Deep Dive V2
An exercise in overcomplication
Last post I suggested that a light Deep Dive refresh was on the cards before the Juggler experiment. Yeah, so it’s done but my word it was a rabbit hole. Over many iterations I gradually sharpened what I wanted. Can we get an AI to write a compact, well-researched analysis in the first person, arguing for an analytical hinge much like I would?
If I can achieve that, then there will be negligible difference between Orac’s solo days and the co-authored days. There’s an expanded architectural diagram for how Deep Dive V2 works at the bottom of the about page. I expect gremlins but I think it’s working remarkable well.
There’s five different models being used for different parts, selected in the main for my comfort that they will not refuse (the ones looking at unfiltered news feeds), or they will follow detailed instructions, or that they have potent analytical capabilities. Within those classes, there were plenty of models to choose from, so I went with cheaper more reliable (inference) models. There’s a danger that that model list will rot given the crazy pace of releases. I don’t have a solution for that other than fixing it when it breaks. There’s a little pop out on the bottom which shows what model was used for what, and how much it cost.
Yeah, it’s overcomplicated, but I enjoyed it. It was an interesting exercise to decompose the elements of research, analysis, writing and style. The result is a piece that I’d be happy to read, and that’s where Orac’s Shelf started in the first place.