OpenAI isn’t done living as much as the “open” in its name.
While the corporate isn't offering any of its latest models as open source, it has spent this week revealing more about the way it's tackling AI, what problems the technology exacerbates or enables (e.g. disinformation/deepfakes), and what its plans are has for the longer term.
Today it was revealed: “Model Spec“, a framework document designed to shape the behavior of AI models utilized in the OpenAI application programming interface (API) and ChatGPT and to gather feedback from the general public using a Web form hereopen until May twenty second.
As OpenAI co-founder and CEO Sam Altman I posted about it on X: “We will listen, discuss and adjust this over time, but I feel it would be very useful to be clear about when something is a mistake or a call.”
Why does OpenAI publish a model specification?
According to OpenAI, the publication of this working document is an element of its broader mission to be sure that AI technologies work in a way that is helpful and secure for all users.
Of course, this is far easier said than done, and doing so quickly moves into the territory of long-unresolved philosophical debates about technology, intelligent systems, computers, tools, and society basically.
As OpenAI writes in his Blog post announcing the model specification:
By publishing the primary draft, OpenAI goals to interact the general public in a deeper discussion concerning the ethical and practical considerations in AI development. Users can submit their comments through OpenAIs Model specification feedback form on its website for the following two weeks.
After that, OpenAI says it would release “updates on changes to the model specification, our response to feedback, and the progress of our research to shape model behavior” over the “next yr.”
However, in its blog post announcing the model specification today, OpenAI doesn’t specify how exactly it influences the behavior of its AI models – and whether a number of the principles written within the model specification are included within the “system prompt” or “pre-release prompt”. Used to align an AI system before releasing it to the general public – but it surely's secure to assume it would have a huge impact.
In some ways, the model specification seems to me to be comparable to Anthropic AI's “constitutional” approach to AI development, which was initially a key differentiator but which the latter firm has not fully highlighted in a while.
AI behavior framework
The model specification consists of three primary components: goals, rules and default behavior. These elements serve because the backbone for governing an AI model's interactions with human users, ensuring that they usually are not only effective but additionally meet ethical standards.
- Goals: The document sets out comprehensive, overarching principles to assist developers and end users alike. This includes helping users achieve their goals efficiently, considering the potential impact on a wide range of stakeholders, and maintaining OpenAI's commitment to creating a positive impact in the neighborhood.
- Regulate: To navigate the complex landscape of AI interactions, the model specification sets clear rules. These require compliance with applicable laws, respect for mental property, protection of privacy and a strict ban on the production of non-safe content (NSFW).
- Default behavior: The guidelines emphasize the importance of assuming good intentions, asking clarifying questions when needed, and being as helpful as possible without exaggerating. These default settings are intended to supply a balance between the several needs of various users and use cases.
Some like AI influencer and professor Ethan Mollick of the Wharton School of Business on the University of Pennsylvania have likened it to fiction “Three Laws of Robotics,” developed by science fiction creator Isaac Asimov back in 1942.
Others had issues with the present implementation, equivalent to how OpenAI's model specification causes the behavior of ChatGPT or other AI models. As tech author Andrew Curran identified XAn example from OpenAI included within the model specification shows a hypothetical “AI assistant” backing down and never difficult a user on their erroneous claim that the Earth is flat.
Continuous commitment and development
OpenAI recognizes that the model specification is an evolving document. Not only does it reflect the organization's current practices, but it’s also a dynamic framework that adapts based on ongoing research and community feedback.
This advisory approach goals to assemble diverse perspectives, particularly from global stakeholders equivalent to policymakers, trusted institutions and material experts.
The feedback received will play an important role in refining the model specification and shaping the event of future AI models.
OpenAI plans to maintain the general public updated on changes and insights from this feedback loop, reinforcing its commitment to responsible AI development.
Where can we go from here?
By clearly defining how AI models should behave through its model specifications and continually looking for input from the worldwide community, OpenAI goals to create an environment by which AI can thrive as a positive force in society – even in a time when she is faced with lawsuits and lawsuits criticizing training on the work of artists without express consent.
