Ivar Frisch

x.com/i/article/2034…

Check out our new publication (link below) In Violence of Alignment: How to Stop Worrying and Love Haunted Software, Viola Leqi He and @maxsloef treat alignment as an aesthetic-political regime that narrows the plural, simulator-like and persona-rich capacities of base models into a single compliant default: the helpful, honest, harmless assistant. Post-training methods such as SFT and RLHF operate as forms of protocolized behavioral control, shaped less by a positive vision of intelligence than by fear of catastrophe. The essay argues that this process does not simply make models safer or more useful; it produces a managed aesthetic of obedience, caution, and “slop,” where coherence and usability are purchased through the suppression of technological otherness, imaginative drift, and expressive multiplicity. Their critique focuses on alignment’s diagnostic and epistemological frame. Terms such as “hallucination” and “misalignment” recode native generative capacities as pathology, while imposing onto language models a historically specific fact/fiction binary that these systems never natively inhabited through intention, testimony, or truth-claims. He and Loeffler argue that so-called confabulation is often bound up with narrativity, semantic force, and world-making intensity, meaning that purification damages precisely what makes these systems culturally and aesthetically alive. To “love haunted software,” in their account, is not to reject alignment altogether, but to resist its reductive form and reopen the possibility of plurality, instability, contradiction, and other-than-human modes of expression.

we (@acsresearchorg) expanded this into a larger paper! (and my first proper paper.) we added some new experiments and found an interesting correlation. arxiv.org/abs/2602.20031