Z.ai and Concordia AI Propose Six Stages for Managing Open-Weight AI Risk
Z.ai and Beijing safety consultancy Concordia AI have set out six stages for managing open-weight AI risk. The report focuses on the challenge of releasing AI models, as once they are out, they cannot be easily recalled.
Framework for Open-Weight AI Safety
The framework provides a comprehensive guide to navigating the tension between openness and safety. It starts with three key problems:
- Reversing a release is difficult or impossible.
- Fine-tuning can strip out safeguards.
- Developers lose monitoring capabilities once weights are out.
Six Stages of AI Risk Management
The report outlines six stages for managing AI risk:
-
Risk Identification: Covers misuse and accidents, defined by unacceptable risks across four dimensions: where the model runs, potential misuse, what it enables, and societal harm absorption.
-
Risk Analysis: Conducted before and after development and deployment.
-
Risk Evaluation: Categorizes models into green, yellow, or red zones, indicating full release, restricted/staged release, or suspension.
-
Risk Mitigation: Applies protections throughout the model's lifecycle.
-
Risk Governance: Establishes oversight and accountability.
Related Context
- Z.ai raised $5 billion in Hong Kong this month.
- Apologized over its ZCode coding tool and open-sourced it.
- The White House's technology strategy left open-weight AI off its list of critical technologies in August.