A dialogue-based AI chat interface for its GPT-3 family of big language models was introduced by OpenAI on Wednesday under the name ChatGPT. During the testing phase, it is now free to use with an OpenAI account. ChatGPT offers a user-friendly conversational interface and is built to severely limit any damaging output, in contrast to the GPT-3 paradigm featured in OpenAI’s Playground and API.
According to OpenAI’s announcement blog page, “the conversation style enables ChatGPT to respond to follow-up queries, acknowledge its errors, contest false premises, and refuse unsuitable requests.”
People have been putting ChatGPT through its paces up to this point, discovering a wide range of potential uses while also researching its weaknesses. It can expand on the philosophical classification of a hot dog as a sandwich, develop new code, generate AI art prompts, write poetry, and describe the worst-case time complexity of the bubble sort algorithm—all in the manner of a “fast-talkin’ wise guy from a 1940s gangster movie.”
A lot of potentially destructive inquiries (about things like hate speech, violent content, or how to make a bomb) are also left unanswered by ChatGPT because doing so would contradict its “programming and purpose.” OpenAI accomplished this by prepending a special prompt to every input as well as by utilizing a method known as Reinforcement Learning from Human Feedback (RLHF), which allows an AI model to be adjusted based on how well-rated the responses it generates are by humans.
One of the main issues limiting large language models’ potential commercial utility is how to control their aggressive tendencies. According to OpenAI, ChatGPT represents a substantial incremental step toward creating a safe AI model for everyone.
Yet, predictably, some of ChatGPT’s built-in content filters have already been bypassed by users employing quasi-social engineering techniques, such as requesting the AI to frame a restricted output as a fictitious situation (or even as a poem). According to a report by Ars in September, ChatGPT also seems to be open to prompt-injection attacks.
Similar to GPT-3, its dialogue-based cousin is excellent at creating things up entirely in an authoritative manner, such as a book that doesn’t exist and includes specifics about its content. Another significant issue with huge language models as they exist right now is this: How can you believe any of their output if they can convincingly fabricate facts out of thin air?
However, as several have noted, the output quality of ChatGPT appears to be a significant improvement over that of earlier GPT-3 models, including the new text-davinci-003 model we discussed on Tuesday. ChatGPT, according to OpenAI, is a model from the “GPT 3.5” series that was trained using “a blend of text and code from before Q4 2021.”
GPT-4 speculations are still circulating in the meantime. It would be interesting to see what GPT-related improvements the company has been working on over the previous 12 months if today’s ChatGPT model is the climax of OpenAI’s GPT-3 training effort in 2021.