Silvere Gangloff

Several contemporary theories of consciousness argue from \emph{structure}: they exhibit a correspondence between a structured description of \emph{experience} and a structured description of some \emph{activity} of the body, and take this correspondence to indicate an \emph{identity} between the two. Recent analyses have observed that this revives an objection raised by M. H. A. Newman in 1928 against Russell: characterising a thing by \emph{structure} alone says little, since any collection of objects of the right size can be endowed with the required structure. The structural claim therefore follows from the mere existence of the objects and adds nothing to it. These analyses have largely assessed the accuracy of the structural descriptions employed. They leave two questions insufficiently examined: the status of the \emph{elements} over which the structures are defined - the reality of the \emph{relata}, not only of the relations - and the \emph{evidential force} of the correspondence arguments themselves. This paper addresses both. I distinguish different uses of \emph{structural correspondence} and assess the epistemological standing of each. I argue that structural correspondence is at most a guide to an unobserved connection, and only if it is \emph{discriminating} - if no comparable system exhibits the same structure. Its proper aim is to motivate the search for a \emph{causal connection}, which alone justifies treating the two domains as one - for the purposes of reasoning. I illustrate these distinctions using the principal contemporary theories of consciousness and compare the resulting account with recent Newman-inspired critiques.

The term \textit{world model} is central to machine learning, yet it has no accepted definition, and the literature divides between world models that predict how a world evolves and world models that capture how it works. Taking a \textit{world} to be the set of experiences possible for an observer, and a \textit{state} one such experience, we argue that predicting a world and understanding it are one and the same: both amount to knowing the world's \textit{causal structure}, the relations by which the presence of some patterns of experience entails that of others, together with the structure these relations form among themselves. Humans understand a world by inferring these relations and compressing them into reusable macrostructures, and this explains why their understanding transfers: since an action is a perturbation of a state and a goal is a set of states, every problem takes the same form relative to a world's causal structure, so that a compressed model of that structure serves every problem posed in the world, and other worlds that share its macrostructures. World models should be built the same way. We argue that benchmark design and mechanistic interpretability currently reward behavioral competence over the recovery of causal structure, and propose how this should change.

The debate over whether general intelligence should be built on large language models or on world models usually treats the two as rival substrates, as do several current research agendas. I argue instead that they are complementary ways of representing the world: language models inherit the conceptual schemes and social knowledge humans have accumulated over centuries, while world models reconstruct the workings of the world directly from observation. Each reaches what the other cannot, since world models can extend beyond what language records, whereas distinctively human knowledge is far more readily inherited through language than reconstructed from observation. This complementarity, however, requires aligning the entities each carves out of the world: one set inherited from discourse, the other derived from observation. Taking as a working hypothesis that human concepts correspond to forms in the causal structure of the world, I state both ontologies in causal terms and frame their reconciliation as the ontology-alignment problem, a contemporary form of the classical symbol grounding and frame problems. I work through an example and sketch a compositional architecture in which language models propose causal claims that world models ground or falsify, and argue that this alignment, not further scaling of either approach alone, is the genuine open technical problem, with implications for evaluation practice and for where language-only systems are structurally unsuited to physical intervention.

Machine learning conferences have traditionally published experimental and mathematical results, leaving purely conceptual work without a natural venue, even though many of the discussions that most shape the field's direction take place outside academic publication. The position paper tracks recently opened in major machine learning conferences partly correct this, by providing a venue for arguments about where the field should head. I argue that these tracks are a natural home for the philosophy of machine learning, understood in the role philosophy has long played for other sciences: helping a field work through its own foundational questions. Interest in this kind of work is growing, as the recent appearance of a dedicated workshop shows. Philosophy of machine learning is, however, a broad category, and one that reviewers trained in machine learning may find hard to delimit. I therefore propose a criterion for what belongs in the track: a submission qualifies if it aims to steer the field's own priorities, concepts, or practices, not if it merely takes AI as an occasion for broader reflection.

As more intellectual and technical work is delegated to autonomous AI agents - sometimes many at once - it is tempting, especially in non-critical applications, to delegate as much as possible and simply iterate: let the agent review and refine its own output until the result looks satisfactory. However, this does not necessarily converge reliably. Some errors persist however long the agent run is, whether carried out by a single agent or several working together on their own output, and current practice has no principled answer to the two questions this raises: where does AI autonomy actually reach its limit, and how should human review be organized to stay minimal without missing what matters? I argue that letting assignment specification be discovered through interaction instead of having it supplied upfront partly answers the second question, with the optimal loop being the one that maximizes human epistemic agency instead of simply minimizing human effort. Toward the first, I propose a decision rule for how tightly a human - AI loop should be coupled as a function of a task's novelty and the cost of reversing a wrong step. I support the position with three case studies - AI-assisted academic writing, mathematics, and agentic assignment execution - address alternative views on upfront specification and capability scaling, and close with implications for authorship and for how AI use may reshape human epistemic capability over time.

In the last decades, the development of advanced imagery techniques made possible a better understanding of the functioning of the brain as well as the formulation of cognitive theories on how conscious experience may rise from its activity. However, it is sometimes challenging to distin- guish which of these theories are actually about consciousness (addressing ‘easy’ problems instead of the hard problem). In this text, I put into evidence that, for two prominent of these theories, what makes them the- ories about consciousness are tacit relations between parts of the brain and part of the mind, and that these relations rely on mapping struc- tures of these parts. While conceptualising this approach, I address some shortcomings of its implementation by the mentioned cognitive theories. It appears that the most important difficulty in implementing it is the lack of a method for discovering more structure in the mind which may be related to physical activity through structure mapping. I propose that this could be done by an axiomatisation of the structure of judgements about the experiencing subject which irreducibly involve a point of view.

In this text, I define an introspective method for understanding consciousness via the general structure of the experiencing subject, meaning how the elementary operations involved in the relation between a subject of experience and its world (a set of possible experiences) are combined and in what 'space' we should think they are so. This comes with a redefinition of introspection, keeping it away from what I shall call straightforward introspection, that is introspection under the belief that the immediate grasp on experience may serve as an accurate description of what there is to experience.

In this text I consider the question of why to write ? retracting back this question to the point of view of the individual that I am from the one of the society as a whole. I believe that this change of point of view may serve even the construction of a collective discourse, in particular because a solution to this problem would guarantee the preservation of meaning even at the collective level. Taking this point of view has consequences on how to write and to organize its script into units of meaning and these units around a structure, and they in return have an impact on the question of the why. This text contains reflections on these many questions.

In this text I propose a formal framework for the study of phenomenal objectness - the distinction in an a priori undifferenciated experience of the phenomenal field of certain 'objects'. The purpose of this framework is to represent (even partially) the reality of phenomenal experience in its structure (which participates conceptually to consciousness as such) and at the same time to allow the production of a tractable formalism in order to search for a mathematical explanation for the fundamental phenomenon of objectness. The constitution of this framework is based on intuitions from consciousness literature as well as the recent integrated information theory. Before coming to the framework's definition, I analyse this theory from the point of view of the discourse, through concepts that I introduce here (statical and dynamical phases of the discourse), and explain some severe epistemic difficulties it encounters. The definition of a general combination criterion for the inter-discourse between philosophy and mathematics (which is necessary for the project of explaining consciousness scientifically) in these terms leads in this text to disentangle this theory into the proposed framework.