Mission control installs a sharper telescope camera. The pictures look amazing, and every object now measures one pixel wider.
By lunchtime, Sky Sorter is calling stars planets all over the place.
"Nothing in the pseudocode changed!" Rocket says. "Why is it getting worse?"
Raven compares old and new cards side by side. "The data changed. A size 3 star now measures size 4."
Nova dims her lights thoughtfully. "A sorter learns from the data it was given," she says. "When the world changes, it may need new data."
The crew starts labeling a fresh batch.
AI systems may be trained on data that changes over time.
When that happens, how well they work can change too, in ways that are hard to understand.
That is one reason teams keep testing, even after a system is finished.
Real AI is a tool people build. It does not notice on its own that the world has changed, and it can be wrong.
| Card | True label | Size on old camera | Size on new camera | Version 3 label (new camera) |
|---|---|---|---|---|
| S1 | star | 3 | 4 | planet |
| S2 | star | 2 | 3 | star |
| S3 | planet | 6 | 7 | comet |
| S4 | planet | 4 | 5 | planet |
| Week review | True or false? |
|---|---|
| Machine learning needs tremendous amounts of training data. | ? |
| The test set is used to choose the sorter's cut-offs. | ? |
| Accuracy is the number right divided by the total tested. | ? |
| Data that changes over time can change how well an AI system works. | ? |
| Once a sorter scores well, it never needs testing again. | ? |
Next week the crew looks harder at its training data.
What if some kinds of snapshots were left out of it on purpose, or by accident?
Outstanding week, explorer! Tomorrow your family trains and tests a guesser together.