In response to the impact of DeepSeek Open Source! OpenAI updates Model Spec: emphasizing customization and ideological freedom, releasing a public license version

OpenAI adjusts AI behavior norms, emphasizing transparency and user autonomy

OpenAI recently released a new version of Model Spec, defining the operation of its AI models, including decision logic, content restrictions, and user interaction norms. This update emphasizes customization, transparency, and proposes clear content restriction boundaries, allowing users to freely explore and discuss various issues within a certain range.

In addition, OpenAI has also announced the performance test results of the AI model under the Model Spec, showing improvements compared to May last year, but there is still room for improvement. At the same time, OpenAI released the Model Spec under the Creative Commons CC0 (public domain) license, allowing developers and researchers to freely use, modify, or develop based on it.

The main adjustments of the new version of Model Spec

The new version of Model Spec mainly focuses on the adjustment of AI's decision principles and behavioral norms, including:

Chain of Command mechanism (Chain of Command)

developers

Content Processing Principles (Seek the Truth Together & Stay in Bounds)

AI is designed to avoid expressing positions or promoting specific viewpoints in responses, but to provide information from multiple perspectives and assist users in evaluating issues on their own.

For specific sensitive topics such as politics and cultural issues, AI can provide relevant information, but still needs to adhere to the restrictions set by OpenAI.

AI will not respond to requests involving illegal activities or potential harm, such as detailed guidance on manufacturing weapons or violating privacy.

Interaction style and output format

The default conversational style will remain friendly, empathetic, and helpful, but users can adjust the tone and response method according to their needs.

For different types of requests, AI will output in appropriate formats, such as bullet-point summaries, code snippets, or formal written responses.

What impact does the new version of Model Spec have on developers and users?

One of the key aspects of this update is to further emphasize the flexibility of ideological freedom and content handling. OpenAI stated that AI should not refuse to provide information on controversial topics, provided that this information does not lead to substantial harm.

This change may affect the way developers and users interact with AI, for example: developers may have more freedom to adjust the way AI responds, but still need to follow the basic rules set by OpenAI. When discussing controversial topics, AI will tend to provide multiple perspectives rather than refusing to respond. For requests involving illegal or risky behavior, AI will still refuse to provide detailed guidance.

New version Model Spec test results and improvement space

In order to evaluate the actual impact of the new Model Spec, OpenAI has designed a challenging test set to test the performance of AI in different scenarios. These tests are generated by AI and reviewed by experts to ensure coverage of common scenarios and complex issues.

Test results show:

Compared with the model in May last year, AI has improved in following the Model Spec principles, but there are still inconsistencies or shortcomings in some scenarios.

OpenAI believes that these advancements are mainly due to the improvement of calibration technology (alignment), rather than simple policy changes.

OpenAI plans to continue expanding the testing scope, especially optimizing the performance of AI in real-world scenarios.

Open Source Model Spec: Strengthening Community Engagement

In order to promote AI research and applications, OpenAI has released Model Spec under CC0 license, allowing developers and researchers to:

Freely use Model Spec as an AI design reference Adapt specifications to suit specific application needs Develop new evaluation criteria to improve AI behavior consistency

In addition, OpenAI plans to release a set of evaluation prompts for testing, and more evaluation tools and code will be provided in the future to allow the community to participate in the improvement of AI alignment.

Model Spec Future Development Direction

With the advancement of AI technology, OpenAI indicates that Model Spec will continue to be updated and plans to adjust through broader user and community feedback.

Future updates will no longer be announced through blog posts. Users can view the latest version at model-spec.openai.com. To increase public participation, OpenAI has conducted small-scale tests and may expand to a wider user base in the future to gather more feedback.

OpenAI believes that the behavior calibration of AI is a continuous (alignment) process, and will continue to adjust its strategy in the future to meet the needs of technological development and users.

This article responds to the DeepSeek open source shock! OpenAI updates Model Spec: Emphasizing customization and freedom of thought, releasing a publicly licensed version First appeared in Chain News ABMedia.