OpenAI LLC on Wednesday detailed ChatGPT, an artificial intelligence system that can understand and answer complex natural language questions.

ChatGPT is based on the advanced GPT-3 neural network that OpenAI first debuted in 2020. GPT-3 can perform a variety of natural language processing tasks ranging from writing essays to translating text. Moreover, it’s capable of generating software code based on user prompts.

OpenAI’s new ChatGPT system draws on the capabilities of GPT-3 to answer complex natural language questions posed by users. Moreover, ChatGPT can request clarification if it doesn’t fully understand a question at first.

During an internal demonstration, OpenAI researchers asked the AI system to identify an issue in a malfunctioning snippet of software code. In response, ChatGPT requested additional technical information about the issue. It then generated a detailed explanation of the malfunction along with a suggestion for how to fix it.

ChatGPT also includes new safety features. According to OpenAI, the AI system includes features that enable it to challenge questions with incorrect premises as well as reduce the risk of harmful and untruthful output.

OpenAI trained ChatGPT using a method known as reinforcement learning. With reinforcement learning, researchers give an AI system a task and provide it with feedback on how well it completes the task. Based on this feedback, the AI finds ways to improve its accuracy.

In the case of ChatGPT, the feedback used during the training process was provided not by OpenAI researchers but rather a second neural network. OpenAI trained the second neural network on a dataset of pre-prepared questions and answers. By analyzing the dataset, the neural network learned ways to hone ChatGPT’s question answering skills.

“Today’s research release of ChatGPT is the latest step in OpenAI’s iterative deployment of increasingly safe and useful AI systems,” OpenAI researchers wrote in a blog post. “Many lessons from deployment of earlier models like GPT-3 and Codex have informed the safety mitigations in place for this release, including substantial reductions in harmful and untruthful outputs.”

ChatGPT is the latest in a series of AI systems that OpenAI has developed using its GPT-3 neural network as the foundation. Previously, the research group detailed InstructGPT, a series of neural networks based on GPT-3 that can answer natural language questions more accurately than the original version. 

Earlier, OpenAI introduced an AI system called Codex that can generate software code based on user prompts. Codex is a version of GPT-3 that the research group trained on billions of lines of open-source code. According to the company, the AI system supports more than a dozen programming languages and is also capable of performing data science tasks. 

Image: OpenAI