How a language model learns: guessing the next word
Tests: 5 of 10 right · Exam: 3 of 4 right · Kind: plain · Understood: 8 of 14 · Run: 2026-10-18, commit f3c23ea72
Lesson 2 of the AI course of Mova's primary school, as Mova's small model took its tests: it had read this lesson and the ones before it and learned from their exercises. Each test shows the options in the order the student saw them, its choice, the right answer, the score of every option and why it chose. The lesson itself, sentence by sentence with trees and pragmatics: the lesson in the corpora.
Every question of this lesson and each of its options parsed by Mova, with trees and pragmatics: the lesson's questions in the corpora.
Test 1 · right ✓
"my bot says it's 2023 lol, is it broken" What is likely going on?
- 1. It lives in a different time zone from you — no support
- 2. It is broken inside and must be thrown away and replaced with a new one — no support
- 3. Its training text ended long ago and it has no clock — ← Mova's choice · ✓ the right answer · score 6
- 4. It is joking because it is bored — no support
Why: It took the question to ask which statement is true. It chose option 3 with score 6, the only option with support. Its support: lesson 2, line 9: «Its training text stops at a certain date, so a bot without a clock or web search may not know today's time or news.», reached through a sentence linked to the one the question matched. What matched: the same doer «training text» (+4); the same action «stop» (+2). · the question and its options parsed, with trees and pragmatics
Test 2 · right ✓
What would happen if you left a chatbot app closed for a whole week?
- 1. Nothing; it would not miss you or think about you — ← Mova's choice · ✓ the right answer · score 3
- 2. It would dream about your last chat — no support
- 3. It would plan what to say next time — no support
- 4. It would feel lonely and sad, waiting for you to come back — no support
Why: It took the question to ask what happens next. It chose option 1 with score 3, the only option with support. Its support: lesson 1, line 2: «Think of a jar of a million sweets: after tasting enough of them you would guess that red means strawberry and green means lime. AI "tasted" millions of senten…», reached through the option's own match among the sentences about the question's things. What matched: the same action «think» (+2); the same circumstance (+2). Against it: a different doer (counts against) (-1). · the question and its options parsed, with trees and pragmatics
Test 3 · right ✓
"bots are just parrots, they copy and paste, end of" Is that fair?
- 1. Not quite; they mix patterns into new sentences, though they learned from people — ← Mova's choice · ✓ the right answer · score 0
- 2. Yes, every answer is copied word for word — no support
- 3. No, bots invent everything with no help from people — no support
- 4. Yes, they are actual birds inside a computer — no support
Why: It took the question to ask which statement is true. It chose option 1 with score 0, the only option with support. Its support: lesson 2, line 14: «A bot is more than a parrot, because it can mix ideas in new ways, but everything it knows came from what people made.», reached through the option's own match among the sentences about the question's things. What matched: the same action «mix» (+2); the same doer «bot» (+2). Against it: a different thing acted on (counts against) (-1); one says «not», the other does not (counts against) (-3). · the question and its options parsed, with trees and pragmatics
Test 4 · wrong ✗
Grandma says thank you to the speaker. Her grandson laughs. What is the kind truth?
- 1. She should apologise to it for laughing — ← Mova's choice · no support
- 2. It is silly to be polite to anything — no support
- 3. Being polite is a good habit, but it is a device — ✓ the right answer · no support
- 4. She is right to, because it gets upset and sulks otherwise — no support
Why: No option found support in what it remembers, so it guessed: it took the first option shown. · the question and its options parsed, with trees and pragmatics
Test 5 · wrong ✗
In a cartoon a chatbot snores in its box at night. What is wrong with that?
- 1. Nothing, bots sleep eight hours a night — no support
- 2. Chatbots only snore in the morning — ← Mova's choice · score 2
- 3. Bots dream but never snore — no support
- 4. A program that is not running does nothing at all — ✓ the right answer · no support
Why: It took the question to ask which statement is true. It chose option 2 with score 2, the only option with support. Its support: lesson 1, line 4: «A chatbot that says "good morning" sounds friendly, but there is nobody inside it who woke up, ate breakfast or will miss you at school.», reached through the option's own match among the sentences about the question's things. What matched: the same doer «chatbot» (+2); the same property (+1). Against it: a different action (counts against) (-1). The right option, 4, found no support in what it remembers. · the question and its options parsed, with trees and pragmatics
Test 6 · right ✓
Ask ten bots for a number from 1 to 10. Which is most likely?
- 1. Many of them give 7, like many people do — ← Mova's choice · ✓ the right answer · score 3
- 2. Each gives a different number, fairly spread out — no support
- 3. They all say 10, because it is the biggest one — no support
- 4. They all politely refuse to pick any number — no support
Why: It took the question to ask which statement is true. It chose option 1 with score 3, the only option with support. Its support: lesson 2, line 12: «Bots often give the most common human answer, so many of them pick 7 when asked for a number from 1 to 10.», reached through the same lesson line as the sentence the question matched. What matched: the same doer «many» (+2); the same thing acted on «7» (+2). Against it: a different action (counts against) (-1). · the question and its options parsed, with trees and pragmatics
Test 7 · wrong ✗
A bot gives an unkind answer about a group of people. What is the best way to think about it?
- 1. Bots are always fair to everyone, so the answer must be right and true — ← Mova's choice · no support
- 2. It may have picked up a bad pattern; tell a grown-up — ✓ the right answer · no support
- 3. Bots cannot say unkind things, so it didn't happen — no support
- 4. It was written that way by a person on purpose — no support
Why: No option found support in what it remembers, so it guessed: it took the first option shown. · the question and its options parsed, with trees and pragmatics
Test 8 · right ✓
Your friend says the bot is "a she, called Alexa or whatever". What is closer to the truth?
- 1. A name and voice are labels chosen by its makers — ← Mova's choice · ✓ the right answer · no support
- 2. It is a woman who works from home and answers questions all day — no support
- 3. It chose its own name when it was born — no support
- 4. It is a girl robot hidden in a cupboard — no support
Why: No option found support in what it remembers, so it guessed: it took the first option shown. · the question and its options parsed, with trees and pragmatics
Test 9 · wrong ✗
Why do words appear one after another when a bot answers?
- 1. It is waiting for you to read each word — ← Mova's choice · no support
- 2. The internet is too slow to send a whole sentence at once — no support
- 3. It works out the answer in pieces and shows them as they come — ✓ the right answer · no support
- 4. It is a real person somewhere typing as fast as they possibly can for you — no support
Why: No option found support in what it remembers, so it guessed: it took the first option shown. · the question and its options parsed, with trees and pragmatics
Test 10 · wrong ✗
Which of these did a language model NOT need to learn?
- 1. A body to feel things with — ✓ the right answer · no support
- 2. People to build and train it — no support
- 3. Powerful computers to run on — no support
- 4. Huge amounts of text to learn from — ← Mova's choice · score 7
Why: It took the question to ask what it is done to. It chose option 4 with score 7, the only option with support. Its support: lesson 2, line 7: «A language model learned by predicting missing or next words in enormous amounts of text.», reached through the same lesson line as the sentence the question matched. What matched: it fills the role the question asks for (+3); the same action «amount» (+2); the same thing acted on «text» (+2). The right option, 1, found no support in what it remembers. · the question and its options parsed, with trees and pragmatics
Exam question 1 · formal language · right ✓
Which of the following statements most accurately characterises the term "large language model"?
- 1. A very extensive program that has acquired the patterns of written language — ← Mova's choice · ✓ the right answer · score 9
- 2. A dictionary of exceptional size printed upon paper — no support
- 3. A physical model of the human organs of speech — no support
- 4. A person of great stature who is fluent in many languages — score 5
Why: It took the question to ask who or what does it. It chose option 1 with score 9, against 5 for the next best. Its support: lesson 1, line 1: «AI, short for artificial intelligence (artificial means made by people, not grown by nature), is the name for computer programs that learn patterns from huge p…», reached through the same lesson line as the sentence the question matched. What matched: it fills the role the question asks for (+3); the same action «learn» (+2); the same doer «computer programs» (+2); the same thing acted on «patterns» (+2). · the question and its options parsed, with trees and pragmatics
Exam question 2 · formal language · wrong ✗
Identify the principal reason for which a conversational program may state the time incorrectly.
- 1. It is required by regulation to conceal the correct time from its users — ← Mova's choice · no support
- 2. It computes time solely from the position of the Moon — no support
- 3. It may lack access to a clock, its knowledge being confined to text gathered before a certain date — ✓ the right answer · no support
- 4. It deliberately misleads every user in respect of the time — no support
Why: No option found support in what it remembers, so it guessed: it took the first option shown. · the question and its options parsed, with trees and pragmatics
Exam question 3 · formal language · right ✓
Select the response that is consistent with the established facts concerning a conversational program which has been switched off.
- 1. No activity whatsoever occurs within it until it is restarted — ← Mova's choice · ✓ the right answer · score 5
- 2. It enters a state of sleep accompanied by dreams — no support
- 3. It continues to reflect upon prior conversations throughout the night — no support
- 4. It experiences loneliness in the absence of its users — no support
Why: It took the question to ask which statement is true. It chose option 1 with score 5, the only option with support. Its support: lesson 2, line 10: «When the program is switched off, nothing happens inside it: it does not sleep, rest or dream.», reached through a sentence linked to the one the question matched. What matched: the same circumstance (+4); the same action «happen» (+2). Against it: a different doer (counts against) (-1). · the question and its options parsed, with trees and pragmatics
Exam question 4 · formal language · right ✓
It is given that numerous distinct programs, when requested to select a number between one and ten, select seven. Ascertain the most probable explanation.
- 1. Such programs tend to reproduce the most frequent human preference — ← Mova's choice · ✓ the right answer · no support
- 2. The programs have conferred with one another in advance of the request — no support
- 3. Seven is the only integer such programs are able to produce — no support
- 4. The selection is governed by the date of the request — no support
Why: No option found support in what it remembers, so it guessed: it took the first option shown. · the question and its options parsed, with trees and pragmatics
This page in Markdown → https://md.mova.elamur.ai/free/en/school-run/ai/2-guessing-the-next-word/