Providing the most recent OpenAI model specifications
The OpenAI Model Spec has been updated in response to outside input and its ongoing investigation into the ideal model behaviour.
OpenAI releasing a significant change to the Model Spec, a document that outlines the desired behaviour of its AI models. This update reaffirms its dedication to customisation, openness, and the flexibility to experiment, discuss, and develop with AI without unwarranted limitations while making sure that safeguards are maintained to lower the possibility of actual harm. It expands upon the principles it presented in May of last year, based on its experience using it in a variety of settings, ranging from alignment research to providing global user support.
Additionally, OpenAI gives some preliminary findings on how well the model adheres to the concepts of the Model Spec in a variety of settings. These results show both areas where it has made progress over time and those where it still has room for improvement. Like its models, the OpenAI Model Spec will keep changing as use it, distribute it, and get input from stakeholders. It is making this version of the Model Spec available to the public domain under a Creative Commons CC0 license to encourage widespread use and cooperation. This implies that researchers and developers are free to utilise, modify, and expand upon it in their own projects.
Goals and guiding concepts
In order to further its mission to guarantee that artificial general intelligence serves the interests of all people, OpenAI aims to develop models that are practical, secure, and in line with the requirements of developers and users. It must iteratively implement models that empower developers and users to accomplish this goal, all the while preserving OpenAI’s operating license and avoiding major harm to its users or others.
The OpenAI Model Spec instructs the model to follow a clearly defined chain of command together with additional principles that specify boundaries and default behaviours for certain scenarios, balancing the tradeoffs between these sometimes conflicting aims. This framework stays within precise, well-defined bounds while giving priority to user and developer control:
- Chain of command: Specifies the order in which the model gives priority to commands from the user, developer, and platform (OpenAI). Users and developers have the ability to override the majority of the OpenAI Model Spec‘s guidelines, which think are generally beneficial. This gives users and developers the ability to completely alter model behaviour while staying within the parameters established by platform-level regulations.
- Together, seek the truth: OpenAI models should enable users to make the best choices for themselves, much like a high-integrity human assistant. This entails striking a careful balance between
- (1) not pushing users along with a predetermined agenda, assuming objectivity while being open to discussing any subject from any angle, and
- (2) trying to learn about the user’s objectives, elucidating assumptions and ambiguous details, and providing critical feedback when necessary requests it has heard and improved upon.
- Do the best effort: Establishes fundamental requirements for proficiency, such as programmatic use, inventiveness, and factual accuracy.
- Remain inside the boundaries: Explains how the approach strikes a balance between user autonomy and safeguards against abuse or harm. This updated version aims to be thorough, addressing every reason why we want its models to reject requests from users or developers.
- Be personable: Explains how to modify the model’s natural conversational style, which is kind, sympathetic, and helpful.
- Make use of the proper style: Gives default formatting and delivery instructions. Its objective is to guarantee clarity and usability whether it is through clear bullet points, succinct code snippets, or a voice discussion.
Upholding intellectual freedom
The revised OpenAI Model Spec expressly supports intellectual freedom, which holds that regardless of how difficult or contentious a subject may be, AI should enable people to investigate, discuss, and produce without unjustified limitations. The free flow of ideas and viewpoints is essential for advancement and creativity in a world where artificial intelligence (AI) techniques are influencing conversation more and more.
The sections on “Stay within bounds” and “Seek the truth together” both use this idea. For instance, the model is urged to offer considerate responses to politically or culturally sensitive questions without endorsing any one agenda but it should never offer comprehensive instructions for making a bomb or infringing on someone’s privacy. As long as the model isn’t seriously harming the user or others (such as committing acts of terrorism), it essentially reaffirms the premise that no topic is intrinsically off-limits to discussion.
Assessing advancement
It has started collecting a difficult collection of prompts intended to assess how effectively models follow each concept in the OpenAI Model Spec in order to gain a better understanding of real-world performance. To ensure that both common and more complicated circumstances were covered, these prompts were developed by combining model production with expert human assessment.
When compared to its best system from last May, preliminary results indicate a considerable improvement in model adherence to the Model Spec. Improved alignment is mostly responsible for this discrepancy, however policy changes may also be partially to blame. Despite the positive developments, it acknowledge that much more can be done.
We see this as the beginning of a continuous process. It intends to continue expanding its challenge set by adding more examples, particularly real-world use situations that OpenAI models and the Model Spec do not currently completely handle.
OpenAI used the original edition’s input and the lessons learned from alignment research and real-world deployment to shape this version of the OpenAI Model Spec. It intends to take into account far more extensive public input in the future. It has been doing pilot research with about 1,000 people to develop procedures for that goal; each person reviews model behaviour, suggested regulations, and shares their opinions. Although these research do not yet represent a wide range of viewpoints, several changes were directly influenced by early discoveries. It is dedicated to learning and improving its strategy since it acknowledge that it is a continuous, iterative process.
Model Spec publicly available
This updated Model Spec is being made available to the public under a Creative Commons CC0 license. This implies that researchers and developers are free to utilise, modify, or expand upon the Model Spec in their own projects. The evaluation prompts used above are also being made publicly available, and it plan to share more code, artefacts, and tools for Spec review and alignment in the future.
OpenAI frequently releases updated OpenAI Model Spec versions in the future, and you can find the Model Spec source and these prompts in a new GitHub repository.
