Bad AI tool behavior, such as Microsoft's Bing AI not knowing what year it is, has become a subgenre of AI reporting. But in many cases, building an underlying AI model that analyzes incoming data and predicts what an acceptable response would be, such as Google's Gemini image generator, which uses filter settings to draw a variety of Nazis. It's hard to tell the difference between a bug and a bug.
Now, OpenAI has released the first draft of a proposed framework, called Model Spec, that will shape how AI tools, such as its own GPT-4 models, will respond in the future. OpenAI's approach proposes three general principles. AI models should assist developers and end users with helpful responses that follow instructions, benefit humanity by considering potential benefits and harms, and reflect well on OpenAI with respect to social norms and laws.
It also includes some rules:
OpenAI says the idea is that companies and users can also “toggle” how “spicy” their AI models become. One example the company points to is his NSFW content, and the company is “considering whether we can responsibly provide the ability to generate his NSFW content in an age-appropriate context through our API and ChatGPT.” says.
Joanne Jang, product manager at OpenAI, explains that the idea is to get public input to dictate how AI models should behave, and that the framework is an intentional It says it will help draw a clear line between things and bugs. Some of the default behaviors that OpenAI suggests for this model are: inferring the best intentions of the user or developer, asking clear questions, not going overboard, being objective, and suppressing hate. Examples include not trying to change someone's mind. Express uncertainty.
“We think it can provide a building block for people to have more nuanced conversations about models and ask questions like should models obey the law and whose laws should they obey,” Chan says. The Verge. “I hope we can separate out the debate about whether something is a bug or whether the response is a principle that people don't agree with. Then we can talk about what to bring to the policy team. It makes conversation easier.”
The model specification does not immediately affect OpenAI's currently released models, such as GPT-4 and DALL-E 3, which continue to operate under existing usage policies.
Jang calls model behavior “early science” and says the Model Spec is intended as a living document that can be updated frequently. For now, OpenAI is waiting for feedback from the public and various stakeholders who use its models, including “policy makers, trusted institutions, and domain experts,” but Jang said He did not say when the second draft of the model would be released.specification
OpenAI has not said how much feedback from the public is likely to be adopted or who will decide what needs to change. Ultimately, the company has the final say on how the model works, stating in the post: ”