A HUMAN VS. MACHINE EXPERIMENT

Outclick the Oracle.

You think you’re random. This 107-parameter brain disagrees. Prove it wrong in 30 clicks.

LIVE NEURAL NETWORK
UNPREDICTABILITY0
MOVE0 / 30
ESCAPE STREAK0
ORACLE HIT RATE—
I have already made my prediction. Go ahead. Surprise me. Pick any symbol. The Oracle predicts first, then learns from your click.
IT REMEMBERS
    Click, tap, or press 1 / 2 / 3. Surprising choices earn more points.

    Best on this device: —

    OPEN THE BLACK BOX

    Every guess. Every weight.

    Untrained brain

    Try repeating Orbit a few times, then switch. Watch the probability rise, the surprise spike, and the weights adapt.

    1. Predict Your last 3 clicks flow forward through the network. It commits to a bet.
    2. Reveal You click and score 100 × (1 − p) points: the less it expected you, the more you earn.
    3. Measure Its loss is L = −ln p. Big surprise, big loss.
    4. Learn Backpropagation nudges all 107 parameters to make your choice likelier next time.

    Live values below are from before the last click’s update, exactly what the network saw when it bet. No future guesses are revealed.

    01

    Remember

    9 inputs
    x = [onehot(c_t) | onehot(c_t−1) | onehot(c_t−2)]

    Each click becomes three switches: Orbit, Prism, Spark. Newest click first; missing history stays 0.

    02

    Mix patterns

    8 neurons · 80 params
    z = W¹x + b¹, h = tanh(z)

    tanh squashes each weighted sum into (−1, 1): a soft “how much does this pattern match?”

    03

    Make a bet

    3 outputs · 27 params
    a = W²h + b², p_k = exp(a_k) / Σ exp(a_m)

    Softmax turns raw scores into probabilities that sum to 1. Its bet is the tallest bar.

    04

    Learn from being wrong

    η = 0.22
    ∂L/∂a_k = p_k − y_k, δ_j = (1 − h_j²) Σ_k W²_kj (p_k − y_k), W ← W − η ∂L/∂W

    Click a symbol to see its loss and a real gradient update.

    ZOOM IN

    Inside one output neuron

    Who voted for it?
    Each bar is one term W·h. Right pushes this symbol up; left pushes it down.

    One real gradient step

    Input → hidden: all 72 weights + 8 biases W¹, b¹
    Hidden → output: all 24 weights + 3 biases W², b²

    positive negativestronger colour = bigger weight biggest changes from your last click

    Each cell shows the weight and, below it, the Δ your last click caused. Rounded to 3 decimals; calculations use full precision.

    Wait. Is it actually learning?

    Yes. Your last three choices become nine inputs to a neural network with eight tanh hidden neurons and three softmax outputs. After each move, backpropagation adjusts its weights using cross-entropy loss. The lines show real weights: blue-green is positive, rose is negative; thicker means stronger. After each click you see the forward pass (small dots running along the weights) and then the learning pass (gold dots flowing backward along the weights that changed most).

    The prediction is locked before your click. The diagram and percentages reveal that past prediction, never the next one. An escape means it guessed the wrong symbol. Points = round(100 × (1 − probability of your choice)); up to 3,000 points across 30 moves. Clicking truly at random averages about 2,000 points, so beating 2,000 means you modeled the model better than it modeled you. Streaks are for bragging rights and do not multiply points.

    Everyone gets the same initial brain for a given UTC day. Shared links preserve that challenge and the score to beat; restarting resets all learned weights. There is no timer. This is a playful experiment in pattern learning, not a test of intelligence or a measure of true randomness. Game choices stay in this tab; only your best score and sound preference are saved locally.