Realistically, at a certain point the training would likely involve interaction with reality (by sensors and actuators), rather than relying on secondhand knowledge available in textual form.

There's only so much you can learn from humans. AI didn't get superhuman in go (game) by financing more new good human go players. It just played with itself even discarding human source knowledge and achieved those levels.

> What happens when there are no more hosts to donate more training-blood?

LLMs have over 1B users and exchange over 1T tokens with us per day. We put them through all conceivable tasks and provide support for completing those tasks, and push back when the model veers off. We test LLM ideas in reality (like experiment following hypothesis) and use that information to iterate. These logs are gold for training on how to apply AI in real world.