1 of 56

How Do LLMs Work Welcome Activity

Introduce yourself in chat: Name, Location, Connection to K-12 CSEd, and answer the question What’s your favorite chatbot?

Slides: tinyurl.com/cstaca260122slides

Welcome to the 3rd CSTA California Statewide community of practice event of the 2025-26 school year. Please introduce yourself in chat: Where are you located and what is your connection to K-12 CSEd; and What is your favorite chatbot?

Link to slides: https://tinyurl.com/cstaca260122slides

FYI: Do not read this aloud

5:00 - 5:07 Welcome & Intros (CSTA GLA Chapter Leader)

5:07 - 5:15 What is a “Generative AI Model”? (David Czechowski & Vicky)

5:15 - 5:25 Seuss-Inspired Language Model #1 (Vicky)

5:25 - 5:50 3 Seuss-Inspired language models with cards & Connect to current LLMs (David)

5:50 - 5:57 Resource Sharing / Q&A (CSTA GLA CL sharing Padlet & inviting questions / Vicky & David answering questions)

5:57 - 6:00 Share Next CSTA CA Event & Transition to Chapter Breakouts

6:00 - 6:30 Chapter Breakouts

2 of 56

Tech Norms

Mute and unmute strategically. During discussions, it is OK to stay off mute to allow for more natural discussions, but please mute yourself if there is a lot of background noise.
Check your display name (Susie Smith - Location)
Show your video, if able.
Participate! Use the chat to ask questions.
If you take a break, don’t forget to mute yourself and turn off your video.
Have water and snacks handy. It is okay to drink and munch while in the session.

3 of 56

How Do Large Language Models Work?

4 of 56

David Czechowski

CS/Technology Teacher

CSTA-New York

davidczechowski@hpcsd.org

www.TeachingIsSTEM.com

Vicky Sedgwick

Retired Elem CS Teacher

AI4K12 K-2 Grade Band Lead

CSTA GLA Chapter President

visionsbyvicky.com

Both

5 of 56

Agenda

Welcome and Introductions
What is a “Generative AI Model”?
Some Seuss-Inspired Models
Connect to current LLMs
Resources and Q & A
Chapter Breakouts�

Slides: tinyurl.com/cstaca260122slides

AI

LLM

Vicky

Here is an overview of our time today.

It is our hope that you will leave with a deeper conceptual knowledge of the Generative AI tools that seem to b everywhere these days. To do this, we want to begin by anchoring the presentation to the technical topic in a broad sense. What is a Generative AI Language Model? Then, we will progress down into simpler language within a specific context, with some Dr Seuss-inspired AI models. These are hands-on models though it’s difficult to make them truly hands-on over Zoom.Then, we will wrap-up our time by returning to the general and technical. And, we will leave you with some questions to consider in the chapter breakouts.

Our time is limited, so we will provide you a Padlet of resources that we encourage you to investigate and share. We are not experts on this topic but we are happy to share our thoughts on anything you are wondering about, especially after we establish a shared conceptual foundation.

And, before we forget, we want to say “Thank you!” We appreciate that each of you has blocked out the time to think about and discuss this topic. You each bring a unique perspective, and we look forward to your questions and thoughts. No expertise is expected.

General & Technical:

What is a “Generative AI Language Model”?

Concrete & Contextual:

Play With Some ‘Seussian’ Models

General & Technical:

Connect to current LLMs
Discuss “So what?” in Chapter Breakouts

6 of 56

Unplugged Introductions to

Generative AI Language Models

How Do LLMs Work?

Vicky

7 of 56

Questions to Be Pondering

How can AI be wrong?

Is AI intelligent?

Where does AI bias come from?

Vicky

??

^🤔

8 of 56

Define: "Generative AI Model?"

David

9 of 56

“AI”

Generative AI Model

In the chat…

How do you respond to a student that asks: What is AI?

David

10 of 56

Generative AI Model

David

The term “AI” is used in many and various contexts. From glucose monitoring, to self-driving cars, to robots, to gaming, etc… This mind map is the Association for Computing Machinery’s classification of AI. Under the ‘big umbrella’ term of “AI” we will be exploring just a small sliver. Today’s presentation is about “natural language processing” AI, specifically “Language Generation” or “Generative AI”.

In fact, I would like to pass along a caution that is counter-cultural: An exciting advance in one niche under the umbrella of AI does not mean all the technologies we categorize as AI has similarly advanced. There is a reason no one is talking about having ChatGPT drive their car. These are different categories with different challenges.

“Mind-map of AI based on the ACM classification.”

Image source: http://dx.doi.org/10.3233/978-1-61499-125-0-361

11 of 56

Generative AI Model

Generative AI caught the world’s attention because the output appeared to demonstrate human-like ability.

Output is more than just pre-planned responses.
Patterns are used to produce content.
Text, images, sounds, video, code, etc.

David

12 of 56

Generative AI Model

“A simplified representation of a system, process, or concept.”

Some models can help us…

Communicate about a thing
Understand how a thing works
Predict how a thing might behave

David

13 of 56

Some Models of�“Green Eggs & Ham”

Vicky

14 of 56

Vicky

15 of 56

Vicky

16 of 56

Some Popular Large Language Models

Vicky

Most people know what these things *do*, but… What are they? They are large language models or LLMs that (attempt to) represent spoken language.

Some models are made by people, for people (like the Green Eggs & Ham models we just saw).

Some models are made by people, some are made by computers.

Some models are for humans to use, some are for computers to use.

Services like ChatGPT, Gemini, Perplexity, Claude, Grok, etc. have captured the headlines with their ability to generate sentences, paragraphs, essays, code, etc. Each of these use their own complex mathematical representation of the English language (language model) that is fine-tuned to predict a desired response to textual prompts.

The complexity and size of these language models have earned the label “Large… Language Models”

17 of 56

Let’s Look at Some Unplugged “Green Eggs & Ham” Language Models

Vicky

18 of 56

A Story with Only 50 Vocab Words

a

am

and

anywhere

are

be

boat

box

car

could

dark

do

eat

eggs

fox

goat

good

green

ham

here

house

I

if

in

let

like

may

me

mouse

not

on

or

rain

Sam

say

see

so

thank

that

the

them

there

they

train

tree

try

will

with

would

you

Vicky

19 of 56

Language

Our Language Models

Dr Seuss’ “Green Eggs & Ham”

Sentence Strips from Green Eggs & Ham
A Deck of Cards
A Stack of 927 Cards
Markov Chain Card Decks

Vicky

The 4 hands-on language models of “Green Eggs & Ham” we are going to “play with” all focus on the language of the book, the words, NOT the story. While a diorama captures visual moments from the book, these are models of the words, punctuation, and relationships between the words of the book, in all or in part.

The first model was created for use with lower elementary students using only some of the sentences from “Green Eggs and Ham”. It is the simplest to prepare but provides the least flexibility in sentence structure. The next 3 models all use prepared word and/or punctuation cards and slowly increase in complexity. As a result, the quality of the sentences they can generate also increases.

If ChatGPT truly worked like any of these, it is unlikely we ever would have heard of it. However, all of these models are far far less resource intensive.

We hope that by seeing multiple language models you (and your students) can develop a legitimate conceptual understanding of Generative AI Language Models. Don’t worry about the big picture yet, we will return to that. For now, we’ll explore each model on its own.

20 of 56

Model #1: Sentence Strips

Vicky

21 of 56

Model #1: 5 Containers & Sentence Strips

21

1

2

3

4

5

Vicky

22 of 56

22

I

do not

Sam-I-am.

that

I do not like that Sam-I-am.

like

1

2

3

4

5

Vicky

23 of 56

23

I

do not

Sam-I-am.

that

I do not like them Sam-I-am.

like

them

1

2

3

4

5

Vicky

24 of 56

24

I

do not

Sam-I-am.

that

I do not like green eggs and ham.

like

them

green eggs and ham.

1

2

3

4

5

Vicky

25 of 56

25

I

do not

Sam-I-am.

that

I would not like them here or there.

like

them

green eggs and ham.

would not

here or there.

1

2

3

4

5

Vicky

26 of 56

26

I

do not

Sam-I-am.

that

I would not like them anywhere.

like

them

green eggs and ham.

would not

here or there.

anywhere.

1

2

3

4

5

Vicky

27 of 56

27

I

do not

Sam-I-am.

that

I will eat them anywhere.

like

them

green eggs and ham.

would not

here or there.

anywhere.

will

eat

1

2

3

4

5

Vicky

28 of 56

28

I

do not

Sam-I-am.

that

like

them

green eggs and ham.

would not

here or there.

anywhere.

will

eat

I do like green eggs and ham.

do

1

2

3

4

5

Vicky

29 of 56

Sentence Strips

Hands-On�scratch.mit.edu/projects/1042008593/fullscreen/

Share the sentence generated for you in chat

Vicky

30 of 56

Model #2: Deck of Cards

David

31 of 56

Model #2: Deck of Cards

50 Unique Words

3 Punctuation Marks

1 Per Card

like

David

32 of 56

Generating a Sentence with Model #2

Randomly draw 1 card at a time.

Stop when a punctuation mark is drawn (. ! ?).

Don’t use any of your intelligence to reject words, change the order, or stop early.

TRAIN

IF

EGGS

THAT

David

33 of 56

Deck of Cards

Hands-On

https://scratch.mit.edu/projects/1186085429/fullscreen/

David

34 of 56

Model #2 Example Sentences

IN WOULD BE LET ON DARK GOOD WILL DO MOUSE ANYWHERE DO !

IN THEM TREE GOOD YOU MOUSE SAY TREE FOX OR WOULD ?

NOT DARK SEE SAY FOX A LIKE IF BOAT ME SO TRAIN GREEN THANK I A HAM TRY MAY SEE BE AND LIKE ?

Gibberish Sentences.

35 of 56

Model #3: 927 Cards

David

36 of 56

Model #3: 927 Cards

One card for each word/punc in the book.

David

37 of 56

Generating a Sentence with Model #3

Randomly draw 1 card at a time.

Stop when punctuation is drawn.

(Don’t use your intelligence to reject any words, or change the order, or stop.)

NOT

EGGS

THAT

David

38 of 56

927 Cards

Hands-On

https://scratch.mit.edu/projects/1078429038/fullscreen/

David

39 of 56

927 Cards

Sentence Generation Improvements

In the chat…

What is different about the sentences that are generated by this model?

David

40 of 56

Model #2 Example Sentences

TRAIN AND NOT I !

SEE NOT IN ON !

LIKE THEM EAT HAM THEM .

Still Gibberish.

41 of 56

Generated Sentences from Model #3

Some observations…

Common book words appear more often.
Sentences are more similar in length to those in the book.
Still gibberish.

NOT

EGGS

THAT

David

42 of 56

Model #4: ‘Markov Chain’ Decks

David

43 of 56

Model #4: Markov Chain Decks

50 Decks

One for each of the unique words in “Green Eggs & Ham”

927 Cards

Each card is put into the deck corresponding to the word that preceded it in the book.

Collection of “Next Words”

NOT

David

44 of 56

Model #4: Markov Chain Decks

Example “You” Deck (34 cards):

LIKE, LIKE, LIKE, LIKE, EAT, EAT, ?, ?, MAY, WILL, MAY, LET, WOULD, ON, COULD, IN, COULD, IN, SEE, DO, WOULD, WITH, COULD, ON, LET, DO, SAY, MAY, MAY, WILL, WILL, SEE, !, SAM

YOU

David

45 of 56

Model #4: Markov Chain Decks

David

46 of 56

Generating Sentences with Model #4

Randomly draw a ‘Starting Word’
Use that word’s deck to randomly draw the next word.
Keep repeating with each new word’s deck.
Stop when a punctuation mark is drawn (. ! ?).

GREEN

EGGS

AND

GREEN

EGGS

Start

David

47 of 56

Markov Chain Decks

Hands-On

https://scratch.mit.edu/projects/890846221/fullscreen/

David

48 of 56

Model #3 Example Sentences

IN A BOX.

COULD NOT WITH A BOAT.

LET ME BE!

49 of 56

Markov Chains

Andrey Markov (1856-1922)

Modeled processes as sequences of steps.

Probabilities assigned to the chance of each possible next step.

Studied letter sequences in Russian Poetry

David

50 of 56

Generative AI Language Models

David

51 of 56

“Corpus”

How much more complete would this language model be with all this data?

How might sentence generation improve?

1 Book

927 words

3

David

52 of 56

“Training”

What happens to the effort needed for building a model as sophistication increases?

LLMs are not training while being used.

David

53 of 56

“Tokens”

How many more language patterns might be detectable if word roots were analyzed, not just whole words?

TRAIN

EAT

S

ING

EN

David

54 of 56

Resources To Use & Share

Vicky

55 of 56

Join us on 12 March 2026 from 5:00-6:30pm PT for …

PLUG IN. UNPLUG. POWER UP LEARNING!

Register: tinyurl.com/cstaca260312

56 of 56

Credits

Many of the images were generated by ChatGPT 4o with prompts like: "Generate an illustration in the style of Dr Seuss of a creature with a hand puppet with a puppet holding a puppet."

Model #1 resources originally created for AI for CA by Vicky Sedgwick with feedback from Mark Loundy and Katherine Goyette. Thanks to Dr. Emily Thomforde for her idea of Sentence Maps of Green Eggs and Ham.