OpenAI published a new version of its Model Spec on 29 October, the document that describes how its models should behave. The text is reorganised around five values: faithfulness to human intent, minimising harm, honesty and transparency, supporting human agency, and care. The revision adds guidance on context, memory, personalisation, oversight and the models' actions as agents. It also drops most of the detailed style guidance and renames the assistant as GPT.

What it means: As models do more than answer questions and start taking actions, such as making purchases, sending messages or changing code, how they decide what to do becomes more important. The Model Spec is the public document in which OpenAI explains the rules it tries to teach its models, and the new structure puts priority on what users actually want, not only what they literally asked for. The new guidance on memory and personalisation matters for personal data protection, because it sets out when the model draws on what it remembers from earlier conversations. For businesses building OpenAI models into their own products, the document helps them understand how the model will react to developer instructions that conflict with end users' interests. The EU AI Act requires providers of general-purpose models to be transparent about how they work, and documents like this form part of that record. Still, a statement of principles does not guarantee the model will always follow it, so businesses need to test its behaviour in their own use cases.

Read the original on OpenAI

Related service: Custom software

Related stories