It starts with what they already claim to be doing - increasingly relying on existing models in non-trivial work related to training, evaluating and optimizing the next, more capable generation of models. As long as the proportion of work keeps shifting towards agents doing more and more of it, and humans less and less, that's RSI at play.
It may be that it turns out LLMs lack some fundamental level of judgement and it plateaus, but frankly I find this notion absurd; LLMs already show better judgement than most people. The alternative is, at some point LLMs will show the ability to futz their way into improvement of the next generation of models even without humans in the loop - even if much less efficient at first, if generation N+1 is more capable than generation N, it'll either take off or burn out.
just because more and more agents are doing human work, that in no way means the model somehow becomes magically more intelligent, it just means the work will stall and continue on at the same level forever
hell even if they hypothetically have an internal model that can output the entire training data set in a better format, there's no scientific evidence that the newer format has new information that is sufficient enough to train a better AI
as a matter of fact the scientific evidence is on the contrary
All intelligence, LLM or not, is bound to plateau around the point where the need to operate within physical reality bottlenecks the speed of feedback. AI is progressing swiftly in the digital realm where feedback is nearly instantaneous, but it's unclear whether that would translate into improvements in the physical world where signals are much noisier and intelligence and judgment are less impactful.
My go-to example: wheels suck at mobility in the natural world. There's a reason no animal uses wheels as their means of locomotion. They only really work on flat, hard surfaces that give a good grip.
Our solution? We didn't build mechanical legs for all-terrain mobility. We paved the world instead. Dirt roads, then brick roads, then asphalt and concrete - since ancient history, we were forcing the world to adapt, so the one means of mobility we could easily built was effective.
We repeat this pattern every time when the world doesn't agree with us, and adapting ourselves to it is too much of a hassle.
Now, AI will likely adopt the same approach to these kinds of problems, which may not end up very nicely for us.
(Note: we are already adapting our own digital worlds to make them easier for AI. Much like building roads for our wheels, we now have a resurgence of CLI, with new tools exposing high-level operations optimized for agents - not humans - to use comfortably.)
With AI it's a bit of a chicken or the egg situation. Before it can ablate the complexity of the world, it needs to be able to deal with it, including adversaries such as ourselves. For this it needs a tool as sophisticated as our bodies, but could AI develop such a tool for itself without the help of this tool? I'm not convinced.
So far, I don't think the models are capable of running away on their own. Of course, it would be playing with fire to not at least consider the risks of such a runaway scenario and build in safeguards against it. But, there is no model that can build a better model on its own, thus far, to the best of my knowledge (which is far more limited than the models, so maybe I should ask them).