Ever wished that automated help systems were more…helpful? OpenAI’s newest innovation, ChatGPT, works to optimize language models for dialogue, creating a more conversationalist tone when interacting with help systems. ChatGPT’s current version is working through complex instructions and, as a result, outputting long-form answers and content.
What is ChatGPT?
While it seems futuristic (and it kind of is), ChatGPT is the sibling model to OpenAI’s InstructGPT, which follows an instruction and offers a detailed response. For example, InstructGPT would receive a prompt like “Explain how a car works.” The model would create an answer like “A car uses a spark-ignited internal combustion engine to move.”
On the other hand, ChatGPT provides a more conversational, in-depth answer to this question. Using dialogue, ChatGPT is uniquely designed to ask follow-up questions, rectify mistakes, and reject inappropriate questions.
How Does ChatGPT Work?
At its core, ChatGPT is built on a technique called reinforcement learning from human feedback (RLHF). This uses human language to align the language model more closely with what a user is asking. With this type of technique comes safeguards like reducing harmful and untruthful statements that ChatGPT would’ve given if it operated like a traditional conversation bot model. However, Twitter threads have shown users getting around these restrictions, prompting some users to question the morality of the model.
ChatGPT mainly learns in three ways:
- A sample prompt is selected, and a human instructor models the desired answer. ChatGPT learns from the human response.
- A new prompt is picked, and ChatGPT offers up several answers. The human labeler ranks them from best to worst. This information trains the reward model.
- A new prompt is selected, and, using the reinforcement learning algorithm, an output is generated. The reward model selects a reward, and it’s updated.
How Does ChatGPT Threaten Google’s Search Business?
While ChatGPT is in its infancy stage, it’s already quite polished. Instead of anyone ever having to search for instructions or information about anything, all they need to do is ask ChatGPT to get an answer. It’s an alternative to Google’s search engine, and it threatens Google’s existence as there is now competition that could wipe out their entire business model.
Industry Perception of ChatGPT
Since its release, ChatGPT has garnered major attention from leading publications and media like The New York Times, TechCrunch, The Verge, and more. Reception has been mixed, with some users worrying about its correctness, leading Stack Overflow to ban answers generated by ChatGPT. In a conversation with TechCrunch, ChatGPT itself reiterated its goal was to help humanity but also referenced its own critiques and shortcomings – specifically, a lack of transparency, focus on short-term goals, limited scope, and conflicting areas of interest. While OpenAI’s goal for ChatGPT is to help humanity through the use of AI, that reality might be a ways off.
As to the future usage of ChatGPT, we’ll all closely watch to see how OpenAI tackles the potential problems of ChatGPT displaying a racist or sexist attitude, sharing incorrect information, and other problems.
How to Use ChatGPT
Experiment with ChatGPT here by creating a free account. Create an OpenAI account, or log in to your current one. OpenAI encourages user feedback, so you can submit any comments to the ChatGPT Feedback Contest. You’ll potentially win up to $500 in API credits.
OpenAI’s mission is to innovate with increasingly safe and useful AI systems, from helpful chatting in ChatGPT to artificial intelligence art generators in DALL·E 2. To stay up-to-date with the latest tech innovation, follow the cloudHQ blog.