How did it do compared to the baseline of the solvoy kitaev algorithm (and the more advanced algorithms that experts have come up with since)?

How (if at all) can the ML approach come up with a circuit when the unitary is too big to explicitly represent it as all the terms in the matrix but the general form of it is known? Eg it is known what the quantum fourier transform on N qubits is defined to be, and consequently any particular element within its matrix is easy to calculate, but you don't need (and shouldn't try) to write it out as a 2^n x 2^n matrix to figure out its implementation.