Review · module 101 of 155
Review: how a language model works
Recall the essentials of tokens, variation, context and hallucinations in one overview.
After this lesson
You explain in your own words how a language model produces an answer and why that determines what you need to check yourself.
In view · How it connects
Three insights about language models
The model predicts a continuation in tokens, based on patterns and your context.
What is not in the context window cannot be used by the model in that answer.
A fluent answer can contain made-up details. Check the important claims.
Part 1 of 3 · Text in small pieces
Your preference for less movement is on. Choose each part yourself.
A model predicts text in small pieces
A language model processes text as tokens: pieces of words, punctuation marks or spaces. Each time, it predicts a likely continuation based on patterns from its training and the text you give it (lessons 1 and 31). That is why an answer can sound fluent and convincing without the model having looked up a single fact. The setting for variation, often called temperature, determines how predictable that continuation is. More variation helps with ideas. Less variation makes answers more consistent, but not automatically more accurate (lesson 32).
What the model sees determines what it can use
For every answer, the model works with a limited context window: fixed instructions, the conversation, your files and the space for the answer (lessons 33 and 34). Whatever is not in it, or has dropped out of it, the model cannot consult for that answer. Fixed instructions and memory features help, but they do not guarantee that every earlier agreement is taken into account (lesson 61). So give important facts again. A document you supply remains content: it does not replace an assistant’s fixed rules or permissions.
Convincing is not the same as correct
A model can fill in details that seem plausible: a date, an amount or a source with a title and author (lesson 35). So split an answer into checkable claims and look up the original source. Do not ask the same AI to confirm itself. Also look at where the computation happens. In the cloud, your input goes to an external server; locally, the model runs on your own device. A local model does not in itself mean that the whole application works offline: search, storage and synchronisation can remain external (lesson 20).
New practice example
From explanation to a usable result.
A language model predicts text. It works with what is in its training and in your current context, and it can easily make something up.
Fictional practice situationFictional: after the course, a colleague makes a cheat sheet. Among other things it says: “A token is a word”, “With temperature zero the answer is correct” and “Local is always offline”.
View a possible solution
This is how it can be done
All three statements need to be corrected. A token is a piece of text; how many tokens a word counts as depends on the model. A low temperature makes the answer more predictable, not more correct. A local model can sit inside an application that still searches online or synchronises.
Why this works
Reviewing works best when you have to explain a correction. That way you notice which concept you really master and which lesson you could reread.
Make it slightly more difficult
Write two correct statements and one incorrect statement about the context window yourself. Have someone else point out the incorrect one.
Extra practice, at your own pace. You do not need to submit anything. This does not count towards an automatic assessment.
Your workplace · learning by doing
Make something. See what changes.
Checking the AI connection… You can already view the written example.
Present your assignment to Ende.
Read the answer, change one thing, and compare.
Previously held
New version
You remain in control
What do you want to improve in your next attempt?
On this device only. AI answers remain only in this page memory. This is self-review, not a grade.
Why this approach works
Reviewing works best when you have to explain a correction. That way you notice which concept you really master and which lesson you could reread.
A step further
Write two correct statements and one incorrect statement about the context window yourself. Have someone else point out the incorrect one.
Try & discover · no obligation
Become the source checker yourself
Only mark statements that the source does not support. Then check your selection.
Fictional cheat sheet from a colleague after the basic course. Which statements does she need to correct before she shares it?
Select which sentences you would correct above. The answer remains hidden until you check your selection.
This trial works locally with fictitious data and does not count towards your lesson completion. It starts anew with a new lesson or reload.
A prompt to get you started
I am reviewing the basics of language models. Ask me five short questions, one at a time, about tokens, variation, the context window, hallucinations and cloud versus local. Wait for my answer after each question. Then say what is correct and what is missing, in no more than three sentences. Do not give all the answers straight away and do not invent sources.
Apply it
A colleague wants sources for a presentation. The AI gives five titles with authors and years. They sound convincing, but no one has opened them yet.
Which approach best fits this situation?
- Use the list, because a language model knows the literature.
- Set the temperature lower and ask for the list again; then the sources will be correct.
- Look up each source in a library or search service and only use what you actually find.
With JavaScript, you can check your approach and receive immediate feedback.
Test your insight
A colleague says: “I set the temperature to zero, so the answer is correct.” What do you reply?
- Right: without variation, a model makes no mistakes.
- A low setting makes answers more predictable, but you still have to check the facts.
- A low setting makes the model look up sources itself.
Extra practice
Without help, explain in three sentences why a fluent answer can be wrong. Then compare your explanation with the three blocks of this lesson and add what you missed.
A model predicts text in tokens based on patterns and the context it receives. It does not look up facts on its own, so a date or source may be made up. That is why I check important claims in the original source and supply the necessary facts myself.
