user 7:15 Evaluating spec compliance: three phases 8:29 Oyster-2: a trustworthy model built on Qwen3 9:26 Co-construction: experts and the public 10:19 Open-sourcing the spec More AI safety research: https://far.ai Alignment Workshop playlist: https://youtube.com/playlist?list=PLBY5kyt_LfFg&si=7grT5LoW2e7ztPVm FAR.AI is a research nonprofit working to ensure the safe development of advanced AI. We host the Alignment Workshop series and publish frontier alignment research.