OpenAI Releases New Version of GPT-3 and Codex That Can Edit or Insert Content Into Existing Text

Text processing is a common task in many machine learning applications. These tasks deal with large amounts of text to conduct classification or translation, which necessitates a lot of back-end labor. It’s challenging to turn text into something that an algorithm can understand.

The OpenAI team has launched new versions of GPT-3 and Codex that can update or insert stuff into it. Instead of only completing the existing text, OpenAI API can now be used to alter existing content. This includes rewriting a paragraph of text or reforming code, thanks to these additional capabilities. The new work has opened up new possibilities while improving existing ones; for example, insertion is currently being tested in GitHub Copilot, promising early results.

Writing is more than just appending. It’s an iterative process in which current text gets changed, whether you’re dealing with text or code. GPT-3 and Codex can now alter the text in two ways:

  1. Editing what’s already there
  2. Adding new text in the middle of the material.

The new insert feature allows users to add contextually appropriate text to existing content. According to the team, providing future information to the model can increase the quality of completions for tasks like composing long-form material, transitioning between paragraphs, following an outline, and guiding the model to a conclusion.

The model creates a completion that is irrelevant to the second section without the context of future sections. The model provides a completion that connects the two sections together after the context of future sections is considered.

Insert is especially helpful while developing code. In software development, usually, the code is added to the middle of an existing file that already has code in it before and after completion. The model in the preceding example completes the missing function pruning while connecting to previously written code. The team has also added a docstring and missing import, which is impossible to do without first understanding the code. Insert is now being tested in GitHub Copilot, with early positive results.

The insert feature is available in beta today in the API, as part of the completions endpoint, and through a new Playground interface. Currently, the feature is compatible with the most recent versions of GPT-3 and Codex.

The team states that editing existing information is an important aspect of writing text and code. In beta, they have released edits, a new endpoint that changes existing text via an instruction rather than completing it.

The existing text is used as a prompt for editing and instructions on how to change it. The edits endpoint are then used to alter the tone or structure of the text and make specific adjustments such as spelling corrections. The edits function well on empty prompts, allowing for text generation like the completions endpoint does. 

API: https://openai.com/api/

Documentation: https://beta.openai.com/docs/api-reference/edits/create

Try Now: https://beta.openai.com/playground

Reference: https://openai.com/blog/gpt-3-edit-insert/

Tanushree Shenwai is a consulting intern at MarktechPost. She is currently pursuing her B.Tech from the Indian Institute of Technology(IIT), Bhubaneswar. She is a Data Science enthusiast and has a keen interest in the scope of application of artificial intelligence in various fields. She is passionate about exploring the new advancements in technologies and their real-life application.