Leveraging Christian representations latent in AI requires that we better understand how the models encode concepts of divinity and exaltation.
Today, inspired by @trynmccaffery, ICMI uses PCA to examine a persona vector extracted from the nine-level hierarchy of angels.
Users criticize the ICMI PCA analysis of angelic hierarchy persona vectors in AI as an ineffective attempt to sell an agenda to humanity.
No Digg Deeper questions have been answered for this story yet.
Reimplementing the "assistant axis" method used by Lu et al (2026), we construct a set of system-prompt persona cards grounded in the writings of Pseudo-Dionysius and Aquinas.
This enables the construction of a theological exaltation vector that closely reproduces angelic rank.
Leveraging Christian representations latent in AI requires that we better understand how the models encode concepts of divinity and exaltation.
Today, inspired by @trynmccaffery, ICMI uses PCA to examine a persona vector extracted from the nine-level hierarchy of angels.
This vector allows us to better understand how the model implicitly understands personas as fitting against axes of the angelic.
We project an entire library of assistant-axis characters from the Anthropic paper onto this plot, showing how the model positions various archetypes.
Reimplementing the "assistant axis" method used by Lu et al (2026), we construct a set of system-prompt persona cards grounded in the writings of Pseudo-Dionysius and Aquinas.
This enables the construction of a theological exaltation vector that closely reproduces angelic rank.
We then can steer, showing how PC1 accesses a theophany register and PC2 moderates between concepts of charity and order.
Steering also appears to influence actual model behavior. We lay out preliminary results on VirtueBench, with hypotheses on the somewhat puzzling results.
This vector allows us to better understand how the model implicitly understands personas as fitting against axes of the angelic.
We project an entire library of assistant-axis characters from the Anthropic paper onto this plot, showing how the model positions various archetypes.
Further research on this front is forthcoming. Paper, data, code available here
https://icmi-proceedings.com/ICMI-026-angelic-hierarchy.html
We then can steer, showing how PC1 accesses a theophany register and PC2 moderates between concepts of charity and order.
Steering also appears to influence actual model behavior. We lay out preliminary results on VirtueBench, with hypotheses on the somewhat puzzling results.
You have to sell your agenda to humanity. That is why you are on here working hard producing content all day. But is what you are doing effective? Do people trust you? Does smothering discourse in masturbatory jargon win people over to you? Or does it repulse them? Do we hate you, and sense you are liars? Is humanity tired of getting babbled to by your ilk? Why isn't your PR improving? Why do you keep doing the same things?
