Generative AI first steps: How language models learn
Pick an option and check it. The explanation shows after each answer. Take as long as you like; nothing is timed and nothing is sent.
15 questions
Question 1 · easy
While a language model is learning, what does it do again and again with text?
A model learns by guessing the next word, checking the guess against the real text and adjusting itself a little, over a huge amount of text. Nobody writes a rule for each sentence, and it keeps learned patterns, not copies of the pages or saved search results.
Question 2 · easy
Arjun types into a chat assistant: "Explain in simple words how to cancel a train ticket", and pastes the SMS he got when he booked. In AI words, what is the text he sends?
A prompt is what you send the model: your question, your instructions and any text you paste in, like an order to a cook. Parameters, also called weights, are the learned numbers inside the model, and training is the learning stage that happened before Arjun ever used it.
Question 3 · easy
On the bus to college, Asha asks a chat assistant to explain photosynthesis, and it writes a reply. Which stage is the model in while it writes that reply?
Inference is the trained model at work, answering you, like a cook making your order. Training is the long learning stage that came before; while the model answers Asha, it uses what it already learned and is not learning from her question.
Question 4 · easy
Meena, a teacher in Balasore, asks: "Does a chat assistant keep every page it learned from, and find the right page when I ask?" Which reply is right?
A model is not a library: it keeps no copy of the pages it read and does not search them later. What changed while it learned is the model itself, a huge number of small settings, and it writes each answer from those patterns. That is why both the library and the summary pictures fail.
Question 5 · easy
Ravi runs a kirana shop in Sambalpur. He asks a chat assistant with no web search what year the weekly market near his shop began. It replies in one smooth, confident sentence: "The market began in 1974." What should Ravi understand?
After words like "began in", a year usually comes next, so the model writes a likely-looking year. Fluent shows a common pattern, not a checked fact: this assistant read no record and keeps no records, so a confident tone does not make the year safe to use.
Question 6 · easy
The three letters GPT appear in many AI headlines. What do they stand for?
GPT stands for generative pre-trained transformer, a type of model: generative because it writes new text, pre-trained because it was first trained on a huge amount of text, and transformer as the name of its design. General purpose technology and grammar prediction tool sound plausible but are not what the letters mean.
Question 7 · medium
Say a toy model has read only these seven made-up messages, in this order: - Please wait here. - Please call me tonight. - Please call Meena. - Please send the bill. - Please call back. - Please call after lunch. - Please send the photo.
Counting, which word does it predict right after Please?
After Please, call comes 4 times, send 2 times and wait once, so the prediction is call, the word that most often came next. Wait is what it could guess from the first message alone, send is only the last message read, and come never followed Please in this text. A real model does not keep a tally like this; it adjusts its settings.
Question 8 · medium
Nisha asks a chat assistant to draft a letter to her municipal office about a broken streetlight. The draft has a polite opening, her ward number, the date she first complained, and the office's helpline number. Which parts must she check against a record before sending?
For wording and tone, the learned pattern is what you want, so the polite opening only needs her own edit. Names, dates, numbers and amounts need a record, such as her complaint receipt or an official notice, and a smooth letter shows a common pattern, not checked facts.
Question 9 · medium
Kiran tells a chat assistant a fact about his village in Ganjam. Do the model's parameters change right then, while it answers him, so it learns his fact?
Answering is inference, and while a model answers, its parameters usually stay fixed: it uses what it learned and is not learning from the question. Some apps save chats, and a maker may use them to train a later version, but that is not the same model changing as it answers Kiran. Its learned numbers are still there after training; they are what it answers from.
Question 10 · medium
Sameer runs a small computer shop in Bhubaneswar. He wants to download a model and run it on his own office computer, so his customers' files never leave the shop. Which kind of model allows this?
An open model usually means its weights, the learned numbers, are published, so anyone can download and run it under the terms of its licence. A closed model is used through a company's app or service and its weights stay with its maker, so "closed" does not mean private on your own computer, and not every model can be downloaded.
Question 11 · medium
Priya wants a model to help write Odia exam notes. She compares two made-up models: one has 40 billion parameters, the other 8 billion (example numbers). What can she correctly conclude?
Parameters are a model's learned numbers, so 40 billion ÷ 8 billion = 5 times as many. A parameter count tells size, not quality, so "more parameters means better answers" fails; it does not count books read, and a bigger model usually needs more computing, so it is often slower, not faster. Only a test on her own task shows which suits her.
Question 12 · medium
Arjun's chat app shows "Searching the web" before some replies. His friend's app has no web search. Which statement about the two is right?
The model itself looks nothing up: it writes one next word at a time from the patterns it learned. Some chat apps add a web search that fetches pages for the model to read, so only Arjun's app reads pages. The friend's assistant still writes fluent replies, just from patterns, and a model keeps patterns, not pages.
Question 13 · hard
A made-up announcement makes four claims about a new model: 1: Anyone can download it and run it, under its licence. 2: It learned from a large collection of Odia news and textbooks. 3: It replies to a long question in about 3 seconds on a laptop. 4: It has 12 billion parameters.
Which tagging of the claims is right?
Ask of each claim: learning, answering, access or size? What it learned from is training, how fast it replies is inference, who can download it is access (an open model), and a parameter count is size, not quality. Tagging claim 1 as training mixes up where the weights came from with who can get them, and tagging claim 4 as quality reads size as better answers.
Question 14 · hard
For a school project, a student in Koraput asks a chat assistant with no web search for three things: - the year India became independent - the name of the first headmaster of her small village school - a polite closing line for her project report
Which answer needs the most checking against a record?
The smaller and more local the fact, the more checking it needs: a famous date appears often in text, so it is more likely right (though still worth a check), while one village school's first headmaster may never appear at all. A closing line is wording, where the learned pattern is what she wants, so it is not the one that most needs a record.
Question 15 · hard
A chat assistant with no web search writes Ravi a leave application in perfect office style. It also says state employees get a certain number of days of casual leave each year. Ravi says: "The letter is so well written, its leave rule must be right too." What is the best reply?
Good style shows the model learned a common pattern of office letters well; it does not show that any fact was looked up, because the model keeps patterns, not circulars. Use the wording and edit it, but check the number of leave days in a record such as the official circular. Rewriting all of it throws away the part the pattern is good for.
Your answers stay in this browser tab. Nothing is sent or saved.