Can Quantum Machine Learning (QML) actually provide a speedup for training LLMs?

All answers to this question.

Currently, we are still dealing with the "input-output bottleneck." Getting classical data into a quantum state (quantum embedding) is slow and expensive. While Quantum Kernels can theoretically find patterns in high-dimensional spaces that classical computers miss, we aren't at a stage where we can train a GPT-sized model on quantum hardware. Most research right now is focused on "Hybrid Quantum-Classical" models where the quantum part only handles a specific, highly complex optimization sub-task.

   Answered 2024-08-16 by Sarah Harrington

Are you following the work on "Quantum-inspired" classical algorithms? Sometimes just thinking like a quantum physicist leads to better classical optimization tricks.

   Answered 2024-08-17 by Kevin Foster

• Kevin, I’ve seen some of the tensor network research! It’s fascinating how those "quantum-inspired" models are outperforming standard GPUs in specific recommendation engine tasks. It makes me wonder if we will see a hybrid chip architecture in the near future where a QPU acts as a co-processor alongside the GPU for these specific bottlenecks.

     Commented 2024-08-18 by Arthur Miller

We’re still in the "proof of concept" phase. QML is great for chemical simulation and drug discovery right now, but NLP is likely several years away.

   Answered 2024-08-19 by Emily Dawson

• True, Emily. The state of the art is really in material science. For those interested, looking into VQE (Variational Quantum Eigensolver) is the best place to start.

     Commented 2024-08-20 by Sarah Harrington