Well, technically, the "AI"s… can generate their own additional training data…

But then trying to train another AI on said AI-generated data… well, then the AI starts to develop toward model collapse, basically, it gets more stupid and incoherent, develops weirder and stronger ‘quirks’.

But yeah, as far as I see it, basically zero chance an LLM advances beyond ‘very fancy autocomplete’ toward AGI or a capacity for actual metacognition, to think about its own thinking and then try and modify that.

Sorry, but you’re not gonna get a super intelligence if it isn’t capable of actually assessing and correcting itself.