Which one among the following with regard to ChatGPT is correct ?

Correct — A, (a) It is based on large language model (LLM). The name says so: GPT stands for Generative Pre-trained Transformer, and a transformer of that kind trained on a very large body of text is what the industry calls a large language model.

What that means in practice. The model is trained on an enormous quantity of text and learns, in effect, how language goes together — which word is likely to follow which, in what context, at the scale of whole documents. The 'transformer' is the neural network architecture that made this possible, published in 2017, whose attention mechanism lets the model weigh every part of the input against every other part instead of reading strictly left to right. 'Pre-trained' means the bulk of the learning happens once, on general text, before any particular task; the system is afterwards adapted, and in ChatGPT's case fine-tuned on dialogue and further shaped by human feedback so that it answers in the form of a conversation. ChatGPT is therefore an APPLICATION — a chat interface with instruction tuning and safety measures around it — sitting on top of a large language model, which is what option (a) says.

The practical consequences follow from the same fact and are worth knowing for their own sake: the model produces fluent text without any guarantee that the content is true, it has no live access to information beyond what it was trained on and whatever tools it is connected to, and its answers vary between runs. Those limits are properties of a language model, not faults of a chat program.

Why the others are wrong

• (b)It is a predictive Artificial Intelligence — This is the option with a grain of truth in it, and the grain is worth naming: the model does work by predicting what comes next in a sequence of text. But 'predictive AI' is a term of art for a different family of systems — those trained to forecast an outcome or classify a case from historical data, such as credit-risk scoring, demand forecasting or fraud detection. ChatGPT belongs to the other family, GENERATIVE AI, which produces new content: text, code, images or audio. Option (a) states what the system is without ambiguity, and on a which-one-is-correct item that is the option to take.
• (c)It is a high level scientific computer language — A programming language is a notation in which a human writes instructions for a machine — FORTRAN, C, Python. ChatGPT is not a notation and has no syntax that a programmer writes in; it is a trained model with an interface. It can WRITE code in such languages, which is probably the association behind this option, but being able to produce something in a language is not being one.
• (d)It is a chatting software — This describes the wrapper and not the thing, and it is the most defensible of the three wrong options because the product does present itself as a chat window. Ordinary chat software transmits messages composed by human beings between them; it originates nothing. Here the other party to the conversation is a model generating its replies, which is the whole of what is interesting about the system. Choosing this option describes the interface and misses the technology the question is asking about.

A large language model is a neural network with a very large number of parameters, trained on a very large corpus of text, whose training objective is essentially to continue text plausibly. The transformer architecture, introduced in 2017, is what allows the training to scale: its attention mechanism lets each part of the input be weighed against every other part in parallel. Text is broken into tokens, roughly word pieces, and the model works over those. After pre-training on general text, a model may be fine-tuned for particular behaviour — in a chat assistant, to follow instructions and answer in dialogue, refined further by training on human preferences. The result generates language rather than retrieving it, which is why it can compose an answer never written before, and equally why it can state something fluent and false.

Key facts

• GPT stands for Generative Pre-trained Transformer; ChatGPT is a conversational application built on such a model.
• A large language model is trained on very large volumes of text and generates text by predicting the next token in a sequence.
• The transformer architecture, with its attention mechanism, was introduced in 2017 and underlies current large language models.
• Generative AI produces new content; predictive AI forecasts outcomes or classifies cases from historical data.
• Pre-training is general and done once at great expense; fine-tuning adapts the model to a task or a style of interaction.
• Instruction tuning and training on human feedback are what turn a raw language model into an assistant that answers questions.
• Language models can produce fluent but incorrect statements, since fluency and factual accuracy are different things.
• Training such models requires large clusters of parallel processors, which is why graphics processing units dominate the field.

Common traps

• Calling a chat assistant 'predictive AI' because it predicts the next word; the term denotes a different family of systems.
• Confusing the application with the model it runs on.
• Treating a language model as a search engine with guaranteed facts.
• Assuming that anything to do with computers must be a programming language.

• practice — not a real PYQ

  In the abbreviation GPT, as used in the name of large language models, the letter T stands for which one among the following ?

  • (a)Translator
  • (b)Transformer
  • (c)Text
  • (d)Training

  Answer(b) Transformer

• practice — not a real PYQ

  Which one among the following is an example of predictive, rather than generative, artificial intelligence ?

  • (a)A system that composes an essay from a prompt
  • (b)A system that produces an image from a description
  • (c)A system that estimates the likelihood of a loan default from past data
  • (d)A system that writes program code from an instruction

  Answer(c) A system that estimates the likelihood of a loan default from past data