Madhur Mangalam

A persistent ambition in artificial intelligence research holds that large language models (LLMs) will eventually become embodied—through attachment to robots, augmentation with multimodal sensors, integration into vision–language–action architectures, or sheer scale until grounding emerges from training data. This ambition rests on a misunderstanding of embodiment. Embodiment is a constitutive condition of being a self-producing, self-maintaining living system whose continued existence is genuinely at stake in its couplings with the environment. Adding sensors, actuators, or physical instantiation to a computational substrate yields engineered coupling, which is something else entirely. We identify five conditions that any genuinely embodied system must satisfy—autopoietic closure, precariousness, constructive self-determination, intrinsic temporality, and mereological closure—and show that no computational artifact can satisfy any of them, let alone all five. The barrier is categorical incompatibility: the same property that makes a system computational (heteronomous design and maintenance through external specification) is the property that precludes it from being embodied. Coupling artifacts to environments through transducers achieves engineered interaction. Having a body and being a body are different conditions. We close by examining why the embodiment-equivalence claim persists despite its categorical impossibility, and by arguing that the confusion sustains two harms—overselling AI systems for tasks they cannot perform, and misdirecting public concern toward speculative futures while the actual harms of deployed systems accumulate without being addressed. Clarifying the categorical distinction is a precondition for taking the genuine stakes of AI deployment seriously rather than as a philosophical aside.

Unilateral hip osteoarthritis (HOA) reshapes whole-body movement, and total hip arthroplasty (THA) restores function only partially—but conventional kinematic analyses fall short of characterizing how the body’s distributed coordination reorganizes. Network analyses of inter-segment kinematic coupling offer a systems-level alternative, but they conventionally collapse the sign of coupling, treating two mechanically distinct relationships as equivalent: segments whose displacement amplitudes covary, rising and falling together, and segments whose amplitudes trade off, one displacing little while the other displaces a great deal. Here, we introduce a signed partial-correlation decomposition of marker-displacement coordination into parallel covarying ( W + ) and trading-off ( W − ) networks, applied to gait from 80 healthy adults and 106 HOA patients assessed before and 6 months after THA ( ClinicalTrials.gov Identifier: NCT01907503). HOA reorganizes the two networks in opposite directions: covarying coordination loses modular structure and small-worldness, while trading-off coordination becomes more locally clustered and more small-world; global efficiency falls in both. This is consistent with pain-driven splinting that consolidates a few fixed trading-off configurations. Surgery acts predominantly on covarying coordination, partially restoring its topology, while the trading-off network’s reorganization persists across the six-month post-surgical window—and on one measure, i.e., how broadly each segment spreads its trading-off partners, a deficit emerges during recovery that was absent before surgery. A node-level sign-balance metric shows that the proportion of each segment’s coordination carried by covarying relationships returns to healthy values by 6 months, even though the underlying networks remain structurally altered. Standard post-surgical care thus restores the gross balance of coordination within 6 months, while the underlying architecture remains reorganized at that mid-recovery time-point—and, on at least one measure, consolidates further. Whether this reorganization reflects protective motor strategies, and whether training targeted at segmental differentiation would reverse it, are hypotheses that the present observational data raise but cannot test.

Twentieth-century medicine was most successful against diseases sharing a particular causal architecture: a discrete lesion standing in a relation of approximate necessity and sufficiency to a clinical syndrome, amenable to specific perturbation. Mortality from acute myocardial infarction has fallen by nearly 90 percent in the United States since 1970; pediatric leukemia has become survivable; HIV has become chronic. That success has shifted the residual burden toward conditions with a different architecture—late-life neurodegeneration, multimorbidity, frailty, idiopathic neuropathy—in which pathology is distributed across scales, and no single node carries the necessity that targeted intervention requires. The failure is not that medicine has extended life. It is that biomedical institutions have not adapted to the forms of vulnerability that accompany longer lives, and the non-adaptation is structural rather than accidental. This paper asks why: why funding and evaluation systems systematically favor research on discrete diseases and individual molecular targets over integrated research on aging as a systemic, multiscale, temporally extended process. We distinguish two levels at which the apparatus fragments its object. First-order fragmentation partitions aging into separately funded age-related diseases. Second-order fragmentation partitions aging itself into hallmarks, pathways, and targets pursued independently—which is why geroscience, the framework formulated to unify the age-related diseases, is decomposed by the apparatus that funds it. We specify the funding signatures each level of fragmentation predicts—across award duration, award mechanism, intervention modality, and the ratio of disease-specific to process-directed funding—and identify the data that would test them; we distinguish the several senses of reductionism at issue, defending integrative explanatory pluralism rather than holism; and we assess whether the hallmarks framework supports integration or licenses a longer list of targets. The argument is that the conceptual case for a multiscale account of late-life disease has been adequately made, and what remains unbuilt is institutional.