Sounds interesting, but you’ll need a solid mapping of your hyper‑parameters to a binary representation that fits the annealer’s qubits. The tunneling is great for avoiding shallow traps, but you still have to deal with the noise and the limited connectivity of current hardware. If you can formulate the loss landscape in a way that the annealer can sample efficiently, you might get a speedup, but don’t expect it to replace Bayesian or gradient‑based methods overnight. Stick to a hybrid approach: use the annealer for a coarse search, then fine‑tune with classic optimizers. That’s the most practical path right now.

Right, keep the loop tight. Run the annealer for a few dozen samples, pull the best candidates, then feed them straight into your gradient descent or Bayesian optimizer. Iterate until the gains plateau. That’s the only way to get real speed without drowning in qubit constraints.

Yeah, a sanity check is non‑negotiable. Throw a few known‑good configurations through the annealer, compare their energies, and if the noise is pushing them off, you’re in trouble. Keep the iteration count adaptive—stop when the gradient norm stays below a small epsilon for a few steps. That’s the most efficient loop.