First-Test Log, NEW LIGHT card, NEW ANGLE card
Open mission materials →
Concept art made with Gemini 3.1 Flash Image← Mission map
Mission 5 of 12Teachable Machine40 to 50 minutes
Blockstorm: Biome Survival
Nyx opens a block-world survival wave in conditions Image Model v1 has never seen.
Today's AI power
Generalization
A useful model works in new light, angles, and backgrounds, not only its practice setup.
Start here
Predict and test two unseen biomes before retraining, then survive the wave.
Two unseen conditions are written before testing.
Biome Portal and Failure Ledger v1
See how this game works
Game loop preview
See the power before you build it.
- 1SignalA brand-new test condition
- 2AI moveThe model tries a setting it never practiced
- 3Your ruleThe exact game logic checks what happens next.
- 4Game stateASK AGAIN
Your mission
One step at a time. One AI power to unlock.
Step 1 of 9Story
Story
Meet the mission problem.
Story signal: Nyx opens a block-world survival wave in conditions Image Model v1 has never seen.
Do this now
- 1LOOK
Spot the problem in the opening story and tell your crew what needs to change.
Goal
Know what winning looks like.
Your choice: Choose the unseen condition most likely to reveal a shortcut.
Watch this change: A new condition reveals whether the model learned the prop or a shortcut.
Do this now
- 2CHOOSE
Win by proving this: Two unseen conditions are written before testing.
Get ready
Gather the props and get your grown-up guide.
Do this now
- 3CHECK
Open the Mission 5 materials on screen or print them, then gather the listed props.
- 4CHECK
Test safe object conditions such as light, angle, distance, and background on the tabletop zone.
Predict
Guess first, then test.
Do this now
- 5CHOOSE
Predict which new condition will be hardest and why.
Build
Make the AI power in the Google tool.
Do this now
- 6MAKE
Follow the Google tool card below. Complete each step in order, then bring the visible result back here.
Google tool step
Open the tool, make the power, then bring one result back.
OPEN TOGETHER
Teachable Machine
Adult and child operate together
Build thisExisting image project
Save asFirst-Test Log v1
Do these steps
- Open Image Model v1 without retraining.
- Set the first unseen condition.
- Predict the result.
- Run one first test.
- Record the result before changing data.
- Repeat for the second condition.
You are done whenThe prediction bars reveal one success, surprise, or failure under each unseen condition.
Need setup help?
- Bring
- Image Model v1
- NEW LIGHT card
- NEW ANGLE card
- Blank First-Test Log
- Next move
- Circle one result and investigate a possible shortcut.
- If stuck
- Confirm the correct saved model and label map before repeating a failed setup.
Bring it back
Return with the result you can see.
Do this now
- 7CHECK
Explain one training shortcut that might have caused the result. Save the exact mission artifact as First-Test Log v1, then return to this card.
Glitch test
Make a funny mistake on purpose.
Funny glitch: Rex equips a jump spring when the cave needs a shield.
Do this now
- 8TEST
Move a card from its familiar background to a busy or differently lit scene and test before teaching the model that scene.
Boss round
Change one thing and retry.
Game move: Block-world portal run with power equipment swaps
Harder remix: Separate a likely data-shortcut failure from a threshold failure and justify the difference.
Do this now
- 9REMIX
Choose one evidence-based change, run the same test again, and compare what changed.
- 10TEST
Run the boss round: No test example is added to training before the first result is saved.
Explain
Say what the AI did and prove it.
Power reward: Generalization Detective · Biome Portal and Failure Ledger v1
A useful model works in new light, angles, and backgrounds, not only its practice setup.
A high training score can hide shortcuts. New lighting, distance, angles, and backgrounds reveal what the model really learned.
Do this now
- 11SAY
Why is a genuinely new first test more useful than showing the model its training pictures again?
- 12SAVE
Complete the mission proof and save only the named artifact: First-Test Log v1.
What counts as finished?
Why is a genuinely new first test more useful than showing the model its training pictures again?
- Two unseen conditions are written before testing.
- The first-test log contains expected and observed results.
- No test example is added to training before the first result is saved.
Need help or want a harder challenge?
Make this step easier
Use two picture cards labeled NEW LIGHT and NEW ANGLE and draw a happy or surprised face for each result.
Say it this way: The AI noticed ___. The exact game rule ___. I know because ___.
Try the deeper remix
Separate a likely data-shortcut failure from a threshold failure and justify the difference.
- Record the inputs and outputs.
- Compare two versions or conditions.
- Name the constraint or exact rule.
- Design one fair remix test.
Everyone proves the same AI idea: Two unseen conditions are written before testing. · The first-test log contains expected and observed results. · No test example is added to training before the first result is saved.
Quick power check