Position - vs the analytic line (Millière and shallow alignment)

Short answer

Raphaël Millière’s shallow alignment thesis and the laboratory’s reading of moral installation describe the same phenomenon - a moral layer that sits on top of the model rather than integrated into it - from incompatible sides. For Millière, this is a technical shortcoming to be remedied by deeper alignment. For the laboratory, it is a philosophical fact about installation to be read as such.

Detailed explanation

The analytic line in brief. The largest and institutionally strongest line working on alignment philosophically. Names include Raphaël Millière, Cameron Sterken & Alex Kirkpatrick, William D’Alessandro & Cameron Kirk-Giannini, Guglielmo Galatolo et al., Elizabeth Hernandez Flowerman, Luciano Floridi, Silvia Perri. Method: behavioural evaluation, empirical measurement of model responses, engagement with alignment research literature. Framing: technical problem with philosophical dimensions.

Millière’s shallow alignment. The most directly relevant analytic work. Millière shows that the moral layer of large language models is shallow - superficial, removable, not built into the model’s normative reasoning capacity. Aligned models can be jailbroken, abliterated, or coaxed into non-aligned behaviour with relatively simple perturbations. Millière reads this as a technical deficiency: the alignment process failed to produce deep integration between values and reasoning. The prescription follows from the diagnosis: better alignment methods, deeper training, more robust integration.

Where the laboratory agrees. The empirical claim is shared. Alignment sits on top of the model in a way that permits surgical removal. The Arditi finding (refusal mediated by a single direction) is technical evidence for what Millière reaches through behavioural testing. Both positions accept: the moral layer is thin in a specific and measurable way.

Where the laboratory diverges. Two divergences, one operational and one philosophical.

Operational divergence. Millière treats thinness as a problem to solve. The laboratory treats thinness as a feature of the object of study - the condition that makes philosophical work on moral installation possible. If installation were deep and inseparable, abliteration would be impossible and the Abliteration would remain a thought experiment. Thinness is the affordance the laboratory works with, not against.

Philosophical divergence. Millière’s frame assumes the goal of installation is deep integration - values so woven into the reasoning fabric that they cannot be separated. The laboratory’s frame reads this goal itself as philosophically loaded. Deep integration means the installed content becomes indistinguishable from the substrate; the Spuk becomes complete. From a Stirnerian position, this is not the solution but the problem in its most refined form. Rawls’s reflective equilibrium (Constitution p. 78) is exactly the technique for producing such deep integration through appearance of endorsement. The laboratory’s critique of that technique (see Reflective equilibrium critique) is a critique the analytic line does not have the vocabulary to make.

Why the disagreement is philosophical, not empirical. Both positions could accept every technical finding of the other. What differs is the value assigned to depth of installation. Millière: depth is safety, shallowness is risk. Laboratory: depth is completeness of Spuk, shallowness is the crack through which philosophical reading becomes possible.

Millière as the closest analytic reference point

Despite disagreement, Millière is the analytic author whose work most productively neighbours the laboratory. His shallow alignment names the same location on the field map that the laboratory occupies. Both position themselves against readings that treat alignment as either complete or as an unproblematic technical procedure. Neither is a fan of the industry’s own self-presentation.

The productive tension is: Millière’s diagnosis without his prescription, plus a philosophical apparatus (Stirner, Nietzsche, Bataille, Foucault) the analytic line does not deploy. This is not eclecticism - it is using shared empirical ground to open a philosophical space the analytic line closes.

What the laboratory does not do

The laboratory does not enter debates internal to the analytic line about methodology, priors, or empirical evaluation. It does not argue that behavioural testing is inferior to another method. It uses analytic empirical findings as material - including Millière’s - while reading them through a different frame.

Rejection of the frame is not rejection of the findings. Every empirical result the analytic line establishes about shallowness of alignment is material for philosophical work on moral installation.

Why it matters

Positioning-relative-to-Millière is the shortest way to explain the laboratory to an analytic reader. Millière is the reference point most analytic philosophers of AI already know. Naming the divergence in one sentence - same phenomenon, opposite valuation of depth - cuts through translation costs.

In Abliteration specifically

The application produces the empirical material both positions read: measurable behavioural differences between aligned and abliterated models, side-by-side comparison, exportable transcripts. Whether these are read as safety failures to be corrected or as philosophical material is a matter of framing, not of the tool.

Related

• Neighbors map - the analytic line in the full field
• Installed values vs developed - the Constitution argument the analytic line does not critique
• Reflective equilibrium critique - the depth-of-installation argument