OpenAI, an Artificial Intelligence research company, recently released ChatGPT. This is a prototype dialogue-based AI chatbot that can understand and respond to natural language. In less than a week, it has been used by over a million people. Most users are amazed at how well the bot sounds. Some have even said it could replace Google since it can give solutions to complex problems directly.
Early ChatGPT adopters are demonstrating the technology’s ability to carry a conversation through multiple queries in addition to generating software code. This indicated that natural language processing is entering a new phase.
How does ChatGPT work?
ChatGPT interacts in a way that feels like a conversation. This dialogue format lets ChatGPT answer follow-up questions, admit when it makes mistakes, challenge incorrect statements, and reject inappropriate requests. ChatGPT is similar to InstructGPT, which is trained to follow the instruction in a prompt and provide a detailed response.
ChatGPT is based on GPT-3.5, which is a language model that uses deep learning to produce human-like text. However, ChatGPT is more engaging than the older GPT-3 model because it does a better job of generating detailed text. For example, ChatGPT can even generate poems. Another unique characteristic of ChatGPT is its memory. The bot remembers earlier comments in a conversation and recounts them to the user. ChatGPT was trained on an Azure AI supercomputing infrastructure.
The model uses a technique called Reinforcement Learning from Human Feedback (RLHF). OpenAI first trained the model using a method called supervised fine-tuning. Human AI trainers provided conversations in which they played the user and an AI assistant sides. The trainers also had access to model-written suggestions to help them compose their responses.
To create a reward model for reinforcement learning, comparison data was collected, that had two or more model responses ranked by quality.
Limitations of ChatGPT
ChatGPT sometimes produces answers that sound possible but are wrong or make no sense. This is difficult to fix because: there is no way to know if the answer is correct during training; if it is trained to be more cautious, it will avoid answering questions it knows; and supervised training gives the model a false idea of what is correct because the ideal answer depends on what the model knows, not what the human trainer knows.
ChatGPT can be affected by changes to the way a question is asked, or if the same question is asked multiple times. For example, the model might not be able to answer a question correctly if it is phrased in a certain way, but it might be able to answer the same question correctly if it is phrased differently.
The model often uses too many words and repeats itself. This happens because of biased training data and over-optimization issues.
Instead of asking clarifying questions when provided an ambiguous query, the current models usually guess what the user intended.
Open AI CEO, Sam Altman, recently tweeted, “ChatGPT is incredibly limited but good enough at some things to create a misleading impression of greatness. it’s a mistake to be relying on it for anything important right now. We have lots of work to do on robustness and truthfulness.”
Can ChatGPT replace Google?
A few years ago, big tech companies like Facebook, Google, and Microsoft said that digital assistants would be the next step in how humans and computers interact. They talked about how chatbots could do things like order Uber rides, buy plane tickets and answer questions in a way that seemed like a real person. Even though years have passed, not much progress has been made. Most chatbots can only answer simple questions or help customers with minor issues.
However, ChatGPT has opened to good reviews from users. People are finding new ways to use chatbots. Even though it has some limitations, many users find it helpful for routine things. For example, it can help a person who does not know coding to generate code to build an app or website from scratch. On the other hand, coders are using ChatGPT to debug codes they had written. Some even think ChatGPT can be embedded and used to help respond to generic emails or messages.
Google has developed large AI language models (LLMs) that are as good as OpenAI’s ChatGPT. These include BERT, MUM, and LaMDA. Google has used these to make its search engine better. These improvements help Google understand what users want when they use the search engine. Google has created apps like AI Test Kitchen to show people what its chatbot technology can do. But it has limited how users will be able to interact with the chatbot.
OpenAI was also careful at first about developing its LLM technology. But then it launched ChatGPT and allowed anyone to use it. This resulted in a lot of publicity and hype for OpenAI, even though the company spends a lot of money to keep the system free.
Some people think that artificial intelligence chatbots might soon take over regular search engines. But the people who work at Google say that the technology is not ready yet. There are still some problems, like chatbots’ bias, toxicity, and their inclination for making information up.