NOVA ACADEMY · INTERACTIVE LECTURE · HOW MACHINES LEARN
{{ stepTitle }}
{{ stepNum }} / 10
How does a machine actually learn?
In the video we saw AI everywhere. Now we open the box: together we will train a cat-detector from scratch — then see how the very same trick, scaled up, becomes something like ChatGPT.
{{ st.t }}
{{ st.arrow }}
Use the ← → keys to move. Every screen has something for the class to do.
EXAMPLE {{ labelCount }}
Is this a cat?
A machine starts out knowing nothing. Humans must first label examples for it — this is called training data.
{{ fbTxt }}
You just built a dataset.
Six labeled examples — that is exactly what data annotators do, millions of times over.
6
OUR CLASS
50,000+
A REAL CAT DETECTOR
Trillions of words
CHATGPT
DATASET SO FAR
{{ lc.tag }}
A computer has no eyes
To the machine, a photo is just a grid of numbers — one number per pixel saying how bright it is. Words, sounds, everything becomes numbers first.
{{ pxHint }}
Remember the pixels? A photo became numbers. Text must become numbers too — so the model cuts every sentence into small pieces called tokens, and numbers each piece.
Nova Academy teaches AI in Kuala Lumpur
Why cut words up? The model owns a fixed "dictionary" of pieces — like a box of LEGO bricks. It cannot add new bricks.
Blue = common words like "teaches" and "in". They appear so often they earned their own brick.
Orange = rare words like "Academy". No brick for the whole word — so it snaps together from smaller bricks it already has: "Acad" + "emy".
That is the trick: with enough small bricks, the model can read any word ever written — even names and words invented tomorrow.
NOW SCALE IT UP — DRAG THE SLIDER
{{ scDials }}
DIALS (PARAMETERS)
{{ scData }}
TRAINING DATA
{{ scTime }}
TRAINING TIME
{{ scMsg }}
JOIN IN
Scan to become an AI trainer
scan with your phone camera — or type the link below
ai-workshop.nova.edu.my/train
{{ joined }}
TRAINERS JOINED
CLASS ACCURACY
{{ clsAcc }}%
LABELS RECEIVED
{{ clsCount }}
{{ clsMsg }}
ANSWERS INDIALS ADJUSTINGDECISION
CAT
{{ catCount }}
NOT CAT
{{ notCount }}
LIVE ANSWERS
{{ poisonMsg }}
PHOTO INPATTERN SPOTTERSDECISION
CAT
{{ outCat }}
NOT CAT
{{ outNot }}
A web of tiny number-dials
Each circle is a neuron — a tiny calculator. Each line has a weight: a dial that decides how strongly a signal passes through. A photo goes in as numbers on the left; a guess comes out on the right.
{{ fireMsg }}
Watch the lines: training makes some dials stronger.
ACCURACY ON NEW PHOTOS
{{ accuracy }}%
{{ accMsg }}
{{ ph.t }}
{{ ph.d }}
MISTAKES (LOSS) · ROUNDS TRAINED: {{ rounds }}
ChatGPT is the same guessing game — but instead of "cat or not", it guesses the next word, having read a huge slice of the internet.
{{ sentence }}
CREATIVITY (TEMPERATURE)
{{ tempName }}
THE MODEL'S TOP GUESSES — CLICK ONE, OR LET IT PICK
The model "wrote" a sentence — one guess at a time. It never looks anything up; it only predicts what word most likely comes next.
A raw model predicts words — but is it helpful? Thousands of people compare its answers and pick the better one. Which would your class pick?
PROMPT {{ rlRoundLabel }} · {{ rlPrompt }}
{{ prefMsg }}
That is how it learns manners.
Every preference nudges the dials again — toward answers people find helpful, honest and safe. This step is called learning from human feedback, and it is what turns a word-predictor into an assistant.
HELPFULNESS
{{ rs.arrow }}
{{ recapTitle }}
{{ recapDetail }}
From 50,000 cat photos to ChatGPT — the same core idea: learn from examples, one tiny nudge at a time.
LECTURER{{ note }}