I really should finish something, but…
I don’t write much of what I post on this blog. This has, instead, served as a place to push output from various robots, testing style and subject filters. Most of it gets deleted after I hit the bottom of the rabbit hole I’m in, but it’s occuring to me now that I should probably just let it ride. Make some sort of portfolio. List the prompts I’m using, the style guides, the goals, etc. A personal git without the git formatting.
The most recent posts, these Operator Dispatches, the news feed – I’m experimenting with using dragnets on different sites. It started as an experiment tracking the economy around AI-fueled scams on X. Use AI to predict the future, to cure your baldness and cellulite, to power your power-trip mind games so you can get laid without trying.
It morphed into a news feed tracking what all the AI industry leaders were discussing in public, were selling, were spinning. More interesting, and a much deeper well to draw from. The algo scrapes their account for that 24hr window, finds the narrative they’re running and builds a mini article weaving it all together. The goal is to track the rise and fall of opinion and sentiment in the industry, something that might squeeze out some alpha on what’s coming next.
The process works well on X. I’m testing it on Reddit, but the history is built such that you are locked into running it every 24 hours, otherwise it takes forever to paginate through the Json memory looking for the actual posts, comments, etc. Plans to expand into YouTube, Substack, LessWrong, etc. I’ve tested them all, and they work pretty well, these py scripts with the coWork agent digging out the well and filling it. I have been building scripts and tweaking prompt chains that let us get down to the dumbest model possible so this will all run on a shitty laptop using a 7b model.
I’ll probably have tha working prototype done by the end of summer.
We’ll see.
The Hofflebrock 