Feed

"These examples aren't just facts," she would say, tapping her stack. "They are choices. Human choices. Someone picked every picture. Someone wrote every label. They decided what was 'right'."

That's why Feed worked so closely with Catch. Catch was her friend over in the DataForge. When data arrived at the AIForge, it always came with notes from Catch. Catch asked the important questions. Who gathered this data? Why did they gather it? Feed needed those answers to make sure her examples were good. They were a team.

Like Sort, Feed grew up in a paper-crafts workshop. The workshop had one big rule. Every paper figure was made to help another. Feed was folded and clipped together for one reason: to help Sort. Sort learned her first rules for sorting things by studying Feed's cards. They were a perfect pair.

When she first arrived at the AIForge, Bit asked her a question. Bit was the head of the academy.

"What is training data?" Bit asked.

Feed didn't hesitate. "It's the examples a model learns from," she said. Her voice was crisp and clear, like folded paper. "The model is only as good as its examples. Good examples, good model. Bad examples, bad model. Garbage in, garbage out."

Bit smiled. "You're hired."

Feed started her first class the same way every time. She would lift her entire stack of cards. With a flick of her paper wrists, she fanned them out. The students saw a rainbow of tiny pictures and words. Each one had a label.

"I am Feed," she'd say. "And this," she shook the cards gently, "is training data. It's what an AI model learns from. The model is what these cards teach it. Simple as that."

She taught her students five big questions to ask about any stack of training data.

"First," she said, holding up one paper finger. "Where did these examples come from? Who collected them? And why? Remember my friend Catch? Her notes are key. Sometimes, the way data is collected can hide a problem."

"Second: Who wrote the labels? What rules did they follow? Were they the right people to be deciding the 'right' answer?"

"Third: What's included? Do the examples show all different kinds of things? Or just the common ones? What about weird, rare situations?"

"Fourth," Feed's voice grew firm. "What's missing? This is the most important question. A model can't learn about something it never sees. If there are no pictures of purple bananas in my stack, the model will never know they exist."

"And fifth: Are the numbers balanced? If I have ten thousand pictures of cats and only two pictures of dogs, what do you think the model will be good at spotting?"

She paused, looking at the students. "It will be a cat-spotting expert. And a dog-spotting disaster."

"And that brings me to my biggest rule," Feed said. She snapped her cards back into a perfect stack. "Garbage in, garbage out. You can't fix bad examples with a fancy model. A wobbly foundation makes a wobbly building."

She looked at them seriously. "And never, ever say the model 'thinks' or 'understands.' It doesn't. It finds patterns. That's all. Be honest about what it's doing."