OpenAI LLC today introduced a new feature that will enable developers to create custom versions of GPT-3, its powerful natural language processing model.

OpenAI provides GPT-3 through an application programming interface with on-demand pricing. Developers can integrate the API into their applications to perform tasks such as translating text and generating software code. The new customization feature that OpenAI introduced today is available as part of the API.

GPT-3 made headlines when it debuted last year because the artificial intelligence is capable of performing a large number of tasks with impressive accuracy. The AI can generate essays based on short text prompts. Developers have also used GPT-3 to write software code, translate text and solve simple math problems automatically, to name a few use cases. 

GPT-3’s versatility is the result of a unique architecture. The AI is based on a Transformer design, an approach to building natural-language processing models that is fairly popular in the industry. But GPT-3 is more sophisticated than the typical Transformer-based neural network. It features 175 billion parameters, the settings that guide how an AI processes data, significantly more than many other neural networks in the same category.

Another factor behind GPT-3’s ability to perform many different tasks is the way OpenAI trained the AI. The lab trained GPT-3 on hundreds of billions of data points collected from multiple sources. The result: GPT-3 can in some cases learn to perform a new task from only a few examples. 

In theory, developers don’t have to train GPT-3 on a large number of examples before integrating it into their applications. In practice, however, increasing the amount of training data used in a project can significantly improve processing accuracy. That’s what the new AI customization feature released by OpenAI today aims to facilitate.

Using the new feature, developers can optimize GPT-3 for a specific use case by training the AI on a custom dataset. OpenAI supports many different kinds of datasets. It also enables developers to continuously train GPT-3 on new information as it arrives. A company using GPT-3 to analyze customer feedback, for example, could retrain the AI every week based on the latest suggestions from users.

“It takes less than 100 examples to start seeing the benefits of fine-tuning GPT-3 and performance continues to improve as you add more data,” OpenAI researchers wrote in a blog post today. “In research published last June, we showed how fine-tuning with less than 100 examples can improve GPT-3’s performance on certain tasks. We’ve also found that each doubling of the number of examples tends to improve quality linearly.”

One OpenAI customer, venture-backed startup Viable, is using GPT-3 to power a service that helps companies analyze customer feedback and identify the most frequently recurring suggestions. OpenAI says that Viable boosted the accuracy of feedback summaries by 24% using the new customization feature.

An indirect benefit of training GPT-3 on custom datasets is it can help companies lower AI costs. 

To perform a task with GPT-3, such as summarizing a research paper, applications must send the AI a text prompt describing what it should do. OpenAI charges for GPT-3 usage based on the number of words in applications’ prompts. Customizing GPT-3 for a specific use case with a specialized training dataset can potentially enable the AI to carry out the task based on simpler prompts that contain fewer words, which reduces API costs.

OpenAI’s latest update highlights the growing industry focus on AutoML technology. Customizing neural networks with specialized training datasets can provide major reliability benefits. However, the task has historically been too complex and costly for many companies. AutoML is the umbrella term for a broad collection of products focused on easing the AI customization process to help enterprises realize a bigger return on their machine learning investments.

OpenAI has been customizing GPT-3 for different tasks even before today’s update. The lab offers several different versions of the AI with varying performance, accuracy levels and pricing. OpenAI has also worked with Microsoft Corp. to develop Codex, a specialized edition of GPT-3 that provides code autocomplete features for developers.

OpenAI’s product strategy has helped it win over big-name customers in the tech industry. The research group says Intel Corp., IBM Corp. and Salesforce.com Inc. are among the organizations using its GPT-3 API in their software projects. 

Image: OpenAI