OpenAI Submits Comment to NTIA on Open Model Weights

OpenAI submitted a comment in response to the NTIA's Request for Comment on dual-use foundation models, stating that multiple pathways exist for safe and beneficial AI.

During the development of GPT-2 in the past, OpenAI adopted a staged release approach to assess the model's potential for misuse. Later, at the GPT-3 stage, OpenAI chose an API-based release strategy to secure research funding through commercialization of the technology and to enable immediate responses to misuse.

OpenAI also recognizes the value of the open-source ecosystem and has released the weights of models such as CLIP and Whisper. Open-weight releases offer the following benefits:

• Facilitating academic research into the internal structure of AI models
• Enabling users and organizations to run models locally on edge devices
• Allowing creative modification of models to suit users' purposes

At the same time, API- and product-based approaches like ChatGPT have the advantage of continuously mitigating risks discovered after release. For example, working with Microsoft, OpenAI was able to detect and block state-sponsored cyber threat actors abusing GPT-3.5-Turbo and GPT-4 — something that was possible precisely because the weights were not directly released.

In conclusion, OpenAI emphasized that both open-weight releases and API-based releases are essential tools for building a beneficial AI ecosystem.