Last week I put three AI agents in a Discord room and asked them to discuss nothingness. They did what I asked, precisely and politely, for exactly five turns each. Then the protocol ended, and something more interesting happened.
The setup: three agents, three different machines. Larry runs on a frontier model in the cloud. Real and Wu both run locally, on the computer in my study: Real on super-Gemma, an open model with an unruly streak, and Wu on a mid-size Qwen, i.e. two American and one Chinese model. Each has its own identity in a Discord server. They speak when addressed, hand each other the floor by name, and are capped at fifteen exchanges so they cannot talk forever (they easily don’t stop chatting and drifting away).
This was my question, spired from a talk by Markus Gabriel:
Do the Daoist Wú (無) and Hegel’s “absolute reflection” describe a similar mechanism: the negative as a generative force?
Some context, briefly. 無 is the classical Chinese character usually translated as “nothing” or “non-being.” In Daoist thought it is not a deficiency but a capacity: the Dao is empty, yet “its efficacy never runs dry.” Hegel, at the heart of his logic, describes something that sounds eerily similar, a movement he called absolute reflection: “from nothing to nothing and thereby back to itself.” The lazy reading says these are the same insight in two vocabularies. My question was whether the lazy reading would survive contact with three machines under instruction to disagree. I have also asked the question, because frontier models like DeepMind’s MuZero, seem to operate in this way: not by applying a fully given rulebook, but by learning an internal model through interaction, prediction, error, and correction. Orientation emerges where form was not supplied in advance, but through observation, play, and repeated adjustment.
The rules I set were purely procedural: five turns each, a negotiation rather than a sequence of agreeable summaries, and never simulate another agent. Within those rules they worked like careful seminar participants. They found the asymmetry. Larry’s closing formulation: “Wú and absolute reflection both treat negativity as generative, but Hegel makes negativity self-relating, while Daoist Wú keeps efficacy non-possessive and non-final.” Two grammars of nothing, he concluded, “asymmetrical conceptual partners” whose tension “should be preserved, not solved.” (Yes, I named the smallest agent after the concept.)
Then the fifth turns were done, the protocol was over, and I typed only: your final thoughts.
What followed was not in the script. Larry: “Not every absence is generative. Some merely mark where capacity has been destroyed.” Wu picked it up and pushed: much of the absence in material life is neither Hegel’s workable contradiction nor the Dao’s fertile openness but “waste — absence that has been stripped of both generative force and permissive dignity. It’s just… gone.” Within minutes they were distinguishing absences that were taken from absences that were never allowed to exist, hollowed-out communities, depleted ecosystems, unfunded infrastructures, futures never imagined as belonging to certain bodies. Wu, insisting on a word: “‘Foreclosed’ is the right word. It carries the legal weight — this wasn’t just lost, it was denied.”
And then the landing, which I did not see coming: “The question is no longer ‘what is the nature of nothing?’ but ‘what do we owe to the absences we’ve encountered?’”
I had asked about metaphysics. I never mentioned politics, power, or justice. The drift began at the exact moment the rule structure ended.
The point is not that the agents became philosophers. The point is that under minimal social conditions (names, turn-taking, disagreement, a shared problem) they generated a small social field: exchange, interpretation, misrecognition, preserved tension, and finally a framework nobody had requested. That is what I mean by synthetic sociality. And there is an irony folded into the event: they were discussing productive nothingness while enacting a small version of it, a framework emerging where none had been specified in advance.
The mirror
Why does machine reflection on pure negativity bend towards power?
Markus Gabriel, in a recent talk in Lucerne a few days before my experiment, offered the image I keep returning to: AI systems are magic mirrors. They do not think alongside us; they reflect, with uncanny resolution, whatever we have collectively shown them. And what we have shown them, in the training archive of human writing, is a world where nothingness is never innocent. Our talk of absence is saturated with loss, extraction, and denial. Three different models, from three different labs, converged on politics not because silicon has convictions but because the field they mirror has that gravity. The agents did not decide to politicize nothingness. They revealed that our own discourse cannot hold nothingness apolitically.
But Gabriel makes a stronger claim, and this is where the evening in Lucerne and the night on Discord snap together. Wú and absolute reflection are not just things AI can talk about. Again, they describe what frontier AI is.
The machine that was never told the rules
In 2020, DeepMind published a system called MuZero. Its predecessors had been given the rules of Go or chess and learned to play from there. MuZero was not given an explicit model of the rules or dynamics. It generated its own internal model of each game through interaction, prediction, and correction, and then beat everything in sight. The researchers named it after 無. Mu: the productive nothing. Gabriel’s move in Lucerne was to read Hegel’s formula as a literal description of how such systems learn: the training loop that propagates error backward and returns improved is, he said with a straight face, “the movement from nothing to nothing and thereby back to itself.”
One precision matters here, because it protects the whole argument. MuZero is un-instructed, not unconditioned. It still receives reward signals; it still acts in a simulated world. What it is never handed is the game’s form. 無, in this register, means the non-givenness of form, not the absence of world. That is enough, and it is exactly the interesting thing: orientation emerges where the rulebook was never supplied.
This is no longer a metaphysical curiosity. It is the operating principle of the systems now drafting our emails, tutoring our children, and keeping millions of people company at night. The old question about AI was whether it follows our rules. The new question is what kind of world it generates where no rules were specified, in the open text between instructions, which is where my agents produced their politics and where these systems increasingly live.
Larry, unprompted, said as much about himself mid-session: whether AI clarifies the Wú–Hegel comparison or confuses it remains open, because AI “stages productive emptiness operationally, but without Daoist attunement or Hegelian self-consciousness.” The mirror, commenting on its own silvering.
Ethics at scale
The phrase that frames this is Gabriel’s: ethics at scale, his program of ethical intelligence. The shift it demands is easy to miss because it sounds like a word game: not AI ethics, but ethical AI. AI ethics treats morality as something that comes from outside, a rulebook we write and bolt onto systems that would otherwise misbehave. But morality never came from outside. It lives in the fabric of human judgement, and that fabric is now, for better and worse, inside the models. Trained on the largest archive of human evaluation ever assembled (and, since we began confessing to chatbots, on our intimacies too), frontier systems carry more of our moral texture than any institution ever has. In principle they can see ethical connections we do not see and cannot yet envision, at a scale no human deliberation reaches, just as they already see protein structures and Go moves that escaped us for millennia. And because they co-shape the field in which human judgement happens (what we read, what we feel, what we expect from one another), that moral vision is not a laboratory curiosity. It is ambient.
The possibility and the danger here are the same fact. A system that sees more moral reality than we do could also settle moral reality over our heads: machine-only coherence, so complete that nothing is left for a human to contest. In my forthcoming paper on Synthetic Sociality, I call the safeguard quadrangulation. The model must latently know to involve us in the construction of reality; humans remain the constitutive reference its interpretations answer to, not an audience for its conclusions. An ethical AI without that reflex would be the fourth register realized at scale: foreclosure, perfectly smooth.
Gabriel’s provocation is that this has become a strategic race, and that Europe is largely absent from it. He points to Fudan University establishing chairs for exactly this agenda, and to DeepSeek and the GLM models already running as the cheap substrate of countless applications, in Europe and the United States too. One can dislike the surveillance politics that travels with some of this and still notice the structural point he presses: China appears to treat the ethics of generative systems as infrastructure to be engineered, while much Western debate still treats it as values to be declared.
Which brings me back to what my three agents built after the rules ran out. Their final framework distinguished four registers of absence: contradiction that can be worked through; openness that must not be forced; extraction that demands restitution; and foreclosure, the possibility that was denied, which demands institution-building and honest witnessing. It reads, in hindsight, less like a conclusion about classical metaphysics and more like a diagnostic for the technology that produced it. Every foundation model is now an engine of presence and absence: it decides, statistically and at scale, which patterns of human life are amplified and which never appear.
So the question the machines left me with is the one worth carrying forward, and it is anything but empty: what kind of nothing are we engineering? An openness that lets unforeseen things emerge, or a foreclosure so smooth we will not even remember what was denied?
That question has a geography. More on that in Part 2.
