We Do · Could This Class Run Its Own AI?

Frozen Numbers, Moving Data

You now know something most adults do not. The model is frozen. Your prompt does not change it. That is a precise technical fact, and precise facts are useful mainly because they let you take a fuzzy worry and find out which part of it is actually true.

Here is a fuzzy worry worth testing. Americans are increasingly uneasy about artificial intelligence. In a survey conducted in June 2026, 52 percent of U.S. adults said they were more concerned than excited about the growing use of AI in daily life, compared with 9 percent who said they were more excited than concerned (Pew Research Center, 2026b). That gap has widened over time. In 2021 the concerned share was 37 percent. The shift among younger Americans is sharper still: 55 percent of adults under 30 now report more concern than excitement, up from 39 percent only two years earlier (Pew Research Center, 2026b).

One specific worry shows up again and again inside that general unease, and it is a worry about data. Roughly seven in ten Americans predict that AI will make their personal information less secure, while just 3 percent expect it to make their information more secure (Pew Research Center, 2026a). Confidence in the institutions meant to manage this is low in both directions: 67 percent report little or no confidence in the federal government to regulate AI effectively, and about six in ten are not confident that the companies building these systems are developing them responsibly (Pew Research Center, 2026a).

Now put today’s lesson against that worry and sort it.

If the concern is that your words are being absorbed into the model, becoming part of what it knows, retrievable later by some stranger who asks the right question, then the concern is aimed at the wrong stage. Inference does not write anything back. The parameters were frozen before the file reached the server. Your sentence does not become a parameter. It cannot.

But notice how narrow that reassurance actually is. It says something about the numbers. It says nothing whatsoever about the trip your sentence took to reach them.

When you type into a cloud model, your text leaves your device. It crosses a network, arrives at a company’s servers, and is processed there. Along that path it can be logged. It can be stored. It can be retained under a policy you did not read, reviewed by an employee, handed over under a legal request, or collected into a pile of text that gets used months later as training material for an entirely different model, in a separate run, producing a separate frozen file. None of that is the model learning from you. All of it is your data going somewhere.

So the public concern is not wrong. It is misfiled. The risk lives in transmission and storage, not in the frozen numbers. And that distinction is not a technicality, because the two problems have completely different solutions. You cannot fix a data retention problem by asking a model to forget, since the model never remembered. You fix it by controlling where the text goes.

Which is exactly the choice built into this course. Two of the three models you will use next week run on a machine in Mr. Muggivan’s house, reached through a tunnel. Your prompt goes there and nowhere else, and it is not retained. The third model, Gemini, runs on Google’s servers, and your prompt makes the full trip. Same question, same moment, three answers, and one of the three took a fundamentally different path to get to you.

That is the comparison you will be running, and it is worth knowing in advance that most of the class will focus on which answer sounded best. Sounding best is the easiest thing to notice and the least informative thing to measure.

Return to the discrimination this course keeps circling. A claim can be true and still fail to answer the question in front of it. "The model does not learn from you" is true. If someone asks whether their conversations are private and you offer them that sentence, you have handed them something accurate that sits beside their question rather than answering it. Knowing the difference between the two is most of the skill.

References

Pew Research Center. (2026, June 17). Americans and AI 2026: Chatbots, smart devices and views on impact. https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/

Pew Research Center. (2026, August 18). Young adults in the U.S. are increasingly wary of AI, concerned it will take jobs. https://www.pewresearch.org/short-reads/2026/08/18/young-adults-in-the-us-are-increasingly-wary-of-ai-concerned-it-will-take-jobs/

The diagrams · 1 of 4

The words, in plain terms

parameter

One number inside the model.

Something like 0.0412. Pull one out and it tells you nothing. Whatever the model knows is spread across billions of them at once.

In a gameNot a setting in the options menu. Closer to one of the millions of numbers a physics engine multiplies without any of them meaning anything on its own.

The largest model on the rig has roughly twenty seven billion.

pretraining

The long stage where the numbers get nudged, over and over, until they predict text well.

Weeks to months on machines that cost more than most buildings. Nobody writes facts in. The facts are a side effect of getting good at predicting what comes next.

In a gameThe long grind before the game ships, not anything that happens while you play.

post-training

A shorter stage afterwards that shapes how the model answers.

Pretraining decides roughly what it knows. This decides whether it answers in full sentences, follows an instruction, or refuses.

In a gameThe patch that changes how something behaves without rebuilding the game.

freeze

The moment the numbers are written to a file and stop changing.

From here the model does not get smarter, does not get worse, and does not update overnight. It stays what it is until a company runs the whole process again.

In a gameThe moment a build goes gold. After that the disc is the disc.

inference

What happens when you type a prompt: your text goes through the frozen numbers and an answer comes out.

Nothing inside changes while it happens. The numbers you use are bit for bit the same as the ones used by the person before you.

In a gamePlaying the game, as opposed to building it.

knowledge cutoff

The date the text it read stopped.

Ask about anything after it and the model may not know that it does not know, so it produces a confident, well-formed, invented answer.

In a gameThe version the strategy guide was written for. Ask it about a later patch and it answers anyway.

Four Stages, One Question

Stage	What happens	How long	Do the numbers change?
Pretraining	The model reads enormous amounts of text and guesses what comes next. Every wrong guess nudges the numbers.	Weeks to months	Yes. Constantly. This is where knowledge comes from.
Post-training	A smaller, targeted stage. Teaches the model to follow instructions, answer in full sentences, and refuse some requests.	Days to weeks	Yes, but far fewer of them, and aimed at behavior instead of knowledge.
Freezing	The numbers are written into a file and locked. The file gets copied to servers and to machines like the one in Mr. Muggivan’s house.	A moment	No. This is the moment it stops.
Inference	You type a prompt. It runs through the frozen numbers. An answer comes out.	Seconds	No. Not once. Not ever.

Two things follow from the bottom two rows. The model has a knowledge cutoff, because the reading stopped on a date. The model is not learning from you, because the learning was over before the file reached the machine you are typing on.

The diagrams · 2 of 4

The words, in plain terms

parameter

One number inside the model.

Something like 0.0412. Pull one out and it tells you nothing. Whatever the model knows is spread across billions of them at once.

In a gameNot a setting in the options menu. Closer to one of the millions of numbers a physics engine multiplies without any of them meaning anything on its own.

The largest model on the rig has roughly twenty seven billion.

pretraining

The long stage where the numbers get nudged, over and over, until they predict text well.

Weeks to months on machines that cost more than most buildings. Nobody writes facts in. The facts are a side effect of getting good at predicting what comes next.

In a gameThe long grind before the game ships, not anything that happens while you play.

post-training

A shorter stage afterwards that shapes how the model answers.

Pretraining decides roughly what it knows. This decides whether it answers in full sentences, follows an instruction, or refuses.

In a gameThe patch that changes how something behaves without rebuilding the game.

freeze

The moment the numbers are written to a file and stop changing.

From here the model does not get smarter, does not get worse, and does not update overnight. It stays what it is until a company runs the whole process again.

In a gameThe moment a build goes gold. After that the disc is the disc.

inference

What happens when you type a prompt: your text goes through the frozen numbers and an answer comes out.

Nothing inside changes while it happens. The numbers you use are bit for bit the same as the ones used by the person before you.

In a gamePlaying the game, as opposed to building it.

knowledge cutoff

The date the text it read stopped.

Ask about anything after it and the model may not know that it does not know, so it produces a confident, well-formed, invented answer.

In a gameThe version the strategy guide was written for. Ask it about a later patch and it answers anyway.

Three Models, Real Numbers

Model	Total parameters	Active per token	File size at Q4	Where it runs	Knowledge cutoff
Gemma 4 E4B	8,000,000,000	4,500,000,000 effective	6.33 GB	Mr. Muggivan’s rig	January 2025
Gemma 4 26B-A4B	26,000,000,000	4,000,000,000	18.2 GB	Mr. Muggivan’s rig	January 2025
Gemini 3.7 Flash	Not published	Not published	Not published	Google’s servers	March 2026

These are the three models you will use on Monday. Two of them sit on a machine you could walk up to and touch. One of them does not.

The diagrams · 3 of 4

The words, in plain terms

parameter

One number inside the model.

Something like 0.0412. Pull one out and it tells you nothing. Whatever the model knows is spread across billions of them at once.

In a gameNot a setting in the options menu. Closer to one of the millions of numbers a physics engine multiplies without any of them meaning anything on its own.

The largest model on the rig has roughly twenty seven billion.

pretraining

The long stage where the numbers get nudged, over and over, until they predict text well.

Weeks to months on machines that cost more than most buildings. Nobody writes facts in. The facts are a side effect of getting good at predicting what comes next.

In a gameThe long grind before the game ships, not anything that happens while you play.

post-training

A shorter stage afterwards that shapes how the model answers.

Pretraining decides roughly what it knows. This decides whether it answers in full sentences, follows an instruction, or refuses.

In a gameThe patch that changes how something behaves without rebuilding the game.

freeze

The moment the numbers are written to a file and stop changing.

From here the model does not get smarter, does not get worse, and does not update overnight. It stays what it is until a company runs the whole process again.

In a gameThe moment a build goes gold. After that the disc is the disc.

inference

What happens when you type a prompt: your text goes through the frozen numbers and an answer comes out.

Nothing inside changes while it happens. The numbers you use are bit for bit the same as the ones used by the person before you.

In a gamePlaying the game, as opposed to building it.

knowledge cutoff

The date the text it read stopped.

Ask about anything after it and the model may not know that it does not know, so it produces a confident, well-formed, invented answer.

In a gameThe version the strategy guide was written for. Ask it about a later patch and it answers anyway.

The life of a model — pretraining and post-training under one arrow labelled numbers changing, and that arrow stops at the freeze. Everything to the right of the freeze is inference.

On the formschematic 11

The diagrams · 4 of 4

The words, in plain terms

parameter

One number inside the model.

Something like 0.0412. Pull one out and it tells you nothing. Whatever the model knows is spread across billions of them at once.

In a gameNot a setting in the options menu. Closer to one of the millions of numbers a physics engine multiplies without any of them meaning anything on its own.

The largest model on the rig has roughly twenty seven billion.

pretraining

The long stage where the numbers get nudged, over and over, until they predict text well.

Weeks to months on machines that cost more than most buildings. Nobody writes facts in. The facts are a side effect of getting good at predicting what comes next.

In a gameThe long grind before the game ships, not anything that happens while you play.

post-training

A shorter stage afterwards that shapes how the model answers.

Pretraining decides roughly what it knows. This decides whether it answers in full sentences, follows an instruction, or refuses.

In a gameThe patch that changes how something behaves without rebuilding the game.

freeze

The moment the numbers are written to a file and stop changing.

From here the model does not get smarter, does not get worse, and does not update overnight. It stays what it is until a company runs the whole process again.

In a gameThe moment a build goes gold. After that the disc is the disc.

inference

What happens when you type a prompt: your text goes through the frozen numbers and an answer comes out.

Nothing inside changes while it happens. The numbers you use are bit for bit the same as the ones used by the person before you.

In a gamePlaying the game, as opposed to building it.

knowledge cutoff

The date the text it read stopped.

Ask about anything after it and the model may not know that it does not know, so it produces a confident, well-formed, invented answer.

In a gameThe version the strategy guide was written for. Ask it about a later patch and it answers anyway.

The arithmetic — four numbered steps from a prompt to one predicted word. Multiply, add, squeeze, score. A response is not written; it is one word at a time.

On the formschematic 14

View

The classwork

Could this class run its own AI? Day 5 of 6. Section 1 goes with the reading on your screen. Section 2 goes with the two pictures and the table. All of it is one activity with one deadline. Finish it in class and there is no homework. Station view is read-only — it is the backup for a form that will not load. Switch back to My view to answer for a grade.

Question 1multiple choice

Which period are you in?

APeriod 1
BPeriod 2
CPeriod 4

Question 2multiple choice

Which stage is the model in when you type a prompt to it?

APretraining
BInference
CFreezing
DPost-training

Question 3multiple choice

A model gives a confident answer about an event from last week, and the answer is invented. Which term explains this best?

APost-training
BParameter
CKnowledge cutoff
DInference

Question 4multiple choice

Two students in different periods send the exact same prompt to the same model. Are they using the same numbers?

AYes, identical
BNo, the model updated between periods
COnly if they used the same station
DThere is no way to tell

Question 5multiple choice

Which stage would you change if you wanted the model to stop answering rudely, without changing what it knows?

AInference
BFreezing
CPretraining
DPost-training

Question 6written

In your own words, what is a parameter?

Paragraph.

Question 7written

The reading says the public concern about data is misfiled rather than wrong. In one sentence, where does the real risk sit?

Paragraph.

Question 8multiple choice

On the first picture, the arrow labelled numbers changing stops at the freeze. What is missing from the right-hand side of that picture?

AAn arrow that changes the numbers
BThe prompt going in
CThe answer coming out
DThe knowledge cutoff

Question 9multiple choice

On the second picture, the zoom ends on a single parameter. If you read that one number on its own, what could you tell about what the model knows?

ANothing; knowledge is the pattern across all of them
BOne complete fact
CThe model’s knowledge cutoff
DWhich stage wrote it

Question 10written

On the table, Gemini’s parameter count is listed as not published. Name one reason a company might not publish it.

Paragraph.

Question 11written

A chatbot uses your name three messages after you gave it. Explain how this happens without the model changing.

Paragraph.

Question 12writtenoptional

Two sentences. Make a claim about whether running a model locally solves the data concern, and give one reason.

Paragraph.

The first item is your name. Answer it — that is what puts you on the work now that email collection is off.

Exit ticket · on your own