From Medianama: 2024-04-08 02:17:21

OpenAI submitted to the NTIA advocating for an iterative deployment approach to AI models. It released model weights gradually to assess and mitigate risks before full deployment. The NTIA consultation sought feedback on AI model openness, benefits, and risks, emphasizing the need for clear standards and government involvement in risk assessment.

OpenAI suggests rigorous risk assessment for highly capable AI models to mitigate catastrophic risks before deployment. However, less resource-intensive models may not require extensive risk assessment due to lower potential for harm. The company argues for a balanced approach to risk management and innovation based on the model’s capabilities and investment.

OpenAI developed a Preparedness Framework to evaluate its models’ risks in high-risk domains and categorize them accordingly. Models with a high or critical risk are not deployed. Specific factors must be considered when assessing open-weight models, including downstream modification and the limitations of system-level safeguards against misuse.

OpenAI emphasizes the need for societal resilience to AI misuse and suggests governmental involvement in evaluating AI capabilities and risks. Measures to limit AI-misuse consequences, strengthen cybersecurity, and biosafety resilience are crucial. Developers must disclose severe risks to the public and work on awareness before releasing models, mirroring the responsible disclosure norm in cybersecurity.



Read more at Medianama: OpenAI argues for step-wise AI deployment to ensure model safety