OpenAI rewrote its Model Spec around five values, with faithfulness to human intent first
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
OpenAI set out how its models should behave through five values and acknowledges that the models do not yet fully match them.
As The Next Web and Unite.AI relay, the updated Model Spec, the public rulebook for how OpenAI's models behave, is built around five values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. The company calls faithfulness to intent the foundation: the model should understand what a person actually wants rather than follow instructions literally. Sections were added on using context (past conversations, memory) and on supporting user autonomy, most of the style rules were removed, and the model is called GPT in the text instead of "the assistant." OpenAI writes that its models "do not yet fully reflect" the Spec, is testing new training methods, and asks the public to say where it disagrees with the balance it struck.
Sources 3