Three ways to see it

1. 1

   You have probably typed into ChatGPT, watched a colleague paste a report into Claude, or seen a Gemini answer pop up on a Google search. The chatbot replied in fluent English, sometimes in passable Urdu, and the answer felt like it came from a thinking mind. It did not. Behind every one of these systems sits a large language model, abbreviated LLM, and the goal of this lesson is to take that label apart, piece by piece, until the word stops sounding like a brand and starts sounding like a thing you understand.

2. 2

   Start with the word large. A model is large when it has many parameters, and parameters are the tunable numbers that store everything the model has learned. GPT-2, released in 2019, had about one and a half billion parameters and could barely write a coherent paragraph. The current generation of leading models works at the trillion-parameter scale. The leap in fluency you have noticed since 2023 is, in good part, a leap in size. Bigger models read more, store more associations, and produce more confident output. They also cost more to run, which is a fact that matters when a Pakistani ministry has to write the procurement.

3. 3

   Move to the word language. The model is trained on text. Trillions of words of text, scraped from the open web, from books, from code repositories, from forums, from news sites, from leaked datasets. The model does not read this text the way a child reads. It chops every word into small pieces called tokens, in English roughly four characters each, in Urdu often a full word or half a word depending on the tokenizer. Every Urdu sentence you have typed into ChatGPT was first cut into tokens before the model saw it. The quality of Urdu output you receive depends, in large part, on how much Urdu text was in that training pile, and the honest answer is: less than five percent of what was in English.

The why-tree