There a huge leap between a procedure allowing a predictive model to iteratively decrease False Positive Rate and having an AGI.
Correct. That’s why the section on classification and RL are separate. Classification tasks are a subclass of RL. A recurrent task need not be a classification task. In fact that I’d go further and say there’s still a huge difference between having an agent that can do RL and having an AGI. That’s why I put such speculation at the end.
Having said all that, it seems plausible to me that a language model might be able to reason about what modules it needs and then design them. I implicitly believe this to be the case, but perhaps I could’ve been more explicit. This is more of an empirical question, but if that were possible the difference between that model and AGI would become much smaller in my opinion.