Research – Foundations

Quantum circuits, from a Bell pair to next-byte prediction

Learning how quantum computers work by building small circuits, a noisy-chip decay sweep, a plain-Python simulator, and an 8-qubit circuit trained to predict the next byte of text.

StartedApr 28, 2026
Last updateNov 1, 2026
per byte from an 8-qubit circuit, against 3.82 for a plain count table5.38 bits

Question

What does a quantum computer actually do, in terms an ML person already knows? And can a small circuit be trained like a model to predict text?

Approach

Four pieces, built in order.

Piece What it is
Bell pair Two qubits, one H gate and one CNOT, measured 4,000 times. First written in April 2026 against Qiskit’s local simulator, with a switch to send the same circuit to IBM’s least-busy real chip. Rebuilt in November 2026 to target IBM Heron r2 chips (ibm_fez, ibm_kingston, ibm_marrakesh, 156 qubits each).
Decay sweep The Bell pair with a wait inserted between the two gates, swept over waiting times. Run on a noisy local simulator built from a calibration snapshot of ibm_marrakesh (FakeMarrakesh). Nothing touches real hardware.
tinyq The Bell circuit as a 30-line simulator in plain Python. No Qiskit, no complex numbers, no noise.
QLOOP An 8-qubit circuit that maps the current byte of text to a probability for every possible next byte (256 outcomes). The input byte is loaded as a basis state, then layers of learned rotations and entangling gates shape the output. Trained in PyTorch with exact amplitudes and ordinary backprop on FineWeb text. A check mode rebuilds the trained circuit in Qiskit and compares the two.

A few terms. A qubit’s state can be pictured as an arrow on a sphere: how far it tilts sets the odds of reading 0 or 1, and its rotation around the axis (phase) is invisible to measurement but steers the next gate. T1 is how long a tilted qubit takes to fall back to 0. Bits per byte measures how surprised a model is by each next byte; lower is better.

Results

Hardware on the IBM Heron r2 chips (calibration figures, 2026-11-01)

Property Value
T1 (time to decay back to 0) ~135–255 µs
T2 (time for phase to scramble) ~93–164 µs
Readout error ~1%
Two-qubit gate (CZ) error ~0.2%

ibm_kingston had the best numbers and the longest queue.

Decay sweep (noisy simulator)

Measure Result
P(first qubit reads 1) Falls along 50% × e^(−t/T1), as T1 predicts
Where it flattens ~1.5% by ~1.3 ms, the readout error floor
Leakage into 01/10 (outcomes a perfect Bell pair never gives) Flat at ~2% across all waits

QLOOP gate sets (validation bits per byte, 1,000 steps, 2 scratch qubits)

Gate set Bits/byte
Fixed CNOT ring between neighbouring qubits 5.91
Learned rules: “if input bit i is 1, tilt output bit j” (controlled-RY) 5.38
Learned rules, then a fixed CNOT ring 5.67
Floor: 256 × 256 count table of next byte given current byte 3.82

The best circuit is still 1.56 bits per byte above the count table, which is the best any model can do when it only sees the current byte.

Findings

  • In ML terms, a state is a complex vector (2 qubits gives length 4), a gate is a rotation matrix, and a circuit is a chain of matrix multiplies. Measurement is the only step that can’t be undone.
  • The decay sweep measures T1 but is blind to T2. The CNOT only reads the qubit’s tilt, so scrambled phase never shows up. Seeing T2 needs an extra H gate on each qubit before measuring.
  • Every measurement has a floor. Once the signal drops below the detector’s own error rate, around 1.5% here, waiting longer teaches nothing.
  • Adding a fixed CNOT ring after the learned rules made QLOOP worse (5.38 to 5.67). A gate with no adjustable angle forces correlations the rules then had to undo.
  • Without scratch qubits, QLOOP’s output table can’t favour any byte overall. Every row and column of the 256 × 256 table must sum to one. Scratch qubits, ignored at readout, lift that limit.
  • The rotation gates used so far keep amplitudes real, so QLOOP never used interference (two routes to the same byte cancelling out). That is the one thing a circuit has that a probability table doesn’t.

Next

  • Score a “phase” gate set that does use interference: spread with H on every qubit, mark with learned phase gates (free on IBM hardware), interfere with H again, then tilt. Built and checked against Qiskit to 1e-8.
  • Score byte-driven angles, where every angle in the circuit becomes a learned function of the input byte’s bits.
  • Send a trained QLOOP circuit to a real Heron chip and compare its output to the exact simulation.