There was some talk about this on the forum last year and it referenced the very same video: AI (Artificial Intelligence)
Applying AI to KiCad is probably best done right now using external scripts applied to KiCad files, or within a plugin running in PCBNEW. Integrating into KiCad itself would probably exceed the bandwidth of the dev crew while slowing the introduction of more commonly-desired features.
AI doesn’t require the use of LLMs with multi-billions of parameters. Protein folding was solved using neural nets 1000x smaller than LLMs. And even if you did need a large LLM, the costs for running them have dropped precipitously over the past few years.
You also don’t need a lot of (or even any) training examples. You just need a way to evaluate the quality of the result. This was done with AlphaZero that trained itself to play Go by evaluating the results of its moves. You would need to build a function that would evaluate a PCB or circuit, and then you could use that to train an AI system. This is not trivial, but it gets around the paucity of high-quality PCB examples.
There are a few companies trying to apply AI to circuit design: JitX, Quilter.