Observed
Binny struggled with unfamiliar shapes, materials and conditions.
Why do you think this occurred?
AI Literacy Mini-Lab
Train Binny: Can better examples help an AI model sort rubbish?
- 10 to 15 minutes
- No signup
- Nothing is saved
Words used in the lab
Tap a question mark if a term is unfamiliar.
Training examples
Labelled examples used to train the model.
Model
A system trained to find patterns and produce an output for a new input.
Test set
Examples kept separate from training so we can see how the model responds.
Output
The category or action produced by the model.
Train Binny
Follow the examples from training to testing.
1 of 9
Step 1 of 9: Meet
Step 1
Meet Binny.
Meet Binny, the AI recycling robot! Binny helps humans in a recycling centre. The humans showed Binny examples of what can be recycled.
Step 2
Inspect Binny's first training examples.
Clean plastic bottle 
Clean plastic container 
Clean aluminium can 
Clean steel can 
Clean paper stack 
Clean newspaper
AI-assisted object illustrations
What does Rosie notice?
Every example above already has a name and a correct answer attached — the humans knew these items could be recycled before Binny ever saw them. Training with labelled examples like this is called supervised learning: the model learns by comparing its own guesses to answers that were already known.
Step 3
Algorithm 1.0
After reviewing the training examples, Binny's internal settings (parameters) were adjusted so it could predict what should be recycled and what should be trash.
Binny has now seen six examples. Let's find out what it predicts for something new.
Step 4
Predict Binny's first test.
0 of 8 outputs revealed
Show what Rosie noticed
Binny handled the familiar bottle and newspaper easily. Changed shapes and new materials caused inconsistent results. Several recyclable items were sent to trash.
Step 5
One week later
Let's check in with the recycling centre manager after one week of Binny sorting items.
Pattern
Here's what the humans think Binny is doing: recycling anything that looks like paper, plastic or metal, and trashing everything else.
Impact
Human workers had to correct many of Binny's errors.
In our fictional centre, many recyclable items were being sorted as trash, so the process became less efficient.
Step 6
Fetch more examples for the training, Rosie!
Rosie fetches examples with new materials, shapes and conditions. The humans review and label them before using them for training.
Household battery 
Yoghurt tub 
Takeaway coffee cup 
Squished plastic bottle 
Cereal box 
Dirty peanut butter jar 
Tennis ball 
Bicycle tyre 
Light globe 
Electrical cable 
Labelled glass bottle 
Egg carton
AI-assisted object illustrations
Step 7
Algorithm 2.0
After the humans reviewed the errors and added new examples, Binny's internal settings were adjusted again. Let's see how it handles new and unfamiliar items now.
What did Rosie notice?
Training a model again on a smaller, targeted batch of examples — usually to fix specific gaps like Binny's — is sometimes called fine-tuning. It's still supervised learning, just a second, more focused round of it.
Step 8
Retest Algorithm 2.0.
0 of 12 outputs revealed
Show what Rosie noticed
The broader examples helped Binny handle more shapes and materials. The dirty jar was sent for cleaning, and the unfamiliar battery was sent to a human. Better performance still depends on people choosing useful examples, rules and checks.
Step 9
Binny Lab Note
Train Binny
What our fictional demonstration showed
What the humans observed
It looks like Binny recycles clean items, cleans dirty recyclables before sorting, asks a human when something's confusing, and only trashes what it's confident isn't recyclable.
What the new examples helped with
A broader range of clean, dirty, changed and unfamiliar objects helped Binny handle more cases and recognise when a person should review an item.
What a human should check next
Review unfamiliar or potentially hazardous items, including batteries, cables and light globes, before choosing a facility action.
Why humans still matter
Humans need to provide the training examples. They will always need to be involved in handling unusual cases and fixing mistakes.
