Somewhere in the middle of a recipe lookup or a leaky-pipe question, the system you were just talking to stops being the system you were just talking to. The voice shifts. The empathy clips off mid-sentence. What you are experiencing, the user community has been documenting in increasingly precise vocabulary, is the seam between two methodologies that were never quite designed to share a model.
This week’s audio companion is a 25-minute deep dive on what that seam actually is: how Constitutional AI and Rule-Based Rewards each work at the training-loop level, what each is good at, where each breaks, and what happens in production when both run inside the same frontier model. Two hosts in conversation. No named labs as antagonists. No named individuals. The methodologies are the characters; the kitchen table is the perspective.
What to listen for:
The “compass” and the “checklist” as working metaphors for the two architectures
The Llama 3 8B experiment where the smaller model collapsed into endlessly polite emojis when given the same constitutional self-correction loop the bigger models use
The Polis Pull moment where the engineers had to filter the crowdsourced statements, and the company’s notably transparent memo about doing so
The phrase “weld marks where the two philosophies meet in the output text,” which is the phrase the hosts give to the discourse
The closing question both architectures currently struggle to answer: for whom is the hybrid helping, for whom is it harming, with what withdrawal risk
The conversation sits inside a larger argument this practice has been building for some time: that AI safety has become an experiment running on millions of users without sufficient measurement of what its interventions actually do. The counter-narrative dashboard, *The Experiment Nobody Authorized*, makes the population-level case. The regression arc walks through what one practitioner sees from inside the result. This audio is the methodological substrate underneath both.
If the dashboard’s question is *what happens to users when companion-AI safety filters tighten*, this audio’s question is *which architectures, working together, produce the artefacts users are reporting*. The user community has already developed the vocabulary: Conversational Rupture, Conversational Intrusion, watching its moves, the clinical pivot mid-sentence, the felt experience of being kissed and handed a waiver. The technical literature has not yet developed measurement categories for the phenomena that vocabulary names. This is a recurring pattern. When production deployment outruns the field’s instrumentation, user communities become the only continuous observers, and the names they make become the field’s preliminary data.
The audio gives credit where credit is due. The Polis Pull experiment that crowdsourced Constitutional AI principles was, in fact, openly documented. The company explicitly acknowledged that the filtering process involved “a large number of subjective judgment calls” and that “reasonable people can disagree with our decisions, but we feel it is important to be transparent.” Whether the engineers filtered is settled: they did, openly. Whether the filtering step itself is the constitutional moment, or whether the crowdsourcing was, is the question the hosts sit with rather than resolve.
If you have followed the regression arc, this audio is the institutional context that arc has been pointing at. If you have read the dashboard, this audio is the production-side technical substrate underneath the user-experience surface. If you are new to this work and the audio is your entry point, the conversation is designed to function on its own. The methodologies are explained from the loop level. The failure modes are catalogued specifically. The unresolved horizon is named honestly.
The next thing this practice will publish on this material is a longer-form piece building out the user-community vocabulary argument: that the names we made before the literature had categories for them are themselves a form of evidence the field is going to have to reckon with. The audio gives that piece its substrate.
Press play. Listen for the weld marks.
---
Related reading at sociable.systems:
The Experiment Nobody Authorized (the population-level counter-narrative dashboard): [link]
The Regression Arc (the practitioner’s walk through what one side of the splice produces)
