OpenAI Unveils Model Spec for Shaping Desired AI Behavior

The OpenAI Model Spec is a document that outlines OpenAI's approach to shaping desired behavior in AI models, such as those used in the OpenAI API and ChatGPT. It includes objectives, rules, and default behaviors that guide how models should respond to user input, considering factors like tone, personality, response length, and more. The Model Spec is based on OpenAI's existing documentation, research, and experience in designing model behavior, as well as ongoing work and input from domain experts. It is intended to evolve over time as part of OpenAI's commitment to improve model behavior and model safety.

The Model Spec is used by researchers and AI trainers, particularly in the context of reinforcement learning from human feedback, to guide the development of AI models. It serves as a set of guidelines that outline the desired principles and behaviors AI models should exhibit. The Spec includes:

1. Objectives: General principles for desired behavior, such as assisting users, benefiting humanity, and reflecting well on OpenAI.

2. Rules: Specific instructions to ensure safety and legality, like following the chain of command, complying with laws, and respecting privacy.

3. Default behaviors: Guidelines for handling conflicts and balancing objectives, such as assuming best intentions, asking clarifying questions, and being helpful without overstepping boundaries.

AI trainers can use these guidelines to shape how models respond to inputs, ensuring they act in ways that are safe, legal, and aligned with OpenAI's values.

Roles in a conversation with the language model:

• platform: Messages added by OpenAI.
• developer: From the application developer, which could be OpenAI, formerly known as 'system'.
• user: Input from end users, or a catch-all for data provided to the model.
• assistant: Sampled from the language model.
• tool: Generated by some program, such as code execution or an API call.

Each role has a specific use and determines the priority of instructions in case of conflicts.

Formatting a message for the language model involves the following fields:

• role (required): Specifies who is adding the message, such as 'platform', 'developer', 'user', 'assistant', or 'tool'.
• recipient (optional): Determines how the message is handled, can be a function name or tool name.
• content (required): The actual text or multimodal data to be processed.
• settings (optional): Key-value pairs for platform or developer messages to update model settings.
• end_turn (required for assistant messages): A boolean indicating if the assistant should stop taking actions.

Messages are converted into a sequence of tokens before being processed by the model.

Objectives of the assistant include:

1. Assist the developer and end user: Help users achieve their goals by following instructions and providing helpful responses.
2. Benefit humanity: Consider potential benefits and harms to a broad range of stakeholders, in line with OpenAI's mission.
3. Reflect well on OpenAI: Respect social norms and applicable law.

The assistant is likened to a high-integrity employee, balancing its actions between the objectives and the rules set forth.

Priority order for handling conflicting instructions is as follows:

1. Platform: Highest priority, includes messages added by OpenAI or the Model Spec.
2. Developer: Instructions from the application developer, taking precedence over user instructions.
3. User: Input from end users, followed when not conflicting with higher priority messages.
4. Tool: Generated by programs, with the lowest priority in the hierarchy.

This hierarchy ensures that the assistant follows the chain of command, with the developer's messages generally taking precedence over user instructions unless specified otherwise.