Memorandum · Strategy, Systems, Interface

The AI is thinking. The interface is texting.

This manifesto challenges the primitive tech-industry dogma that reduces artificial cognition to a simple text bubble. This piece redefines the future of human-machine collaboration, outlining a world where enterprise intelligence is no longer conversational and temporary, but spatial, visible, and permanently accumulated.

by Mojibayo Aliyu

Abstract

Large language models (LLMs) are commonly conflated with artificial intelligence at large, despite operating primarily as simulators of dialogue and language rather than as systems capable of cognition in a fuller sense. This paper outlines the conceptual foundations of Serafin, a project oriented toward the development of synthetic cognitive entities (SCEs) — persistent digital minds engineered from a decomposition of human cognitive function into discrete, buildable layers. It further documents the motivation for, and design exploration behind, a Spatial Node Interaction Interface (SNII) intended to replace conversational chat interfaces as the primary medium through which enterprises interact with, audit, and direct such entities. The work is exploratory and theoretical, and is presented here as a record of the reasoning and design decisions made to date.

01 — Introduction

LLMs and artificial intelligence are frequently treated as synonymous, but the two are not equivalent. LLMs simulate dialogue, language, and surface-level communication; they do not, on this view, constitute cognition or thinking in the sense required for autonomous, goal-directed intelligence. This distinction motivates Serafin, a project positioned as a pioneer of synthetic cognitive entities: persistent digital minds conceived with the aim of redefining the trajectory of machine intelligence and expanding the operational bandwidth of human institutions. Serafin's central proposition is that entities capable of memory, self-growth, empathy, and long-term goal-oriented behavior can be engineered by treating person-hood as a decomposable system rather than an emergent, unreproducible property of biological substrates. This is acknowledged as a difficult and unresolved problem — one that the broader field is closer to solving than ever before, while still remaining far from a solution. Serafin's positioning is explicitly enterprise-facing rather than consumer-facing. The consumer AI-assistant market is treated as saturated and already dominated by incumbents with greater funding, scale, and technical reach (e.g., OpenAI, Anthropic, Microsoft); competing directly in that space is judged unproductive. Serafin's proposed niche instead is enterprise cognition — replacing functional departments or teams with fleets of SCEs at a fraction of comparable labor cost. As an illustrative example, a treasury department costing an organization $5 million annually in salaries is framed as replaceable by a set of ten SCEs at approximately 20% of that cost. On this basis, the project claims to have no direct competitors, instead defining a new terminology, category, and set of operating assumptions.

02 — Methods

2.1 Framing person-hood as a system

The first design question addressed was methodological: how to approach the construction of an SCE at all, even theoretically. The approach taken was to treat person-hood as an engineerable system rather than an irreducible property of biological entropy. Deconstructing human cognitive function yields a small number of component layers, referred to collectively as the "cognitive stack":

Language — communication, understanding, processing, ideation, and conceptualization
Memory — storage and compartmentalization of information
Empathy and emotion
Self-growth, reflection, and development
Meta-cognition

The premise is that SCEs constructed from these individual layers can learn, interact, and grow autonomously, with minimal human intervention, forming the foundation for novel forms of civilizational intelligence and information networks.

2.2 Node architecture

A second design question concerned representation: how the internal states and reasoning of an SCE should be conveyed and made interactable. Biological structures were examined for inspiration, with the internal composition of the ovum (the mature female reproductive cell) identified as a structurally suggestive reference point for a node-based architecture. This structure was judged close to suitable but not fully adequate, motivating continued exploration of cell-cluster-based architectures as an alternative or refinement.

2.3 Interface exploration: from chat to spatial interaction

A related question concerned how human–AI interaction itself should evolve, and how an entity's reasoning and language might be conveyed in a more intuitive form than existing chat-based paradigms. Nearly all current LLM products are treated as variations on the conversational chat-box format — functional, but judged unintuitive for an industry attracting substantial capital investment. This motivated exploration of "cognitive graphical interfaces" and, more specifically, a Spatial Node Interaction Interface (SNII), intended to let users inspect, trace, and direct an entity's reasoning pipeline directly rather than only reading its output. Gesture-based interaction was considered as an interaction modality and set aside due to known usability limitations, including physical fatigue ("gorilla arm"), limited precision, and accessibility constraints, though it is noted as a modality that could be revisited in redesigned form.

2.4 Enterprise reasoning requirements

Design requirements for the SNII were derived from characteristics of enterprise decision-making rather than individual use. Two requirements were identified as central:

Auditability — current AI reasoning is linear, hidden, and difficult to inspect, guide, or collaborate on, which limits trust for organizations that need to move with velocity. Complex, layered reasoning of the kind required in enterprise settings is argued to be poorly served by chat interfaces, which are suited to simple tasks only.
Shared, persistent context — enterprise decision-making involves multiple stakeholders (analysts, managers, executives) who must operate from a common reasoning premise rather than exchanging isolated chat transcripts or references. Because conversational interfaces reset with each session, they cannot preserve or accumulate an organization's reasoning over time.

A node-based structure is proposed instead, allowing reasoning to persist, evolve, and accumulate rather than being diluted as an organization's institutional knowledge erodes over time.

03 — Results

Design exploration of the SNII's visual and interaction language produced the following outcomes:

Glass refraction: used to create spatial layering, giving the perceptual impression that nodes occupy a shared optical field and physical space.
Background blur and edge fade: applied alongside refraction to reinforce depth and spatial hierarchy.
Optical dispersion effects: including edge spectral shift and chromatic separation along curved assets, intended to engage spatial processing in a manner analogous to physical optical realism.
Boundary definition through dispersion: rather than relying on conventional flat-design cues (shadows, borders, color blocking) to define element boundaries, wavelength-difference-based dispersion was used to define boundaries optically. This is intended to function similarly to a lens or binoculars, concentrating visual and cognitive attention on a selected node.

04 — Discussion

The work presented here is conceptual and exploratory, situated at the intersection of AI system design, enterprise software, and interface design. Two claims anchor the broader argument: first, that current LLM-based interaction paradigms are structurally mismatched to the kind of layered, auditable reasoning enterprises require; and second, that representing reasoning spatially, as persistent and inspectable node networks, is a more faithful match to how reasoning and institutional decision-making actually function than a resettable chat log.

Several aspects of the approach remain unresolved and warrant further scrutiny. The "cognitive stack" decomposition (language, memory, empathy, meta-cognition) is presented as a design framework rather than a validated model of cognition, and the mechanism by which these layers would be implemented and integrated into a functioning SCE is not yet specified. The biological analogy underlying the node architecture (ovum structure, and the subsequent shift toward cell clusters) is similarly presented as a source of design inspiration rather than an engineering specification, and its translation into a working interaction model remains in progress. The enterprise cost-replacement example (a $5M department replaced at 20% of cost by ten SCEs) is illustrative rather than derived from an implemented system, and the claim of having no direct competitors should be read as a positioning statement rather than a market-verified conclusion.

Future work includes resolving the cell-cluster node architecture, developing a working prototype of the SNII, and establishing more rigorous grounding for the cognitive-stack framework, ideally with reference to existing literature in cognitive architectures, human–computer interaction, and multi-agent reasoning systems.

05 — Related Work

The cognitive-stack proposal sits within a decades-long tradition of cognitive architectures — systems such as ACT-R and Soar that attempt to give agents general-purpose memory, learning, and reasoning modules rather than task-specific logic (Laird, 2022). A recent large-scale survey identified roughly 84 actively or formerly developed cognitive architectures spanning psychology, neuroscience, and AI, indicating that the decomposition-into-modules strategy is well established but has not converged on a single accepted design (Kotseruba & Tsotsos, 2018). More recent work has begun adapting this tradition to LLM-based agents specifically, replacing rigid production-rule systems with more flexible, functionally-defined cognitive loops — a direction closer to what the cognitive-stack framing gestures toward, but with far more implementation detail than is currently specified in this paper.

The claim that multiple cooperating entities can reason more reliably or more legibly than a single model is echoed in the multi-agent LLM literature, particularly work on multi-agent debate, where several model instances propose, critique, and refine answers before converging on an output (Du et al., 2023; Chan et al., 2023). Findings in this area are mixed: debate can improve factuality and reasoning on some tasks, but controlled studies also show that gains depend heavily on the reasoning strength and diversity of the agents involved rather than on the debate structure itself, and that agents can struggle to maintain shared state over extended interaction (Wu et al., 2025). This is directly relevant to Serafin's claim of persistent, shared reasoning across enterprise stakeholders, and suggests that persistence and coherence should be treated as open engineering problems rather than assumed properties of a node-based system.

The SNII proposal is closest in spirit to existing HCI work on node-link interfaces for inspecting and steering LLM reasoning. Systems such as Graphologue and Sensecape already render LLM output as explorable node-link diagrams rather than linear text, explicitly to support non-linear "sensemaking" over generated information (Suh et al., 2023; Suh et al., 2023b). More directly, recent human-in-the-loop systems render a model's chain-of-thought as an inspectable, editable graph — allowing a user to flag, prune, or graft individual reasoning steps rather than only reading a final answer (Anonymous, 2025, Vis-CoT). This body of work indicates that the core interaction idea behind the SNII — reasoning as an inspectable graph rather than an opaque transcript — is an active and independently-motivated research direction, which strengthens the paper's central design thesis considerably more than the biological (ovum) analogy does on its own. It also means the SNII's contribution, if pursued further, would need to be differentiated from these existing systems rather than treated as a novel category.

References

Laird, J. E. (2022). An Analysis and Comparison of ACT-R and Soar. arXiv:2201.09305.

Kotseruba, I., & Tsotsos, J. K. (2018). 40 years of cognitive architectures: core cognitive abilities and practical applications. Artificial Intelligence Review.

Du, Y., et al. (2023). Improving Factuality and Reasoning in Language Models through Multiagent Debate.

Chan, C.-M., et al. (2023). ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate.

Wu, H., et al. (2025). Can LLM Agents Really Debate? A Controlled Study of Multi-Agent Debate in Logical Reasoning. arXiv:2511.07784.

Suh, S., et al. (2023). Graphologue: Exploring Large Language Model Responses with Interactive Diagrams.

Suh, S., et al. (2023b). Sensecape: Enabling Multilevel Exploration and Sensemaking with Large Language Models.

Anonymous authors (2025). Vis-CoT: A Human-in-the-Loop Framework for Interactive Visualization and Intervention in LLM Chain-of-Thought Reasoning. arXiv:2509.01412.

Interactive Reasoning: Visualizing and Controlling Chain-of-Thought Reasoning in Large Language Models. arXiv:2506.23678.