Glean
FEATURE SELECTION — *a computer can't look at everything; what you let it look at is what it learns from — so choose the clues on purpose, and drop the ones that only sit near the answer.*
Loading audio…
Press play to listen along. The line being read lights up as you go.
Show full transcript
Loading transcript…
Glean is a bright-eyed magpie with a patched satchel and a habit that surprises everyone: she picks a thing up, turns it over, and then — this is the part people miss — puts most of it back down. Feathers, buttons, bottle-caps, facts. She gathers a heap and keeps only the few that carry real meaning.
Her craft is feature selection — deciding which clues a model is even allowed to look at. "A model can't attend to everything; there's too much," she says, tapping the satchel. "So somebody picks: out of a thousand things you could measure, which handful actually help?" She spreads a few kept clues on the ground. "Here is the whole secret. The clues you feed it are the world it gets to see. Feed it the ones that cause the answer, and it learns something true. Feed it clutter, and it learns clutter — confidently. Choose them like it matters, because nothing about the model matters more."
She learned it the hard way, building a little machine to tell a wolf from a husky. It scored perfectly on her test photos — then failed on new ones, wildly, embarrassingly, and she couldn't see why until she opened it up and looked at what it had actually been staring at.
"It wasn't looking at the animal at all," she said, feathers drooping. "Every wolf photo in my pile happened to have snow in the background. So it learned 'snow means wolf.' I built a snow-detector wearing a wolf costume." Nobody had chosen that clue on purpose — the machine had simply grabbed the loudest, easiest thing in the frame and leaned its whole weight on it. Glean has a name for that trap now: a clue that only rides along with the answer instead of making it. The snow doesn't cause the wolf; the two just kept company in her photos. "A clue that sits near the answer by luck," she says, "is a coincidence with good manners. It'll shake your paw and betray you the first time the weather changes."
When she was twelve, Glean walked to NeuralQuest, and Sift the old owl asked her a question at the gate.
"Of everything you could measure, how do you know which few to keep?"
"You ask what actually causes the answer," Glean said, "not just what happens to sit beside it. Snow sits beside wolves; it doesn't make one. I keep the clues that carry the real signal and drop the ones that only got lucky." She paused. "And there's a sneakier trap than luck. Sometimes a clue is in the pile that gives the answer away — a clue that quietly already contains the thing we're trying to guess. Those feel amazing and teach nothing. I hunt those hardest of all."
Sift smiled. "You are the one. Nothing gets fed to a model in this place until you've decided what's worth feeding it."
In her workshop, clues hung from strings like laundry. She sat a kid down with a pile of facts about houses and one job: guess the price. "Look at all of these," she said. "Number of rooms — keep it, that really shapes the price. Colour of the front door — drop it, that's noise. Size of the yard — keep it, that carries signal. The owner's favourite ice cream — drop it, unless you want a machine that swears vanilla-lovers pay more." Then she pulled one card aside and held it up like a snake she'd caught by the neck. "And this one. 'Final sale amount, rounded.' Somebody left it in by accident. It's basically the answer with a hat on. Train on that and the machine looks like a genius and has learned nothing — the day the hat comes off, it's helpless." She taught the whole habit then, as one steady rule: more clues is not better — the right clues are; watch for clues that only sit near the answer instead of causing it, because those are the snow-in-the-background traps that break the moment the world shifts; hunt the clue that secretly smuggles in the answer, because that one flatters you and teaches your machine to cheat; and remember a clever model will always seize the laziest clue you leave lying around, so the kindest thing you can do is hand it only the ones that carry true meaning. "Feeding a computer everything," she said firmly, "feels generous. It's actually how you teach it to trust coincidence and call it knowledge."
"So leaving stuff out isn't lazy," the kid said, surprised. "It's… taking care? It's deciding what's real before the machine ever gets a vote?"
"It's the whole job," Glean said. She swept the door-colour and the ice-cream and the sneaky answer-in-a-hat gently off the table, keeping only the clues that told the truth, and let the cleared space breathe. Under the quiet she felt the calm that always came once the last useless clue was set down — not the itchy overwhelm of a heap too big to hold, not the guilt of a machine misled by clutter, but a light, airy, uncluttered gladness. It was the relief of a table holding only what mattered, and the settled knowing that whatever learned from it now would learn something true, and would keep being true even when the snow melted and the weather changed. That spare, kept-only-what-counts feeling, steadier than any bulging satchel, was to Glean exactly why she picked things up in the first place — so that, so often and so happily, she could put most of them back down.
The NeuralQuest ensemble
Glean is part of NeuralQuest's distributed-narrative cast. Each character embodies a different curricular primitive; together they teach the full subject.
-
Tag
Labeling — the cheerful labeler who treats every label as a human choice and meaning-making act ('every label is a choice — and you're the one making it')
-
Drill
Training loops — the focused practitioner who treats iteration as rhythm, not race; explicit teacher of when-to-stop ('once, again, again — different this time? Then again')
-
Skew
Bias + data fairness — the bias-vigilance anchor who always asks 'whose data is in here, whose is missing, who decided'; appears in every kit from kit 5 onward
-
Veer
Generalization vs overfit — the wandering scout who treats generalization as travel ('trained here, tested here — now go somewhere new, does it still know the way?')
-
Weigh
Ethics + decisions — the reflective elder who carries the ethics gate at the AI-in-society capstone ('can we build it? Yes. Should we? That's a different question')
-
Foretell
Prediction — the model learned on the past; now it guesses about something it's never seen (the leap is the whole point and the whole risk)
-
Odds
Confidence — I'm not sure, I'm 80% sure; those are different, and the difference is the whole point
-
Rue
Loss — how far off was that guess? the exact size of the miss tells the model which way to change
-
Verge
Decision threshold — the model gives a number, I draw the line; move the line and you choose what you'd rather be wrong about