Lecture 1-2: How a Language Model Makes Words
Contents 31
- Lecture 1-1: What the Word "AI" Points To
- Lecture 1-2: How a Language Model Makes Words
- Lecture 1-3: Sounding Right Is Not Knowing
- Lecture 1-4: Reading a Conversation Closely
- Lecture 1-5: Practice Lab, Read an AI Answer
- Lecture 2-1: Invented Facts
- Lecture 2-2: Old Knowledge, Missing Sources, Uneven Coverage
- Lecture 2-3: Scripture and Theology, Where to Be Most Careful
- Lecture 2-4: Building a Checking Plan
- Lecture 2-5: Practice Lab, Find the Errors
- Lecture 3-1: Asking Well
- Lecture 3-2: Checking What You Are Given
- Lecture 3-3: What Must Never Be Typed In
- Lecture 3-4: Keeping Notes and Giving Credit
- Lecture 3-5: Practice Lab, Mend the Prompts, Mark the Secrets
- Lecture 4-1: Hand It Over, Help With It, Keep It
- Lecture 4-2: Office Work and Communication
- Lecture 4-3: Teaching and Study Materials
- Lecture 4-4: Where Teachers and Leaders Draw the Line
- Lecture 4-5: Practice Lab: Hand Over, Help, or Keep
- Lecture 5-1: Preaching and AI: A Theological Question
- Lecture 5-2: Care and Counselling
- Lecture 5-3: When Members and Young People Use AI
- Lecture 5-4: The Minister's Own Life
- Lecture 5-5: Practice Lab: Judge the Scene, Write a Position Paragraph
When a chat assistant answers you in good sentences, it is natural to picture a clever person on the other side of the screen who has looked into the matter. That picture is wrong, and a minister who holds it will trust the wrong things. This lecture offers a better picture, in plain words and without mathematics.
The picture is an analogy, and every analogy fails somewhere. We use this one because it predicts the right kind of mistakes, and we say at the end where it breaks down.
What these 16 minutes give you
- Explain in everyday words how a language model learns which words tend to follow which.
- Distinguish training, which happens before you arrive, from use, which happens while you type.
- Explain why the same question can produce different answers.
- Explain why a long conversation can lose its early parts.
- Say where the "next word" analogy helps and where it fails.
It is the glory of God to conceal a thing, but the glory of kings is to search out a matter." — Proverbs 25:2
World English Bible
He who gives answer before he hears, that is folly and shame to him." — Proverbs 18:13
WEB
The first verse treats searching out a matter as honourable work. The second names the opposite: an answer given before the matter has been heard. A language model does not search out a matter in the sense of that verse. We will see why.
1. A phone that guesses your next word
Most people have used a phone keyboard that suggests the next word. Type "Thank you for" and it offers "your," "the" or "coming." The phone has not understood your message. It has noticed that in the writing it was built from, certain words tend to follow certain others.
Now imagine that device made enormously larger. Instead of looking at the last two words, it looks at your whole request and everything written so far in the conversation. Instead of being built from some short samples, it was built from a vast body of text: books, articles, web pages, forum posts, and much else. And instead of suggesting a word for you to choose, it chooses the word itself, adds it, and goes on to the next, until the answer is finished.
That is the central picture. A language model is a program that continues writing in the way that its training makes most fitting. Everything it says to you is built up one small piece at a time.
2. Training and use are two different times
Training is the long process before the model ever reaches you. During training the system is shown a very large amount of text. It repeatedly tries to guess what comes next, is told how far off it was, and adjusts itself a little. After an enormous number of rounds, it has become very good at producing text that fits the patterns of the writing it was shown. What it keeps from training is not a library of pages. It is a very large set of adjusted numbers that capture patterns. Nobody can open it and point to the place where a fact is stored.
Use is what happens when you type. You give it words (called the prompt) and it continues. At that moment it is not reading the world, not consulting an encyclopaedia, and not remembering the sermon you heard last Sunday. It is using the patterns it already has.
Two consequences follow.
First, what the model "knows" about the world stops at the point where its training text stopped. We come back to this in Week 2.
Second, using it does not teach it in the way that a conversation teaches a student. Within one conversation it can use what you have said earlier. Whether a service keeps your words and uses them later is a different matter that depends on the service and its terms. That is one reason Lecture 3-3 is about what must never be typed in.
3. Why it picks one word and not another
At each step, many words could follow. The model weighs them. Some are very likely, some are possible, and most are unlikely. It then picks, and the pick has some chance in it. It does not always choose the single most likely word. If it did, its writing would be flat and repetitive.
(An invented example.) Suppose the words so far are "The shepherd lifted the lamb onto his." Very likely next: "shoulders." Possible: "arm," "back." Unlikely: "bicycle." The model is far more likely to say "shoulders" than "bicycle," but it is not forced to say "shoulders" every time.
This is why you can ask the same question twice and receive two different answers. It is not changing its mind. It is making a chain of weighted picks, and the chain can run differently each time. A person who asks a chat assistant for a short prayer, then asks again, will not be given the same words. For a first draft of a notice that variety is useful. For a question of fact it is a warning: if two runs disagree, at least one is wrong, and agreeing runs are no proof, because the pattern may simply be a common one.
4. It is built from pieces, so it can build things that never existed
Because the model assembles text out of patterns, it can write a sentence that no one ever wrote. It can write a poem in the style of a hymn, or a letter to a new member, or a summary of a paragraph you gave it. This is the source of its usefulness.
It is also the source of its main danger. A pattern-follower will produce something that fits the pattern of a book title, a Bible reference, a court judgment or a quotation, even when no such thing exists. We take that up in Week 2.
5. A limited memory for a long conversation
A model can use only so much text at one time. The conversation so far (your messages and its replies) has to fit within a limit. When a conversation grows long, the earliest parts may be dropped or blur, and the model may stop following something you said at the start. You may have told it at the top, "Write for teenagers, in plain English." Forty messages later, the answers drift back to a different style.
Practical consequence: if something matters, say it again, near the question. And do not assume that "it remembers what I told it earlier" for a long exchange. The size of the limit differs from one service to another and changes over time, so this course gives no figure.
Not a recording
A chat assistant does not keep a perfect record of your conversation. It uses what is within its limit at that moment. Treat it as a reader who may have lost the first pages.
6. What the analogy shows, and where it fails
The next-word picture is useful because it explains the kind of errors you will meet. A pattern-follower is good at what is common and fluent. It is weak at what is rare, recent, exact or contested, and it does not know when it has crossed from the first to the second. It can give you a smooth answer to a question that has no good answer.
But the picture also fails in three ways. Do not stretch it.
It makes the system sound like a trick. The better the training and the larger the system, the more it can do: translate, summarise, answer follow-up questions, follow detailed instructions, write computer code. These abilities are real. To say "it only guesses the next word" is not to say that it can do little. It is to say how it does what it does.
It sounds as if nothing is going on inside. Researchers are still studying what happens inside these systems, and plain statements about what they "really" do are often too confident, both the grand ones and the dismissive ones. This course makes no claim about inner life. It makes a claim about reliability: whatever is going on, fluent output is not evidence of truth.
A real service is more than a bare model. A chat service may add instructions you cannot see, may connect the model to a search tool, may filter some replies, and may update its model from time to time. Lecture 1-4 takes this up. So the next-word picture describes the core of the tool, not every feature of every service.
7. What this means for a minister
You do not need to remember how it works in detail. You need three habits.
- Expect fluency without verification. A well-made sentence does not show that a fact has been checked.
- Expect variety. If an answer matters, ask again or ask a different way, and compare.
- Say it again. In a long exchange, repeat what the answer must keep to.
Diagram 1. From training to a finished answer
Diagram 2. Words you now have
| Word | Plain meaning |
|---|---|
| Language model | A program trained on a vast body of text to continue writing |
| Training | The long process, before you arrive, that adjusts the system to the patterns of that text |
| Prompt | The words you give it, plus the conversation so far |
| Response | What it writes back, one small piece at a time |
| Limit of the conversation | The amount of text it can use at one moment |
In short
- A language model is trained on a vast body of text to continue writing in the way that fits the patterns in it.
- Training happens before you arrive. Use happens when you type a prompt.
- It builds an answer one piece at a time, choosing among likely words with some chance. So the same question can give different answers.
- It does not look a fact up. Nobody can point to the place inside it where a fact is stored.
- Only so much of a long conversation can be used at once. Say again what matters.
- The next-word picture explains the kind of error to expect. It fails if it is used to say the tool is trivial, or to say what is "really" inside it.
- Whether a service keeps what you type is a matter of that service's terms, not of the model's nature.
| Front | Back |
|---|---|
| Language model | A program trained on a vast body of text to continue writing in a way that fits the patterns of that text. |
| Training and use | Training is the long process before you arrive; use is what happens while you type. |
| Why answers differ | At each step it chooses among likely words with some chance, so a chain of choices can run differently each time. |
| Limit of memory | Only so much of a long conversation fits at once, so early parts can drop out. |
| Where the analogy fails | It can make the tool sound trivial, and it says nothing certain about what is inside. |
One sentence to remember A language model continues words in a way that fits its training; it does not look anything up, and a smooth answer is not a checked one.
This week's action Notice the word suggestions on a phone for one day and write down two that were right and two that were wrong. No AI tool is needed.
Check yourself
Not "did you read it" but "can you explain it". Ticks are kept on this device only.
Reflection and discussion
- Think of the suggestion on a phone keyboard. What does it get right, and what does it get wrong about you?
- If you asked a chat assistant the same question twice and received different answers, what would you do next?
- Proverbs 18:13 speaks of an answer given before hearing. In what way is a fast, fluent answer from a model like that verse, and in what way is it unlike it?
- What would you say to a colleague who says, "It is only guessing words, so it is useless"? What would you say to one who says, "It has read everything, so it must be right"?
- What might you repeat near the end of a long conversation so that the model keeps to it?