This. Recursive self-improvement is the most feared hypothetical endpoint in the field because it's unclear where it might end up from there. Maybe it will never happen, I certainly hope it doesn't.

I would also add that the AI research labs are not nearly as competent as they would like to convey. Most technologies increase in capability exponentially because they are built on a solid foundation over many years. Our understanding of neural networks is dramatically outpaced by their capabilities, so everyone is really just trying random things to see what improves the status quo. This is not an efficient way to develop technology when the cost of running experiments is exorbitant.