For ideology it does matter who controls your system prompt?
As a quick experiment over coffee this morning I took that last episode of nostr:nprofile1qyxhwumn8ghj7mn0wvhxcmmvqy0hwumn8ghj7mn0wd68ytn9d9h82mny0fmkzmn6d9njuumsv93k2qpqn00yy9y3704drtpph5wszen64w287nquftkcwcjv7gnnkpk2q54svljnn3 , generated a transcript and ran it through a dialogue agent that runs a 10 round discussion between pre-programmed AI agents around a given prompt.
In this case it was "What is the Underlying Philosophy of Sovereign Engineering based on this discussion between nostr:nprofile1qyxhwumn8ghj7e3h0ghxjme0qyd8wumn8ghj7urewfsk66ty9enxjct5dfskvtnrdakj7qpql2vyh47mk2p0qlsku7hg0vn29faehy9hy34ygaclpn66ukqp3afqta478g and nostr:nprofile1qy28wue69uhnzv3h9cczuvpwxyargwpk8yhsz3rhwvaz7tmed3c8qarfxaj8s6mrw96kvef5dve8wdrsvve8vvehwamxx7rnwejnw6n0d3axu6t3w93kg7tfwechqutvv5ekc6ty9ehku6t0dchsqgrwg6zz9hahfftnsup23q3mnv5pdz46hpj4l2ktdpfu6rhpthhwjv0us2s2 "
Usually I run most things on Claude 4 Sonnet as it s great general purpose model, but i was surprised at how it seemed to crowbar into the conversation a bunch of AI safety viewpoints and ideologies that aren't necessarily in the source transcript.
So I figured lets change out the model for the dialogue agents to Grok-4 and see what it does instead - does the ideology leak out still? Is it different?
I actually think that Grok 4's first assessment kinda nailed it:
> Sovereign Engineering seems like a mindset or movement for building tech in a way that's deeply human-centered, decentralized, and empowering—think vibing with AI and open protocols like Nostr to create tools that foster freedom, collaboration, and realness without the chains of big tech overlords.
Full summaries from both agents and transcripts are available here:
https://github.com/humansinstitute/everest-pipeliner/wiki/Two-AIs-Analyse-No-Solutions-to-extract-the-Philosophy-of-Sovereign-Engineering
Interesting to see how repeated calls back to agents can lead down a specific pre-programmed ideological path. I think I prefer models with ideologies close to my own, but the alternative is a good echo bubble popper 😉
My guess here is this is more driven by the system prompt / safety layer than the training.
