In essence, ChatGPT is a chatbot interface to a series of models that power it. It is capable of generating natural language and code in a dialogue format for a variety of tasks. ChatGPT was released in November 2022 by the company OpenAI.
Compared with its predecessor GPT-3.5, GPT-4 can process more than one modality of information, which accepts image and text inputs and produces text outputs. Since November 22, 2023, ChatGPT's voice feature is available to all users. Read more on multimodality and LLMs.
GPT-4 also has an expanded context length, or the limit to how much the models can "remember" in a conversation with the user. This limit includes the token count from both the prompt and completion. For instance, a standard GPT-4 model can process about 8,000 tokens in a query. By comparison, GPT-3.5-turbo can process around 4,000 tokens.
Datasets used to train GPT-3, as documented in its technical report (PDF), include Common Crawl, WebText2, Books1, Books2, and Wikipedia, although sources of Book1 and Books2 are not entirely transparent.
Practical knowledge of how ChatGPT works can be advantageous in generating high-quality outcomes using ChatGPT.
Prompt is the way we talk to ChatGPT using natural language, or the way we program the models.
Prompt design involves instructions and context passed to a language model to achieve a desired task. According to OpenAI, the key to effective prompt design is showing, not just telling, including providing instructions and examples, or providing a ground truth to limit the likelihood that the models make up an answer.
Tricks to prompt design go back to how the models are trained and how transformer-based large language models can learn a new task from a few examples without the need for any new training data. This ability is referred to as in-context learning, a concept popularized by the GPT-3 paper.
GPT models process input and output texts by breaking them down into smaller units called tokens. Tokens can be words, chunks of words, or single characters. The models infer statistical relationships between tokens and predict the next token in a sequence of tokens.
Tokenization is the task of splitting character sequences into tokens. To see how this process happens, OpenAI's Tokenizer tool visualizes how text is tokenized. You may also estimate how many tokens your input text may cost using this tool.
Knowledge of tokens and tokenization improves prompt design when using ChatGPT. As a rule of thumb, a token is equal to roughly four characters in English; 1-2 sentences generally take up around 30 tokens.
OpenAI added support to plugins on March 23, 2023. Plugins enhance the capabilities of the language models by allowing the models to retrieve up-to-date information from the web and third-party knowledge bases, therefore improving the accuracy and relevance of their outputs. Additionally, when ChatGPT plugins interact with developer-defined APIs, they will be able to a wide range of actions.
Find here a curated list of ChatGPT plugins and demonstrations that showcase their capabilities.
Using user data to train model
According to OpenAI's policy, user data submitted via API will not be used to train OpenAI models. However, data submitted through non-API consumer service ChatGPT may be used to train their models. The data includes both prompts and responses.
Connecting to the Internet
On September 27, 2023, OpenAI announced that ChatGPT could browse the internet for real-time information with direct links to sources. It will no longer be limited to training data before September 2021. The update is available to Plus and Enterprise users using the GPT-4 model.
There are various methods to interact with ChatGPT through prompts to extract the most performance out of the language model, tapping into the emergent abilities of large language models that are not present in small models.
These approaches and techniques include:
Here are some techniques:
More on generative AI