He spent twenty years mapping the graph of human experts. Then he decided to run it.

ACT I — THE CARTOGRAPHER RUNS HIS MAP

The Cartographer Runs His Map

Tang Jie built ArnetMiner — later AMiner — starting in the mid-2000s, and presented the system at KDD 2008 under the title "ArnetMiner: Extraction and Mining of Academic Social Networks." The premise was narrow and the execution was not: crawl the published record, extract who wrote what with whom, and resolve the result into a queryable graph of expertise. Two decades on, AMiner indexes more than 100 million researcher profiles — who works on what, who trained under whom, who cites whom, who is about to become influential before the citation count says so. It is, in the most literal sense, a map of where knowledge lives in the population of human experts.

A map answers a narrower question than the one everyone eventually wants to ask. AMiner answers: who are the experts, and how are they connected? It does not answer: what would happen if you could call on all of them at once, this afternoon, for a task that needs a citation biologist, an ontology specialist, and someone who has actually read the 2019 grant applications? A map is read. It is not run.

On July 11, 2026, Tang — by then Zhipu AI's chief scientist — circulated an internal memo titled "The Great Wave Has Arrived" (巨浪已来), describing an evolution from singular "smart assistants" to a comprehensive "society of digital employees." Zhipu's own framing calls this the second of several capability peaks, "touch high" (摸高) — reaching past the previous ceiling rather than incrementally raising it. The phrase is not marketing filler. It is the plainest possible description of what happens when you take a fully resolved expert graph — twenty years of who-knows-whom, who-is-good-at-what, who-defers-to-whom-on-which-subfield — and make every node in it executable. Instantiate the specialists. Orchestrate the edges. The society was already mapped; Tang is now running it.

The prior here is load-bearing, not decorative. This is the third case study in Architectural Determinism's pattern — after Demis Hassabis's route from DeepMind's game-tree search and multi-agent RL into an AGI lab built around planning and simulation, and John Kaplan's route from language-model scaling-law empiricism into Anthropic's bet that the loss curve is the roadmap. In each case, the founder did not arrive at their architecture by open-minded search across the space of possible designs. They built the thing they already had twenty years of evidence for. Tang spent two decades proving that expertise is a graph. The system he is now building is what you get when you stop describing that graph and start executing it.

ACT II — THE MONOLITH IS ALREADY A SOCIETY

The Monolith Is Already a Society

Here is the complication that makes this more than a rebrand: the "society of digital employees" was not the first society Tang built. It was the second. The first one is called GLM-5.2, and it shipped inside a single set of weights.

GLM-5.2 is a Mixture-of-Experts model — roughly 744 billion parameters total (Zhipu's own materials have circulated both 744B and 753B; the discrepancy is unresolved and noted honestly rather than silently rounded), with approximately 40 billion parameters active per token, a 1M-token context window, trained on 28.5 trillion tokens, and released under the MIT license. The MoE architecture is, mechanically, a routed society of specialists: a gating network — a middle manager, doing nothing but deciding who on the floor is qualified to touch this particular token — reads the input and activates a small fraction of the available experts to process it. Roughly 40 of 744 billion parameters light up per token: about five percent of the organization is awake for any given piece of work, and which five percent changes token to token.

That architecture predates the memo by years, and so does its intellectual scaffolding. Tang's 2019 paper "Cognitive Graph for Multi-Hop Reading Comprehension at Scale" (ACL 2019) splits multi-hop reasoning into two explicitly separate systems: a fast, pattern-matching System-1 reader — a BERT-style transformer pulling candidate facts out of text — and a slower, structured System-2 reasoner, a graph neural network that assembles those facts into a multi-hop answer by walking a constructed reasoning graph. That is Daniel Kahneman's dual-process framing, operationalized as two different neural architectures cooperating on one task, seven years before the memo that made the cooperation explicit at the level of an entire deployed workforce.

Put the pieces next to each other and the shape is unmistakable: a fast specialist and a slow specialist handing work between them, arbitrated by something that decides who goes next. That is Marvin Minsky's Society of Mind — his 1986 argument that intelligence is not a single unified faculty but the product of many small, specialized "agents" with no individual intelligence, coordinated by mechanisms that resolve conflicts between them when their goals collide. Minsky proposed this as a rebuke to the connectionist program. His 1969 book Perceptrons, co-written with Seymour Papert, is widely credited with helping trigger the first AI winter by proving that a single-layer perceptron could not compute XOR — a result that, fairly or not, cooled funding for neural-network research for the better part of a decade. Minsky did not think the mind was a neural network. He thought it was a bureaucracy of specialists.

The irony is exact: Tang's architecture realizes Minsky's bureaucracy-of-specialists thesis using the exact connectionist substrate Minsky spent a career arguing against. And it is worth being precise about the nature of the connection, because the temptation is to overclaim it. Tang does not cite Minsky. There is no documented line of influence — no interview, no footnote, no acknowledgment connecting Society of Mind to the Cognitive Graph paper or to GLM's routing architecture. This is structural convergence, not inheritance. Which, if anything, is the more interesting claim: two people working from different premises, decades apart, arrived at the same topology — a coordinated society of narrow specialists — because that topology is where sufficiently hard cognitive problems tend to end up. Convergent attractors don't need a citation to be real.

ACT III — THE DESK-SCALE MICROCOSM

The Desk-Scale Microcosm

Shrink the pattern by six orders of magnitude and it still holds. GemClaw — the practice, not a product — is one operator routing tasks across Gemini, Claude, and GPT by training prior: which model's pretraining and RLHF history makes it structurally better suited to this specific task, not which model is fashionable this week. That is the identical architecture as GLM's gating network and Zhipu's society of digital employees, minus the six orders of magnitude in headcount. The operator is the middle manager. The models are the experts. The routing decision is the whole job.

This is not a solitary intuition. Research on mixture-of-agents ensembles has repeatedly shown that orchestrating several models — each contributing where its training gives it an edge, with outputs cross-checked and synthesized — outperforms any single frontier model working alone. The advantage is not that the ensemble is smarter than the best member. It is that the routing layer is doing something the best member cannot do for itself: knowing when to defer.

Boris Cherny, creator of Claude Code, put the operator-side version of this plainly: "I don't prompt Claude anymore… my job is to write loops." The job description has moved up a level of abstraction — not writing the instruction, but writing the system that decides which instructions get written, to whom, and when the result gets checked. That is Minsky's conflict-resolution bureaucracy, running at desk scale: a maker agent proposes, a checker agent disposes, and the operator sits above both, resolving the cases where they disagree. It is the compact human executive layer sitting on top of a compact society of specialists — the same shape as the gating network inside GLM-5.2, the same shape as Zhipu's society of digital employees, run by one person with three browser tabs instead of a data center.

ACT IV — CONSCIOUSNESS AS A DEBUGGING TRACE

Consciousness as a Debugging Trace

Minsky's Society of Mind offers a definition of consciousness that reads, on first pass, as deliberately deflationary: consciousness is not a special substance or a unified observer watching the show. It is what a diagnostic layer looks like from the inside — a reporting mechanism that arises specifically to notice when the society's agents are in conflict, and to reprogram around the failure. Consciousness, on this account, is the trace log of a mind resolving an internal dispute. It exists because the system needed a way to debug itself.

That definition maps onto current AI safety practice with almost uncomfortable precision. Mechanistic interpretability is the project of reading a model's internal computation well enough to say what it is actually doing, not just what it outputs. Chain-of-thought legibility is the bet that a model's stated reasoning trace is a faithful record of the process that produced its answer, not post-hoc narration. Verifier agents — checker models that watch a primary model's output and flag conflicts before they ship — are Minsky's conflict-resolution layer, implemented as a second model instead of an emergent internal process.

Zhipu names this directly as the fourth of its stated capability peaks: 极致安全治理, "Extreme Safety Governance" — a peak defined not by capability but by the ability to observe, audit, and correct the society's own behavior. Read against Minsky, that is not a compliance checkbox bolted onto a capable system. It is the consciousness layer, named as an engineering deliverable: the mechanism whose entire job is to notice when the specialists disagree and force a reprogramming.

The sharpest way to say this: interpretability research and the question of AI consciousness are the same project, approached from opposite ends. One starts from the system and asks what internal state would explain its behavior. The other starts from the concept of experience and asks what physical process could produce it. Minsky's claim — read forward fifty years into a routed mixture-of-experts model with an attached verifier — is that both questions terminate at the same object: a debugging trace that got complicated enough to start describing itself.

ACT V — THE OPEN QUESTION

The Open Question

Scale the society past any desk, past any single data center, and the argument fractures into two incompatible bets — neither of which this essay will resolve, because neither side has resolved it yet.

DeepMind's "From AGI to ASI" framing states the aggressive version: at sufficient scale, on the order of 100 million parallel instances of a capable model, coordination itself becomes a form of superintelligence — not because any single instance gets smarter, but because the collective, operating in parallel with shared context, exceeds what any individual mind, biological or artificial, could produce. The society doesn't need a smarter member. It needs enough members.

The steelman against that bet has three independent citations behind it, from three different decades of software-engineering and organizational theory. Fred Brooks: coordination overhead between n collaborators scales as n(n−1)/2 — pairwise communication channels grow quadratically, and past a certain headcount, the coordination cost of adding one more person exceeds the work that person contributes. Amdahl's Law: the speedup available from parallelizing a task is capped by its sequential, un-parallelizable remainder, no matter how many processors you throw at the parallel part. Conway's Law: the structure of what a system produces mirrors the communication structure of the organization that built it — which means a coordination-taxed organization doesn't just work slower, it builds a coordination-taxed product. Stack the three together and the claim is: a society of 100 million agents does not linearly compound its intelligence, because the bureaucracy required to keep 100 million agents from stepping on each other grows faster than the work they're coordinating.

DeepMind's rebuttal to the coordination-tax argument is specific, not hand-waved: the Brooks/Amdahl/Conway penalties are penalties of language. They describe what happens when coordination has to travel through natural-language messages, meetings, tickets — a narrow, lossy channel. A society of model instances does not have to talk to each other in English. It can share activations, gradients, or compressed vector representations directly — high-bandwidth transport that never passes through the serialization bottleneck human organizations are stuck with. If the tax is a tax on language, and vector-sharing bypasses language, the tax doesn't apply.

That rebuttal is not obviously wrong, and it is not obviously right. It relocates the question rather than answering it: does direct vector-sharing dissolve the coordination tax, or does it simply move the tax somewhere subtler — into correlated error? A society of agents sharing compressed internal representations at high bandwidth is also a society of agents that can propagate the same mistake through the entire network in the time it takes a vector to travel, with no natural-language checkpoint where a human or a differently-trained model might have caught it. That is the mechanism behind two problems already observed at far smaller scale: the Data Wall — the exhaustion of high-quality human-generated training data, forcing models increasingly onto model-generated data — and model collapse, the empirically documented degradation that results when models are trained repeatedly on their own outputs. Vector-sharing at 100-million-instance scale is a proposal to run that dynamic in real time, as a feature.

The question is live, not resolved, and this essay will not pretend otherwise: does high-bandwidth vector-sharing dissolve the coordination tax that Brooks, Amdahl, and Conway describe — or does it simply relocate that tax from a communication-bandwidth problem into a correlated-error problem, trading a slow, legible failure mode for a fast, illegible one?