a neural network, hand-written from scratch — no libraries — learning to predict graduate earnings from real U.S. Dept. of Education data. every neuron visible.
FORECAST A SCHOOL
ask the model — updates live as it learns
$– /yr
predicted median earnings, 10 yrs out
—
real data: U.S. Dept. of Education · College Scorecard (data.gov) · … schools · model + math hand-written, no libraries
DATA.PUNK · WEEK 3
GLASS BOX
Everyone else this week asked an AI about the government's data. This is the opposite: a neural network built from nothing — the forward pass, the backprop, the optimizer, all hand-written, no TensorFlow — that learns from the data itself, live, in your browser. And unlike the models the government (and the AI labs) keep in black boxes, you can see every neuron and every weight as it thinks.
What it learns: the U.S. Dept. of Education publishes what graduates of every college actually earn. Watch the network teach itself to predict it — from a school's cost, size, selectivity, and type — and watch its guesses snap onto the truth.
Press train. The web of light in the middle is the network; the cloud on the right is every school, its dots falling onto the diagonal as the model gets it right; the curve is its error, dropping live.