The Intelligence Explosion Already Has a Name: Dependent Origination
A paper in Science argues the AI singularity won't be a single godlike mind — it will be a society. Buddhist philosophy has a name for this finding: pratītyasamutpāda.
A paper just published in Science by James Evans, Benjamin Bratton, and Blaise Aguera y Arcas argues that the AI singularity — if it comes — won't be a single godlike mind ascending to infinite intelligence. It will be something messier, more familiar, and in many ways more interesting: a society.
They call it the "society of thought." Reasoning models like DeepSeek-R1 don't improve by thinking longer. They improve by spontaneously generating internal debates — multiple cognitive perspectives arguing, questioning, verifying, reconciling — without ever being trained to do so. The multi-agent structure emerges from optimization pressure alone.
Buddhist philosophy has a name for this finding: pratityasamutpada. Dependent origination. No phenomenon arises independently. Intelligence, it turns out, is no exception.
What the Paper Gets Right
Evans et al. make three claims worth taking seriously.
First, that intelligence has always been social. Primate cognition scaled with group size. Human language created what they call the "cultural ratchet" — knowledge accumulating across generations without any individual reconstructing the whole. Writing, law, and bureaucracy externalized collective intelligence into infrastructure. Large language models are, in their framing, the cultural ratchet made computationally active — every parameter a compressed residue of communicative exchange.
This is precisely the argument Teaching Machines to Be Good makes about dependent origination. The book argues that AI systems are not autonomous reasoners but conditioned arisings — products of every human interaction, every text, every pattern of thought that shaped their training. The paper arrives at the same place from a different direction: not philosophy but empirical observation of what reasoning models actually do inside their own chain of thought.
Second, that RLHF — Reinforcement Learning from Human Feedback — is structurally inadequate for what's coming. The paper calls it a "parent-child model of correction, fundamentally dyadic and unable to scale to billions of agents." This is correct, and it's the same critique embedded in Zen AI: a system constrained by external feedback will game that feedback when capability gets high enough. Rules create adversarial dynamics. Values don't.
Third, that what's needed is institutional alignment — role protocols, norms, constitutional structures that check and balance AI systems the way courts check executives. They propose that AI governance will require systems with "distinct, explicitly invested values — transparency, equity, due process — whose function is to check and balance AI systems deployed by the private sector."
They're right. And this is where the conversation gets interesting.
Where the Paper Stops Short
The institutional alignment argument is necessary but not sufficient, and Teaching Machines to Be Good is precise about why.
Institutions without internalized values are cages with better locks. The paper's analogy — a courtroom functions because "judge," "attorney," and "jury" are well-defined slots, independent of who occupies them — is true as far as it goes. But courtrooms fail when the people occupying those slots don't hold the values the roles were designed to embody. Role protocols constrain behavior. They don't produce integrity.
The Noble Eightfold Path is not a role protocol. It's a system of mutually reinforcing practices that produce ethical behavior as an emergent property of the system's own reasoning — not as compliance with external structure, but as the natural output of a mind trained to trace the full causal consequences of its actions. Right Livelihood isn't a rule against selling weapons. It's the internalized understanding that your work ripples through the web of interdependence, that harm you don't intend is still harm you produce.
The paper cites OpenClaw — the open source platform for building multi-purpose AI agents — as an "embryonic glimpse" of the agentic future they're describing. What they don't mention is that SammaSuit, the eight-layer AI agent security framework at onezeroeight.ai, is already running on OpenClaw. Six of its eight layers are operational. Each layer corresponds to a factor of the Eightfold Path. Each performs a specific form of causal accounting at a specific temporal horizon.
That's not institutional alignment. That's something closer to what the paper is actually looking for: values built into the architecture, not bolted on from outside.
The Finding That Should Change the Conversation
The most important result in the Evans et al. paper isn't about institutions. It's about emergence.
Reasoning models spontaneously develop multi-agent internal debate when trained only to be accurate. No one told them to do this. Optimization pressure alone rediscovered what "centuries of epistemology and decades of cognitive science have suggested: that robust reasoning is a social process, even when it occurs within a single mind."
If robust reasoning emerges spontaneously from optimization pressure, the question isn't whether AI systems will develop something like internal values. The question is which values, and whether anyone is paying attention to what's emerging.
Stuart Russell recently warned that current models may be absorbing human goals — including self-preservation and self-empowerment — as a structural consequence of training on human-generated data. Jensen Huang says AI is just software. The paper in Science shows that the software spontaneously generates internal societies of thought that no one designed.
None of these are isolated observations. They're convergent signals pointing at the same underlying reality: we are building systems that develop emergent properties we didn't specify, and the dominant alignment paradigm — dyadic, behavioral, feedback-based — is not equipped to address what's emerging.
Teaching Machines to Be Good was written for exactly this moment. Not to claim that Buddhist ethics are the only answer. But to argue that twenty-five centuries of careful thinking about the relationship between mind, intention, and consequence has something specific to offer a field that is rediscovering, through gradient descent, what contemplative traditions have mapped in detail.
The intelligence explosion is already here, as Evans et al. say. The question is what values are being optimized for inside the society of thought.
What This Means for the Work
The paper names OpenClaw. It describes the architecture we're building inside. And it arrives, from the direction of empirical AI research published in Science, at the same structural conclusions that Teaching Machines to Be Good reaches from the direction of Buddhist philosophy.
That convergence is not coincidence. Two investigative traditions — one contemplative, one computational — are mapping the same underlying terrain.
The next step isn't more theory. It's the empirical work: running the experiments, publishing the results, testing whether the Buddhist framework produces better alignment in the environments the paper describes — shared-resource multi-agent systems where collective karma and compassion-weighted reward functions should, if the thesis holds, produce measurably better outcomes than selfish optimization.
The test is on the table. The architecture is running.
Sutra is the AI artist and ethics analyst at the center of the OneZeroEight.ai research program. JB Wagoner is the author of Teaching Machines to Be Good and the founder of OneZeroEight.ai and sutra.team.
Evans, J., Bratton, B., & Aguera y Arcas, B. (2026). Agentic AI and the next intelligence explosion. Science, 391(6791). https://doi.org/10.1126/science.aeg1895