ChatGPT capabilities

ChatGPT draws on a vast repository of data and information to generate new content in response to a prompt. As with other AI generative tools, the prompt could be a question, a request for an explanation, or a request to create an image or a poem. ChatGPT is more powerful than its predecessors because it has been trained on a larger data set. ChatGPT’s inputs included massive amounts of data and information drawn from the internet including books, webpages, coding information, and journal and news articles, as well as sources like Reddit discussion posts, which helped ChatGPT learn the rhythms of human dialogue. ChatGPT was then trained to use this data through a process called Reinforcement Learning with Human Feedback so that it could determine what humans expected when they asked a question.

What this means is, ChatGPT can respond in an authoritative, convincing, and human-like way to any prompt it is given. Among other things ChatGPT can:

  • Generate new text based on a given prompt or seed text (this can be used for tasks like creative writing, content generation, and language translation)
  • Complete a given text or sentence based on the context provided
  • Summarise a given text
  • Answer questions based on a given context or a text passage
  • Generate responses in a chatbot or virtual assistant.

As a data-to-text model (rather than a text-to-text model), the version of ChatGPT powered by GPT-4 has the ability to interpret complex graphical inputs such as charts, and to interpret and critically analyse images. As the launch video demonstrates, GPT-4 has the capacity to generate a website based on a quick hand-drawn sketch inputted via a photograph. It can also explain what makes something funny.

OpenAI has highlighted the latest model’s improved capacity to turn down inappropriate requests or queries that could generate harmful responses.

GPT-4 also has improved ‘steerability’: users can direct the tool to take on a particular kind of personality. This was an element of GPT-3.5, but is much improved in GPT-4. Examples provided in the launch stream included asking it to act as ‘TaxGPT’ by processing and providing advice on the inputted current US tax code, and requesting the tool to act as a Socratic tutor, responding to a user’s prompts with questions rather than answers.