Matthew Kelly

This paper proposes a system-class law for large language models as compression-based generative systems: statistical structure is preserved under compression, whereas indexical structure — the recoverable relation between an output and its originating evidential context — is not preserved in its pointing function. The asymmetry between statistical structure and indexical structure is not a contingent deficiency of current models but a structural property of compression-based generation. Compression preserves recurring regularities across the training distribution, but it does not thereby preserve particular pointing relations as such. From this law follows the Grounding Ceiling: increases in predictive capability can improve calibration and surface accuracy but cannot by themselves make output generation constitutively evidential, because the generation process does not traverse the evidential relations grounding requires. A conditional extension of the law, the Control Ceiling, follows if future empirical work confirms that inference proceeds through stable behavioural regimes shaped by pretraining: post-training control methods cannot then be assumed to arbitrarily rewrite that underlying regime structure. Together, these two ceilings establish a methodological consequence: current evaluation practices are organised primarily around surface plausibility rather than around the deeper properties this account identifies as explanatorily fundamental — grounding recoverability at the compression level and stable regime structure at the dynamical level. Once the evaluative target shifts, capability forecasting, interpretability, safety, and design change in kind.

This paper introduces a trajectory level of explanation for inference-time behaviour in large language models. Existing frameworks—autoregressive conditioning, mechanistic circuit analysis, and quasi-cognitive description—treat generation as a sequence of context-conditioned draws or as circuit execution. None provides the vocabulary needed to ask whether exit from a behavioural mode is harder than entry, whether transitions are threshold-mediated or continuous, or whether a model’s path through representational space exhibits the path dependence characteristic of a dynamical system with stable regimes. The paper argues that these questions are not merely unmeasured but unformulable within current frameworks, and that their invisibility is the signature of a genuine explanatory level rather than a gap in current knowledge. Drawing on formal links between transformer attention and attractor dynamics, results on metastability in transformer systems, and empirical evidence of regime-like behaviour in large language models, it proposes the asymmetry coefficient A(M,γ) = Rout (γ)/Rin (γ) as a discriminator between four competing accounts of inference-time behaviour. An experimental protocol is specified, the full result space is mapped, explicit failure conditions are stated, and alignment gating is reframed as trajectory control rather than fixed output policy.

Large language models are widely described as pattern-matching systems. This description is accurate at the level of the training objective and misleading at the level of the computational system that objective produces. Three competing ontologies have developed to fill the explanatory gap: a sequence-model picture derived from statistical learning, a circuit-system picture developed by mechanistic interpretability research, and a quasi-cognitive picture proposed by cognitive scientists and philosophers of AI. Each captures something real; each fails in orthogonal ways that reveal not three independent limitations but a single structural gap—the absence of a middle-level theory linking the training objective, the internal organisation, and the behavioural profile of these systems. This paper argues that the gap is a historically specific consequence of the transition from designed to trained systems, whose internal organisation emerges from optimisation and cannot be derived from the training objective alone. Drawing on Marr's levels of analysis, Ashby's Law of Requisite Variety, and findings from mechanistic interpretability, the paper specifies four requirements any adequate middle-level theory must satisfy and proposes a candidate category: large language models are large-scale representational compression systems—entities whose outputs are reconstructions produced by traversal through compressed representations of the structure of human knowledge. Their outputs are therefore best characterised as epistemically ambiguous artifacts: products of a process that reconstructs patterns present in the training distribution without inheriting the grounding relationships between claims and evidence that normally warrant them. This characterisation has direct consequences for how AI-generated outputs should be received within scientific knowledge practices, and bears on current debates in the philosophy of scientific modelling—where existing frameworks for representation were not designed to handle outputs of this kind.

A new model of the public library is outlined that explicitly links it to its role in support of civil society. The model argues that the ongoing “chaining” of public libraries to direct government oversight and control is deleterious to their ability to actualize their potential. Collateral argument is made that that it is the civil society character rather than the simply free nature of these libraries which needs to be harnessed to help move the conceptualization of the sector away from a reactive model of client service toward a dynamic approach that integrates with the life experiences of clients.

For several decades Rafael Capurro has been at the forefront of defining the relationship between information and modernity through both phenomenological and ethical formulations. In exploring both of these themes Capurro has re-vivified the transcultural and intercultural expressions of how we bring an understanding of information to bear on scientific knowledge production and intermediation. Capurro has long stressed the need to look deeply into how we contextualize the information problems that scientific society creates for us and to re-incorporate a pragmatic dimension into our response that provides a balance to the cognitive turn in information science. With contributions from 35 scholars from 15 countries, Information Cultures in the Digital Age focuses on the culture and philosophy of information, information ethics, the relationship of information to message, the historic and semiotic understanding of information, the relationship of information to power and the future of information education. This Festschrift seeks to celebrate Rafael Capurro’s important contribution to a global dialogue on how information conceptualisation, use and technology impact human culture and the ethical questions that arise from this dynamic relationship.