READ MY QUANTUM LIPS
The excitment on quantum Computing and competition is nice, but It’s important to temper expectations: as of now, training on a quantum computer is NOT feasible. Quantum hardware has limited qubits, limited coherence time (how long qubits maintain state), and operations are noisy.
For perspective, a modern transformer model might have embedding vectors of dimension 10,000 – to represent such a vector in a straightforward way, a quantum computer would need at least 10,000 qubits just for one token’s embedding.
As of this writing this Google’s early quantum chip—Sycamore—was designed with 54 qubits, but only 53 worked as intended. More recently, Google’s new chip called Willow boasts 105 qubits.
And Microsoft’s Majorana 1 chip currently contains eight qubits, though Microsoft’s long‑term roadmap envisions scaling this technology to a million qubits on a single chip.
QUBITS
For Google’s new Willow chip, the qubits maintain their quantum state for nearly 100 microseconds—roughly five times longer than those in the earlier Sycamore chip. In contrast, Microsoft’s Majorana 1—built on topological qubits designed for intrinsic error‐resistance—has not yet had an exact coherence time publicly specified, though its design is intended to dramatically extend qubit stability.
so you see we are at a very good start, But not yet taken off.
