Unveiling the Magic Behind ChatGPT
ChatGPT is an artificial intelligence (AI) chatbot developed by OpenAI. It’s built on top of this company’s GPT-3.5 and GPT-4 large language models (LLMs), which OpenAI fine-tunes for transfer learning. These methods include reinforcement and supervised techniques, which allow ChatGPT to make detailed, articulate responses to prompts. However, its tendency to confidently make factually incorrect responses is one of its drawbacks.
History
Versions of GPT prior to GPT-3 received limited public exposure. OpenAI published its first paper on GPT-1 in 2018, and GPT-2 was released in February 2019. GPT-3 was initially released as only an autoregressive LLM in June 2020, but was later fine-tuned. This updated version was released as GPT-3.5 on March 15, 2022, which is the basis for ChatGPT. This chatbot was released to the public in November 2022.
Language Models
A language model consists of a neural network with many parameters, typically billions in the case of an LLM. They first appeared in 2018 and were initially trained for specific tasks, such as entity recognition, mathematical reasoning and sentiment analysis. However, LLM now perform well at many different tasks, based on the amount of resources available to them.
For example, an LLM used to generate text will typically begin with simple tasks like predicting the next word in a sentence. Given a sufficiently high parameter count and training, such an LLM can simulate much of the semantics and syntax of a human language. Training also allows LLMs to learn many facts about the world.
Applications
Popular examples of applications for ChatGPT include story-writing, explaining difficult concepts and writing computer code.
ChatGPT can quickly generate short stories, especially if the audience is young children. It does best with a simple vocabulary and narrative, but ChatGPT can provide interesting alternatives to traditional bedtime stories. It can also explain complex concepts such as a mechanical device. Furthermore, users can prompt ChatGPT for greater detail on particular points. In addition, ChatGPT can provide code snippets to solve a specific problem, provided the user can describe the problem correctly.
Limitations
ChatGPT has several limitations that users should be aware of before using it. For example, it can write text that sounds plausible, but is factually incorrect. This behavior is technically known as hallucination and is common among LLMs. It occurs when the reward model ChatGPT uses to account for human oversight is over-optimized. This phenomenon is an example of Goodhart’s law, which holds that a quantity ceases to be a good measure when it becomes a target.
In addition, ChatGPT has limited knowledge of world events that occurred after September 2021, which marks the end of its training period. It also has a tendency to provide longer answers than most users prefer, because the human reviewers wanted answers with greater detail for training purposes. The data used in training also demonstrates algorithmic bias, especially with respect to attributes for people.
Future Developments
The next version of the ChatGPT family of LLMs is GPT-4, which has already been officially launched. It’s currently available only for ChatGPT, but that will soon change. In addition, Bing Chat is partially based on GPT-4, although it doesn’t have access to elements like visual input at this time. Insiders are also reporting that OpenAI could finish training GPT-5 later this year.
The integration of ChatGPT into other applications promises to be a major development in AI, which is already backed by a multi-billion dollar investment by Microsoft. This funding is already starting to pay off with the premium version of Teams, Microsoft’s proprietary business communication platform. Teams Premium uses ChatGPT to automate some tasks, like transcript generation. Microsoft 365 Copilot also integrates ChatGPT directly into Office applications like Outlook, PowerPoint and Word.
