OpenAI is sharing the first draft of its Model Spec, a document that specifies how models should behave in the OpenAI API and ChatGPT, to be used as guidelines for researchers and AI trainers who work on reinforcement learning from human feedback.
The spec, according to OpenAI, is based on the organizations’ research into model behavior along with input from domain experts to inform development of future models. OpenAI said the spec will likely change over time, as more is learned.
Among the objectives of the work is to give developers and users resources and to help them follow instructions, to consider benefits and harms to humanity, and to respect social norms and laws, according to an OpenAI blog post.
The document also includes instructions that OpenAi says “address complexity and help ensure safety and legality,” such as follow the chain of command, comply with laws, respect creators and protect privacy.
Some guidelines for default behaviors include assuming best intentions, ask questions for clarity, be helpful but don’t overstep, be objective and encourage fairness and kindness while discouraging hate.
According to OpenAI, the work will continue to hear from policymakers and domain experts regarding how they understand and support the approach and objectives, and to learn if there are additional objects, rules and defaults to consider.
