OpenAI Launches Core LLM Safety Features for Developers

-

(Photo = Shutterstock)

OpenAI is finally delivering probably the most requested feature from developers. Included on this update is an API function that matches the output of a Large Language Model (LLM) to a JSON file, a document transfer format.

VentureBeat reported on the sixth (local time) that OpenAI launched the ‘Structured Outputs within the API’ feature, which ensures that the output of a generative artificial intelligence (AI) model matches the information stored in a JSON schema file.

JSON is an information transmission standard consisting of key and value pairs created for simple data exchange and storage. The receiving side of JSON data needs a method to validate that the information is structured based on the principles. JSON schema is a rule that checks whether JSON-formatted data is structured based on the set rules.

For instance, should you provide an API in the shape of JSON data to an external service, you could need to examine whether the right data is coming and going, each from the angle of the information receiver and the information provider.

JSON schema files are essential for AI development. They’re used to explain the content, structure, and data sorts of JSON documents that developers input into LLM. In addition they describe constraints on output based on this data.

A JSON document is a file that uses the JSON format to store and exchange text-based data. The format is popular since it is simple to make use of, flexible, and compatible with many programming languages, and it has grow to be an industry standard since it is simple for each humans and machines to read.

Nonetheless, LLM continues to be lacking in the flexibility to read JSON files in comparison with humans or general machines. When using JSON data, it often causes hallucinations or outputs toxic content. It’s because LLM can only partially follow the instructions within the JSON file.

For that reason, ‘structured output from API’ has grow to be probably the most requested feature amongst developers using OpenAI’s LLM. By ensuring that LLM output matches the content of the JSON file, developers can construct more consistent and secure generative AI applications.

(Photo=X)
(Photo=X)

OpenAI CEO Sam Altman said via X (Twitter) that the feature was released “by popular demand,” and in a blog post noted that “our strongest LLM yet, GPT-4o, achieved an ideal rating of 100% within the ‘Structured Output from API’ assessment.”

OpenAI first added support for JSON documents last yr when it introduced JSON mode, but it surely received mixed reviews because of issues with ensuring reliable output. The update says it is going to make it much easier for developers to constrain LLMs and ensure they match JSON schemas.

Moreover, structured output from the API has been shown to assist LLM higher understand more complex JSON schemas. OpenAI emphasizes that while each JSON mode and structured output from the API ensure valid JSON, only structured output from the API ensures schema compliance.

Now, developers do not have to fret in regards to the LLM they use missing required keys or mistaking invalid values.

Structured output evaluation results from API (Photo = OpenAI)
Structured output evaluation results from API (Photo = OpenAI)

OpenAI emphasized that “safety is OpenAI’s top priority,” and that “the structured output features in the brand new API adhere to established safety practices and permit the model to reject unsafe requests.”

Structured output from APIs is now available for GPT-4o, GPT-4o Mini, and a number of other custom versions. It may be used with a wide range of APIs, including the Assistants API, the Batch API, and the Chat Completions API, and is compatible with each text and vision inputs.

Reporter Park Chan cpark@aitimes.com

ASK ANA

What are your thoughts on this topic?
Let us know in the comments below.

0 0 votes
Article Rating
guest
0 Comments
Oldest
Newest Most Voted
Inline Feedbacks
View all comments

Share this article

Recent posts

0
Would love your thoughts, please comment.x
()
x