A few years ago it was discovered that LLMs can implement gradient descent in their processes. Producing a full trained neural network would be an extension of this.

See: https://arxiv.org/abs/2212.07677