Vibeleaderboard
← Course overviewModule 01 · Understand AI · Beta
Lesson 1.6

Module contents

Understand AI

Lesson 6 of 6

What does the application add?

Two applications can use the same model and still behave differently, because each decides what information reaches the model and what the system can do beyond generating text.

Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.

The visible conversation is not the whole system

In lesson one, software meant instructions a computer carries out. The application is software that receives your request, prepares the context for the model and handles what comes back. The chat screen shows part of that process, and only part.

The application can supply instructions you never typed. An organization might configure its assistant to answer only from supplied documents; another might set it up for open-ended writing. With the same model, those different instructions and different supplied information can produce different behavior. The model’s context, from lesson three, is whatever the application actually supplied, which may include things you cannot see and exclude things you can. 15

Remembering information is an application capability

Imagine you tell a fictional garden assistant that you prefer short replies. The application could save that preference and include it in the context of future requests. The model would then receive the preference again each time. No change to its learned values is needed for this example; the application is doing the remembering.

Real memory features vary. OpenAI’s documentation, for example, describes using information from earlier chats and other connected material to personalize ChatGPT when memory is enabled, with controls for the user. That describes one product’s feature. It does not mean every application remembers information, remembers everything, or handles deletion and later training in the same way. 13

This gives you a useful question when a system surprises you: did the model change, or did the information supplied to it change? A stale saved preference calls for a different fix from a weakness in the model, and the answer tells you where to look.

Three fictional applications, one model

Here is an authored comparison. Three fictional applications all use the same fictional model, Oak.

  • Sprout drafts a reply from whatever text you paste. It supplies nothing else.
  • Trellis first looks up the organizer’s current notice in an approved document store and supplies it to Oak before drafting.
  • Harrow can, after you approve a draft, post the reply to the garden’s shared message board.

Ask all three: “Has the organizer cancelled Saturday’s session?” Sprout's context holds only the old note. It can explain the rain condition or ask for a current notice, but the old note cannot confirm whether the organizer has since cancelled the session. Trellis can answer from the current notice, because its software retrieved it and put it in the context. That answer is only as good as the retrieved document: fetching it alone does not show that it is the right notice or that the reply interprets it correctly. Retrieval and search bring evidence into the context; the resulting claims still need to match that evidence. 14

Now say “Post the reply.” Producing the words of a message is different from delivering it. Sprout and Trellis can only draft; neither has a connection to the message board. Harrow has that connection, the permission to use it, and a rule that you approve first. Even then, a generated sentence saying “Posted” is a claim, not proof that the action happened. A confirmation from the board software is a different kind of evidence, and it is the kind you would want.

Line the three up and the pattern is clear. The model generated words in every case. What differed was what the application supplied before generation and what it was able to do afterward. Later modules examine how software takes such actions and how to know they succeeded.

Drafting and sending are different operations

  1. 1 →GenerateProduce the message text
  2. 2 →ActSoftware sends through an authorized connection
  3. 03CheckInspect the result from that service
The model producing the word “sent” is not, by itself, evidence that a message was delivered.

Transfer task: a choir rehearsal reply

Try the module’s distinctions on a situation that is not the garden. Everything in this task is an authored illustration.

You help run a community choir. In this fictional situation, the model finished training months ago. The organizer decided on a new venue this morning and wrote the following note, which is all you have:

Thursday: rehearsal moved to the school hall, 7 p.m. Bring your music folder. No rehearsal next week.

You paste it into a fictional chat application and ask for a friendly reminder for members. The application shows:

Rehearsal is on Thursday at 7 p.m. in the school hall. Music folders will be handed out at the door, and next week we are back at the usual time.

Write your own explanation before reading the guide below. Cover four things. First, which sentences or phrases are supported by the note, and which are unsupported or contradict it? Second, what did the application do here, and what did the model do? Third, what is the difference between the model’s learned values and the note, and why could this morning’s change of venue not already be in the learned values? Fourth, for each addition, what kind of check would settle it?

Self-review guide. A complete explanation identifies “Thursday,” “7 p.m.,” and “school hall” as supported by the note. It identifies “music folders will be handed out at the door” as unsupported: the note tells members to bring folders, and nothing says spares exist, so the draft has changed an instruction into a promise, the same shift as the gloves in lesson four. It identifies “next week we are back at the usual time” as contradicting the note, which says there is no rehearsal next week; the note directly gives a different schedule. A complete explanation says the application accepted the request, supplied the note as context and displayed the result, while the model generated the wording, and that the venue change reached the reply only because the note was in the context; the model’s learned values were fixed before this note existed. For checks, it names the choir organizer or the current schedule as the source that could settle the folder question, and it points out that reassurance alone would not settle either question. A recheck needs to use the note or obtain appropriate evidence. Naming these terms without saying why each one applies is not yet an explanation.

This task gives you practice and a way to notice gaps. Explaining the distinctions here is practice in reasoning about an AI response; operating real systems involves further skills taught later.

Carry the foundation into the rest of the course

You can now take apart something people bundle together as “the AI.” Training set the model’s learned values from past data. A request supplies context. The model generates a reply one token at a time from both, and that reply can vary. The application decides what goes into the context and what the system can do around the model. Checking a claim means matching it to the right evidence.

To understand software that takes action, you need to know what software can read, change and run: files, programs, saved state and requests between computers. That is the next module. After it, the course explains agents, which use models to carry out work through a sequence of actions, and harnesses, the software that manages those actions and makes tools available. Understanding both comes before operating either, and nothing needs installing to follow the explanations.

Intel is the site’s collection of reports, talks and other material about AI. The editorial summary of a talk by Andrej Karpathy about building with agents stresses keeping your own understanding when delegating work. That is the question this module has trained you to ask: which part of the system did the work, and what shows that it succeeded? You can also glance at harnesses in the tools directory. For now, treat them as software environments with different capabilities; later modules supply the criteria for choosing one.

Carry this forward

  • An application can add instructions, saved information, retrieved documents and connections to other software without retraining the model.
  • Applications using the same model can behave differently because they supply different context and can take different actions.
  • Generating a proposed action, executing it and confirming its result are separate steps with different evidence.
  • Understanding software that acts requires the files, programs, saved state and requests taught next.

Lesson check

Two fictional applications use the same model, trained months ago. Sprout drafts from the text you paste. Trellis also retrieves a cancellation notice that the organizer first wrote today. You ask both whether Saturday is cancelled, and only Trellis mentions the cancellation. Which explanation fits?

Choose an answer for each question.

This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.

You have reached the end of the first module. The module quiz below checks how these ideas fit together.

Module 1 quiz: put the ideas together

1. An organization releases a new model, but your application still lists its previous version. What should you conclude?
2. In a simplified training process, predictions are compared with known examples and internal numerical values are adjusted. Which part is training?
3. The model version was trained months ago. You ask about a notice first written this morning, but the app has no search access and you have not supplied the notice. Which statement is justified?
4. A first reply omitted the rain condition. You retry with exactly the same supplied information and the next reply includes it. What does that prove?
5. An application remembers your preference for brief answers in a later conversation. Which mechanism could explain this without retraining the model?
6. A fictional library note says “Closed Monday; opens Tuesday at noon.” A draft says “Open Monday and Tuesday at noon.” What needs correction?
7. After a very long conversation, an assistant stops following an early instruction. Which explanation fits what you learned?

Choose an answer for each question.

This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.

Sources · Reviewed 2026-09-11
  1. 5. Google: Language models and tokens
  2. 13. OpenAI: Memory explained
  3. 14. OpenAI: Search in ChatGPT
  4. 15. Anthropic: what enters the context window

Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.