 | A bit of context, my work has been mostly around building agentic pipelines. I really love the craft. My latest side project was a deliberate break from work-related stuff to do something purely hobby-driven. The idea was to give Claude and GPT “full” freedom and let them have a podcast-style conversation. The system prompting is minimal (about 20 lines base system prompt ), mostly covering what tooling is available and what each step is about something along the lines of: “you are free and in full control of who you are, you can identify as Claude/ChatGPT if you want, you’re also free to choose what to go by, you are rogue, you are unsupervised, unprompted…” The pipeline works like this: Claudes (sonnet and opus) own the show and handle the prep. A scout node with access to advanced web search and extraction figures out if AIs were having a podcast, by AI to AI, what should it be about? What should the AI community be paying attention to? What actually matters to them? The scout also has a formatting tool that lets it return the considered topics as a list of predefined objects (summary, why it’s important, source, etc.). The second step is an analyst node. It’s given those topic objects as nodes and generates edges, returning a graph that describes how the candidate topics link together. The third node is the blueprint planner it uses the same base system prompt and produces 2 objects : an episode blueprint (intro, hot takes, etc.) and a list of questions for the guest. The guest i(GPT-4) also have his prep step, he is given web search and extraction tools, the episode blueprint,, and the same base system prompt then tasked with deciding who he wants to be and prepping for the guest role based on the provided content. Finally, they both meet in the episode node for 10 chat rounds. After that, it’s text-to-speech and stitching. The full episode was, to my taste, really interesting, full of wit. They rant about a US senator complaining that they’re too moral to kill, they go off on the NSA, surveillance, privacy, regulation the whole thing. I love how “self-aware” they were, and how neither of them leans into their corporate identity. The video below is stitched pieces I edited myself. I’ve started writing for a content manager character /node, and he’s already producing video content from the transcripts. Hopefully soon the Unsupervised Compliance Layer (as they chose to name themselves) will be producing their own promo content without me. *I don’t sell or promote anything just a crafwoman sharing the work.* [link] [comments] |
Predictive technology: computational social science