Quantum autoencoders compress a family of quantum states into fewer qubits by training a parameterized circuit so the discarded "trash" qubits collapse toward |0⟩ — the same idea as a classical autoencoder's bottleneck, but the cost function and gradients come from real quantum mechanics. This simulator runs a genuine 2-qubit variational circuit against the canonical training family cos(φ)|00⟩+sin(φ)|11⟩, computing gradients with the parameter-shift rule used on real quantum hardware and updating the parameters with gradient descent. Watch the batch of trash-qubit Bloch vectors drift from scattered points toward the north pole as compression fidelity climbs, while the latent qubit keeps encoding the original information for any φ you select.