#LLMs Unplugged

Teaching Resources for a ChatGPT World

#Acknowledgement of Country

a(nother) talk about AI*/LLMs

they’re A Thing now

#Hi, I’m Ben

#The problem

everyone uses ChatGPT (classrooms, office work, group chats)

most folks don’t have a good mental model of how it works…

…and this is a problem

#The gap

some feel this “don’t know how it works” anxiety more than others, e.g. classroom teachers

CS Unplugged has been teaching computing concepts without computers for 20+ years

algorithms & data structures
traditional ML/AI
LLMs?

#What would “LLMs Unplugged” look like?

end-to-end generation (training → generation)
modular and low-friction
broad accessibility (no coding, minimal maths)

(spoiler) it’s live at llmsunplugged.org

#Historical roots

#Markov (1913) and Shannon (1948)

did this work by hand

counting letter sequences in Pushkin’s Eugene Onegin (Markov)

generating synthetic text through weighted sampling (Shannon)

#The core insight

language models work by

counting patterns in existing text
exploiting these patterns to generate new text

modern LLMs: same approach, vastly greater scale

#The core mechanic

#Training: counting word pairs

Text: “See Spot run. See Spot jump.”

Preprocessed: see spot run . see spot jump .

	`see`	`spot`	`run`	`.`	`jump`
`see`		II
`spot`			I		I
`run`				I
`.`	I
`jump`				I

the grid is the model

#Demo time

llmsunplugged.org/lessons/training

llmsunplugged.org/lessons/generation

#Why this works

the randomness explains why LLMs give different responses to the same prompt
the weighting explains why output is usually coherent
this is how LLMs work, just smaller

#Lesson structure

#Fundamentals

two core lessons (do in order):

Training—counting patterns into a grid
Generation—weighted sampling with dice

available in grid and cutouts variants

about 30mins each (but flexible)

#Extensions

Pre-trained Generation

More Context

Sampling

Tool Use

In-context Memory

Induction Heads

Word Embeddings

LoRA

RLHF

Synthetic Data

llmsunplugged.org/lessons/pretrained-generation

#Reception

#Notes from the field

over 500 participants over the past year

senior executives and public servants
undergraduates (all disciplines)
high school teachers
school-age students

LLMs are “just” probability and randomness at scale—not reasoning, not understanding, but sophisticated pattern matching

#The shareback moment

the room comes alive during the generation lesson when groups share the output from their model

“fish fish fish red one fish two fish”

“I do not like green eggs and you may see me be I do not…”

#Resources

#What’s available

lessons with instructor notes
interactive widgets
videos (coming soon)
software tools for custom n-gram booklets

all CC BY-NC-SA licensed

llmsunplugged.org

#Next steps

continued development (esp. extensions)
teacher training
evaluation (inc. via partnerships)
strengthen curriculum links (e.g. ACARA)
translations
keen to hear your experiences

#This Friday

Lifting the Veil—a hands-on AI workshop for educators with Brimbank Tech School at VU Sunshine Campus

Friday 13 February, 10am–3pm

#Thanks