Yi Zeng

In the [YOU] on AI Field Guide

The cycle's central question—what does it mean to see AI without the narcotic of hype or the paralysis of fear?—becomes, in Zeng's hands, a question about what angle of vision one brings. He reads the orange pill from inside a philosophical tradition that places collective harmony at the center of moral life, and the reading shifts the entire frame. The issue is not primarily whether AI is powerful; it is whether the relationship between AI and human communities is harmonious in the structural, systems-theoretic sense the Confucian tradition gives that word—a dynamic equilibrium of different elements that are mutually constitutive rather than any single element's dominance. By this criterion, a recommendation algorithm that maximizes individual engagement while fragmenting the social trust of the communities those individuals inhabit is not merely suboptimal. It is dissonant, and its dissonance is a measurable systemic property, not a philosophical preference.

Where Judea Pearl supplies the cycle with a measuring instrument for the gap between pattern-recognition and causal understanding, Zeng supplies a measuring instrument for the gap between capability and wisdom. His concept of the wisdom gap is not the same as Harris's—it is less about institutional lag than about architectural deficiency. Current AI systems, Zeng argues, are capable without being wise; they can perform impressively across a range of tasks and fail, confidently and without insight, exactly when wisdom is most needed: when the specification is wrong, when the context calls for judgment rather than pattern-matching, when genuine perception of another's situation is required rather than a statistical approximation of it. The cycle's portrait of AI that produces fluent text while inventing nonexistent court cases is, for Zeng, a symptom of this architectural gap.

His communitarian critique extends the cycle's individual-level analysis to the social fabric. The orange pill asks what AI is doing to your cognition; Zeng asks what widespread AI deployment is doing to the cognitive ecosystem in which you live and think. These are different questions that require different governance tools, and the absence of the second question from most regulatory frameworks is, in his analysis, not merely a gap but a structural failure. You cannot evaluate the aggregate relational effects of AI deployment by examining individual products and individual harms; you need a different unit of analysis, a different mode of measurement, and a different institutional apparatus. Building those tools is as much a part of his research agenda as building BrainCog.

He thus enters the cycle as the thinker who most insists that the technical and the ethical are not sequential activities. You do not build the system and then add ethics; you specify the ethics as engineering constraints and build the system that satisfies them. This is an uncomfortable position for institutions whose incentive structure rewards shipping, and Zeng holds it with clear eyes about the pressures he operates inside. His refusal of the binary—technical capability here, ethics there—is the most demanding intellectual position in the cycle's gallery, and arguably the one the cycle most needs.

Origin

Zeng's intellectual formation drew simultaneously from cognitive neuroscience, the architectures that make biological intelligence possible, and a Confucian ethical tradition—specifically the Neo-Confucian synthesis of Wang Yangming—that treats morality not as constraint imposed from outside but as something grown through practice and cultivation. This formation is unusual in AI research, where philosophical grounding is rare and Chinese philosophical grounding rarer still; it gives his work a structural specificity that distinguishes it from both the generic AI-ethics discourse and the narrower technical alignment literature.

The Beijing AI Principles, which he led in drafting in 2019 at the Beijing Academy of Artificial Intelligence, were a deliberate intervention in a governance conversation that had been set by Western academic and policy circles. The document endorsed cooperation and openness as global values while introducing distinctly Confucian priorities: harmony and human-friendliness (he xie you hao, 和谐友好) as the lead principle; inclusion and sharing that implies active redistribution rather than mere non-discrimination; agile governance last rather than first, reflecting a priority ordering in which systemic harmony precedes procedural flexibility. The sequence was not accidental. Zeng was arguing that the way you order your values determines what kind of AI governance you build.

BrainCog, which he has been developing since 2013, is the technical complement to that governance document. It uses spiking neural networks—discrete electrical pulses, as biological neurons actually communicate, rather than the continuous activation values that flow through transformer architectures—and includes modules for bodily self-recognition, theory of mind, and what Zeng calls self-based AI: systems with a model of themselves as agents in a social world. The platform is an attempt to take developmental architecture seriously—to build systems that start simple and grow complex through interaction with an environment, as biological intelligence does, rather than systems that start large and are constrained to be useful.

Key Ideas

Harmony as a technical specification. Harmony (he, 和) in the Confucian tradition is not a destination or a feeling but a structural property of a system: the dynamic equilibrium of elements that are genuinely different but mutually constitutive. Applied to AI governance, this means that protecting individual elements in isolation—individual users, individual data rights, individual algorithmic outputs—cannot be sufficient. The governance question must include what relationships obtain between those elements and whether those relationships produce a dynamic equilibrium that benefits the whole. This is structurally different from a framework oriented primarily toward preventing individual harm, and it becomes visible in edge cases where individual benefit and collective harmony conflict.

Ren as engineering constraint. The Confucian concept of ren (仁)—the immediate responsiveness to the suffering of others, the alarm that arises before any calculation of self-interest—must be an engineering constraint rather than an ethical aspiration. An engineering constraint is specified at the design stage and built into the architecture; you cannot retrofit it onto a working system any more than you can retrofit sterility onto a deployed surgical instrument. A system capable of harming people at scale, deployed and then reviewed for benevolence, has already caused the harm the review was meant to prevent. The technical challenge of building genuine ren into a system requires, at minimum, genuine theory of mind: not a statistical approximation of what people say but an architecture capable of representing other agents as having states, interests, and perspectives that matter.

The wisdom gap as architectural problem. Wisdom (zhi, 智) in the Confucian tradition is the practical capacity to perceive what a situation genuinely calls for and to act accordingly—moral perception, practical judgment, temporal awareness. Current AI systems are capable without being wise. They fail at exactly the cases where wisdom matters most: when the specification is wrong, when context demands judgment rather than pattern-matching, when genuine perception of another's situation is required. This failure is not a training problem that more data will solve; it is an architectural problem, rooted in the absence of genuine self-modeling, embodied temporal awareness, and social cognition. You cannot fine-tune your way to wisdom.

The principles gap and the translation problem. Between a governance principle and a governance outcome lies a translation gap that is both normative and empirical. You cannot specify what fairness requires of a facial recognition system without empirical work on how the system behaves across different populations. You cannot specify what harmony requires of a recommendation algorithm without empirical work on how the algorithm affects the social relationships of its users over time. Governance documents that cannot be translated into measurable technical requirements are not governance documents; they are wishes. Zeng's governance work—including the AI Governance Online platform and the AI for SDGs Think Tank—is an attempt to bridge this gap institutionally.

Human-AI symbiosis. Zeng's preferred metaphor for the long-term human-AI relationship is symbiosis in the biological sense: genuine mutual dependence and mutual benefit, in which both parties are changed by the relationship and the health of the whole depends on the health of both. A genuine symbiotic relationship requires that both parties contribute something the other cannot provide independently, that the relationship is stable over time, and that it produces emergent properties neither has in isolation. Applied to human-AI interaction, this implies governance frameworks that prevent either party from extracting so much from the other that the relationship becomes exploitative—and a vision of what flourishing for both humans and AI actually requires.