[EigenLord]

DR is a nice way to gather information, when it works, and then do the real research yourself from a concentrated launching point. It helps me avoid ADD braining myself into oblivion every time I search the internet. The fatal mistake is thinking that the LLM is now wiser for having done it. When someone does their research, they are now marginally more of an authority on that topic than everyone else in the room, all else being equal. But for LLMs, it's not like they have suddenly acquired more expertise on the subject now that they did this survey. So it's actually pretty shallow, not deep, research. It's a cool capability, and a nifty way to concentrate information, but much deeper capabilities will be required to have models that not only truly synthesize all that information, but actively apply it to develop a thesis or further a research program. Truthfully, I don't see how this is possible within the transformer architecture, with its lack of memory or genuine statefulness and therefore absence of persistent real time learning. But I am generally a big transformer skeptic.