I had a long ELI16 session with Claude about it, and the way I understand it is that they

- use Ising machines to describe a certain problem into clauses, storing system state (e.g. spin of something) in variables

- then use a neural network layer where each neuron determines the value of one clause

- then for each state item, use the neuron output to determine if flipping that state would improve the overall system score

- and then use FN-like "noise" to determine whether to flip or no

If the energy landscape of the problem is pretty local, this is guaranteed to find a good solution to the system, using way less compute than brute-forcing it.