Epoch Idle: AI Training Sim

You are handed a single neuron and a dataset it cannot fit. Epoch Idle is an idle game about the real loop of machine learning - scaled down, sped up, and made tactile. Every purchase runs a slice of actual gradient descent. Every number on screen is live math, not a counter with a skin. TRAIN Start with one perceptron and straight-line data. Watch the loss curve bend as you tune weights, learning rate, and hardware - then hit the LINEAR // CEILING wall and learn why activation functions exist. DEEPEN Unlock ReLU, Sigmoid, and GELU. Add Conv2D kernels for segmented data and fight the overfitting they invite. Route multi-head attention over tangled regions no kernel stack can tile. End in a one-way Transformer paradigm shift. PRESTIGE - THREE DEEP • Backprop Step: collapse the run into ∇ gradients and spend them down the optimizer tree (SGD → Momentum → RMSProp → Adam). • Checkpoint Freeze: bank every mastered dataset as permanent Model Weights ω. • Paradigm Shift: reroute your multipliers into the Transformer regime. LEARN WITHOUT NOTICING Loss ceilings are real formulas. Overfitting follows the real 2σ² estimation law. Every paper you publish arrives with its real abstract - from the perceptron to "Attention Is All You Need" - and 28 tappable concepts explain the math behind every system you use. FAIR BY DESIGN • Zero forced ads. The one rewarded ad is an optional ×2 TPU Overclock. • Offline catch-up up to 24 h (extendable to 48) - deterministic, chunked, honest. • Full tutorial that fades when you've proven the mechanic. • No accounts, no cloud, no tracking: your run is a local file you can export from Settings. Train, publish, shift paradigms - the gradient never sleeps.