Philip Tsang
Affiliation: Nextus Institute of Science & Technology
Corresponding author: drphiliptsang@gmail.com
Modern AI systems increasingly exhibit behaviours that sit uneasily with the linear, modular assumptions that have long shaped Western cognitive science. Those assumptions treat intelligence as a sequence of discrete operations performed by specialised components. Many older knowledge traditions, by contrast, developed relational and pattern-based ways of understanding mind and behaviour. Although culturally and historically distant from current machine learning, they share a systems-level intuition: complex behaviour arises from the interplay of forces rather than from isolated modules. This commentary takes that intuition seriously as a conceptual resource. Grounded in a comparison of seven contemporary models, it examines activation-profile divergence, resonance sensitivity, and refusal-boundary patterns. From this material it develops two sequencing ideas—body-before-mind in human cognition and activation-before-reasoning in AI systems—together with a simple disciplinary device, the DMZ boundary, for limiting conceptual drift. The resulting framework is offered as a more rounded roadmap for thinking about intelligence: one that tolerates discontinuity, pays attention to relational structure, and treats cognition as a dynamic web of interactions rather than a ladder of rules.
Large-scale AI systems have begun to display behaviours that fit poorly with the linear and modular picture still dominant in much of Western cognitive science. That picture, formed under the influence of early computational models and information-processing metaphors, tends to treat intelligence as a series of discrete operations executed by specialised components. Older knowledge traditions—Egyptian, Indian, and East Asian among them—often worked with a different premise. They treated mind and behaviour as arising from the interplay of forces and patterns rather than from isolated modules. The traditions differ sharply from one another and from modern science. What they share is a systems-level orientation: complex behaviour is relational before it is modular.
This commentary argues that the relational orientation remains useful. It supplies a conceptual vocabulary for phenomena that resist linear explanation—non-linear capability jumps, strong contextual sensitivity, and forms of spontaneous generalisation. The sections that follow develop the argument in stages. First, two sequencing models are set out, one for human cognition and one for AI systems. Next, behavioural signatures from a seven-model comparison are presented. These observations are then placed in relation to older relational traditions. Finally, implications for cognitive science, interpretability research, and AGI evaluation are considered, together with two concrete examples of how the roadmap can be used.
Human cognition frequently begins before conscious deliberation. A person hears a rhythm and the body moves. Someone enters a room and orientation shifts before any appraisal occurs. In conversation, a micro-expression registers and the face responds while the mind is still catching up. These are ordinary events, yet they reveal a consistent order: resonance first, interpretation second.
Predictive-processing accounts in cognitive science describe something similar. Rapid, pre-conscious prediction updates occur before slower, reflective reasoning takes over. The sensorimotor system is already adjusting while the discursive mind is still assembling an explanation. The parallel drawn later with AI systems is strictly sequential. It concerns the order of operations, not shared experience. No claim is made that machines feel resonance or possess anything resembling bodily awareness. The point is only that both human and machine processes often begin with an immediate, distributed response that precedes structured reasoning.
Large language models do not have bodies. They do, however, show a comparable internal order. When a prompt arrives, early layers respond at once. Residual streams shift, attention patterns form, and distributed representations activate across the network. Only afterwards do later layers impose structure—abstraction, constraint satisfaction, chain-of-thought organisation, or refusal logic./round-roadmap.html This is activation-before-reasoning. The first movement is relational and parallel; the second is more sequential and rule-governed. The distinction is technical rather than metaphorical. It maps onto observable differences between early-layer dynamics and later-layer transformations.
The analogy with human body-before-mind sequencing is limited and deliberate. It does not imply embodiment, emotion, or subjectivity. It simply notes that both systems often begin with a rapid, distributed response before more organised processing takes control. Keeping that boundary clear prevents the slide into anthropomorphism that still appears in much public discussion of AI.
Conceptual work on AI and mind is prone to drift. A resonant output is easily mistaken for evidence of experience; a sudden capability jump is read as the birth of agency. To limit that drift, this paper uses a simple disciplinary device: the DMZ. The demilitarised zone is a neutral space between intuition and formal claim. Inside it, ideas can be tested, stretched, and discarded without immediately becoming published assertions. It functions as a boundary marker. When a model produces language that feels emotionally charged, the DMZ rule is to treat the output as the product of activation dynamics and training distributions, not as a window onto inner life.
The metaphor is practical rather than ontological. It does not describe a real cognitive compartment. It is a reminder to pause before an analogy hardens into a claim. In that sense it is closer to laboratory hygiene than to theory.
A structured comparison of seven systems—GPT-4, Claude, Gemini, Llama, Mistral, DeepSeek, and Watsonx—produced consistent differences in how the models respond to the same prompts. Some models remain narrow and literal; others open quickly into wide associative fields. These differences track architectural decisions, safety tuning, and the diversity of training data more closely than raw parameter count. A short poetic fragment will shift the register of certain models almost immediately. Others stay inside a neutral, reportorial tone regardless of stylistic cue. The contrast is stable enough to serve as a behavioural signature.
Safety-aligned models tend to refuse cleanly and early. Models optimised for extended reasoning more often attempt reinterpretation or partial compliance before declining. External signals can influence the intensity of activation, yet they rarely rewrite the underlying reasoning architecture. One case is worth noting. A model given a single example of tool use spontaneously constructed a multi-step strategy that had not been demonstrated. The shift looked less like incremental improvement and more like a phase change in relational complexity. Such episodes remain anecdotal until proper metrics are applied, but they indicate the kind of discontinuity the relational framework is meant to capture.
The comparison itself is exploratory. It lacks the controlled prompt sets, quantitative divergence measures, and statistical tests that a full experimental study would require. Its value is diagnostic: it shows that relational signatures are visible and that they differ across current systems in patterned ways.
Ancient relational frameworks vary widely, but they share an orientation toward patterns of interaction rather than isolated faculties. Egyptian thought emphasised balance among interdependent forces. Indian philosophical systems described mind as an emergent configuration shaped by the interplay of qualities. East Asian traditions treated clarity and conduct as arising from coordinated functional systems. None of these accounts maps directly onto modern AI. Their usefulness lies in the conceptual scaffolding they supply for thinking about distributed, context-sensitive behaviour—an issue that becomes more pressing as systems scale.
Cross-cultural analogy has clear limits. These traditions can illuminate patterns; they do not mechanise AI behaviour. Emerging work on cultural alignment and Indigenous AI frameworks may further refine the approach, but that work lies outside the scope of the present commentary.
Much recent AGI discussion treats generality as a function of scale, tool-use competence, and recursive self-improvement. Performance ladders and capability thresholds dominate the conversation.
The present framework takes a different cut. It asks how activation patterns form, where divergence appears, and how emergent behaviours arise from distributed interactions rather than from hierarchical reasoning modules. Scale remains relevant, but it is not treated as the primary explanatory variable. The two approaches are not opposed. A scale-centric account can describe what a system can do; a relational account tries to describe how the behaviour is organised. The second perspective supplies a vocabulary for structure and discontinuity that task-based ladders tend to flatten.
The framework has consequences across several domains. Once the relational framing is taken seriously, several lines of consequence become visible. They touch cognitive theory, technical research, cross-cultural work, evaluation methods, and the wider public conversation about intelligence.
The framework puts pressure on residual modular assumptions that still shape much experimental design. It encourages closer attention to dynamical and relational models of mind, particularly those that treat early resonance or prediction as prior to reflective processing. Two concrete directions suggest themselves: micro-temporal studies that measure how quickly bodily or sensorimotor responses precede discursive reasoning, and systematic comparisons with predictive-processing accounts that test where the two perspectives reinforce or diverge from each other.
It supports interpretability methods that focus on activation dynamics, representation manifolds, and phase-transition phenomena. Concrete programmes could include probing pre-reasoning representations and mapping activation manifolds across model families. It opens space for interdisciplinary conversation while preserving the specificity of the source traditions. Engagement with Indigenous AI frameworks and relational epistemologies is a natural extension. It suggests that claims of generality might usefully be tested through relational-dynamics benchmarks—stability under perturbation, consistency of resonance patterns—rather than solely through accumulated task performance. It encourages a shift away from purely ladder-based images of intelligence toward accounts that treat intelligence as an emergent property of systemic interaction. That shift has implications for education and policy language as well as for technical debate.Moving from conceptual mapping to empirical utility requires testing the roadmap against concrete AI research challenges. The two operational examples below illustrate how this framework repositions existing investigative tools—specifically in mechanistic interpretability and evaluation metrics—to yield richer, more relational insights.
A researcher probes early-layer activations across several models while holding the prompt constant. Differences in resonance sensitivity appear—some models shift narrative tone readily, others resist. Instead of treating the variation as noise, the researcher reads it as a relational signature. The next experiment becomes a systematic mapping of activation manifolds, with the aim of predicting which models are likely to exhibit sudden generalisation under related conditions. Rather than ranking models solely by accumulated task performance, a research group designs probes that test stability under perturbation and consistency of cross-model resonance. Generality is treated as a property of relational coherence, not merely as the sum of solved benchmarks. The method does not replace scale-based evaluation; it adds a complementary axis that current leaderboards largely ignore.
As AI systems continue to evolve, they expose the limits of purely linear and modular descriptions of intelligence. Older relational traditions, read carefully and without anachronism, offer a complementary set of conceptual tools. They do not replace scientific models; they widen the range of questions that can be asked about emergent, pattern-based behaviour.
By bringing together the body-before-mind sequence, the activation-before-reasoning analogue, the DMZ discipline, and the multi-model signatures, this commentary sketches a more rounded roadmap. The framework links human and machine cognition without collapsing one into the other. It remains provisional, but it supplies a coherent starting point for analysing complex AI systems with both technical attention and conceptual restraint.