OpenAI API Release

OpenAI has launched a general-purpose API that allows developers to access AI models from the GPT-3 family. This release marks a shift toward providing a "text in, text out" interface that can be applied to virtually any English language task, lowering the barrier for creating AI-powered products.

General-Purpose Capabilities and Interface

The OpenAI API provides a flexible interface where users provide a text prompt and the API returns a text completion that matches the the provided pattern. This general-purpose approach differs from traditional AI systems designed for single use-cases.

Key functional capabilities include:

Few-Shot Programming: Users can "program" the model by providing a few examples of the desired task within the prompt.
Fine-Tuning: Performance on specific tasks can be improved by training the model on a provided dataset of examples, regardless of size.
Human Feedback: The system can learn from feedback provided by labelers or users.

Deployment and Model Architecture

The API currently runs models with weights from the GPT-3 family, incorporating speed and throughput improvements. OpenAI has designed the system to be simple for individual users while remaining flexible enough for professional machine learning teams to focus on research rather than distributed systems infrastructure.

Safety and Misuse Mitigation

Because the potential applications of generative AI are unpredictable, OpenAI is launching the API in a private beta. The organization is implementing several safety measures to prevent misuse:

Production Review Process: Before applications go live, they undergo a mandatory review evaluating the risk level, the openness of the application, and the plan to address potential misuse.
Access Termination: Access is terminated for use cases involving harassment, spam, radicalization, astroturfing, or other causes of physical, emotional, or psychological harm.
Constraint Implementation: To reduce risk, OpenAI encourages constraints such as human-in-the-loop systems, content filtration, input/output length limits, and topicality limitations.
Academic Research: Through an academic access program, OpenAI is collaborating with third-party researchers from institutions like the Middlebury Institute, University of Washington, and the Allen Institute for AI to study potential misuses.

Addressing Model Bias

OpenAI acknowledges that API models exhibit biases reflected in generated text. To mitigate these effects, the lab is taking the following steps:

Developing usage guidelines for developers.
Creating tools to help users surface and intervene against harmful bias.
Conducting internal research into fairness and representation.
Collaborating with developers to implement human-in-the-loop monitoring for sociotechnical systems.

Strategic Rationale for the API Model

OpenAI chose to release the technology via an API rather than open-sourcing the models for three primary reasons:

Funding: Commercializing the technology provides revenue to cover the costs of ongoing AI research, safety, and policy efforts.
Accessibility: The underlying models are large and expensive to run, requiring significant expertise to deploy. An API makes these powerful systems accessible to smaller businesses and organizations that lack the infrastructure of large companies.
Control: An API allows OpenAI to respond more effectively to misuse by adjusting access over time, which is not possible with an open-source release where access cannot be revoked.