The billion dollar question is whether this works "out of the distribution". I.e., whether LLMs can only find and use the specific ideas buried in training data, or whether they can learn to apply the "thinking process" to a new problem. IMO this is still unanswered (due to these recent controversies).

But regardless of the answer, it seems we have a planet-scale positive feedback loop here. LLM became good (enough) by training on generally available data (books, internet, github) + RLFH, so experts tried to use them on hard tasks, which required lots of hand holding. These conversations became part of the training data, and the next generation of frontier LLMs were better. So, more experts used them on harder tasks, again requiring hand holding. These conversation became part of the training data... etc.

In a nutshell, top human minds across the world are pouring their skills into LLMs just by using them. This is not "continuous learning", but if you re-train on the most recent sessions every, say, quarter (which seems to be happening?) you get close to that in practice.

In modern RL, we also train deep nets on some (often non trivial) loss function. And RL is generating its training data. Hence, it blurs the line with SSL. I'd say, however, it's more complex and more computationally expensive. You need many / long rollouts to find a signal to learn from. All of this process is automated. So, from this perspective, it blurs the line with UL too :-) Though it dependence on the reward is what makes the difference.

Overall, going from more structured to less structured, I'd order the learning approaches: SL, SSL (pretraining), RL, UL.