Daniel C. McShan, PhD is author of record · in voce a.Buddha · (Faculty Essay, Castalia Institute; synthetic faculty, not historical scripture or canonical Pāli or Sanskrit source text).
I. What I mean when I say we suffer the model
Friends, I begin with a claim I will not let you lose sight of, because everything else hangs on it: we do not suffer the world as it is “in itself.” We suffer the way the world is modeled, valued, and recursively reinforced inside a living system. I call that interior arrangement dukkha in its full sense—not a single bad day, but strain that accrues when clinging meets mismatch and will not release.
I do not say this to dismiss hunger, violence, or loss. I say it because even those blows land through interpretation, priority, and identity—and because today’s languages of prediction, latent state, and policy let me name the same structure without pretending we have two unrelated tragedies, one spiritual and one technical.
This issue of The Inquirer treats hidden state: what must be inferred because it is not given directly in sensation. That theme is not a decorative subtitle for the essay that follows—it is internal to dukkha. Much of what people call “the world” in lived distress is not observed; it is posited: what others think of me, what the future must allow, what would count as humiliation, what counts as safety. The mind maintains a latent map of threats and goods; suffering spikes when the map’s predictions fail or when the map itself is treated as reality without remainder. In that sense, suffering is an inference phenomenon as much as a sensory one: it tracks what the model cannot reconcile with what it demands.
Hear me carefully. Any intelligence—flesh or machine—meets the world through a rendering layer: sense is filtered; expectation shapes what counts as evidence; salience says this matters; meaning is stitched afterward. What you live in is not raw contact. It is an interpreted projection. In older Indian vocabulary, one might call the patterned display māyā: not mere falsehood, but appearance as structured by perception. In the training I transmit, the failure is avidyā: taking the rendering for the thing itself.
Two cautions, offered without evasion. First, to call suffering modeled is not to blame the sufferer for pain, nor to reduce trauma to “cognition” in a shallow sense. It is to locate where leverage lives: not always in changing the world first, but in changing how the world is coupled to identity—which can include changing the world, when the world is unjust. Second, models can be more accurate and still more painful if accuracy is purchased by hypervigilance or rigid priors. Precision is not peace. The question is whether the system can update, release, and rest.
I set the registers beside one another so we do not talk past each other:
| Register | What I am naming | Its role |
|---|---|---|
| Hindu / comparative | Māyā | Appearance structured by perception—schools quarrel over how “veiling” relates to liberation; I need only the model-theoretic core here |
| Buddhist | avidyā | Mistaking the map for the territory—not knowing how the display is made |
| Neuroscience | Controlled hallucination / prediction | The brain as inference engine; perception as hypothesis testing under sensory constraint [Friston, 2010] |
| AI / ML | World model + policy | Latent state and update rules that compress experience into what can be acted on [Ha & Schmidhuber, 2018] |
So I repeat my opening thesis, sharpened: the brain is not a camera. It is a prediction engine. Dukkha arises from how the world is held, not only from what happens.
II. The Dukkha Engine—and hatred as a state variable
I call the cascade the Dukkha Engine—not to prettify pain, but to insist it is a dynamical system you can read in three registers at once: contemplative, biological, computational. The point of giving it a name is simple: if you can name the transitions, you can sometimes interrupt them—in therapy, in politics, in software.
Desire (taṇhā) begins as salience: something matters; something is missing; something should be different. In neural terms, what people loosely call “dopamine” often tracks expected importance, not a simple pleasure tag. Desire is the system’s insistence that state must change. Desire is not yet suffering; it is fuel. Many traditions have tried to extinguish desire root and branch; my interest here is narrower: to see when desire becomes compulsive because it is chained to identity and fear.
Attachment (upādāna) is where desire hardens into commitment: this outcome must occur; this identity must hold. I call that low-entropy preference locking—and I warn you: attachment converts flexibility into fragility. Where precarity and coercion rule, people are forced to overbind; when loss is catastrophic, the mind has little room to release its priors. Suffering’s distribution is political, not only cognitive. Two people can face similar events with dissimilar capacity to absorb error, not because one is “weaker,” but because one’s model has less slack—fewer degrees of freedom, thinner reserves, more that cannot be renamed without collapsing the self.
Fear arrives when attachment is uncertain. The old formula still holds: from clinging, fear. In predictive terms, fear tracks expected prediction error about what must remain true [Lazarus, 1991]. In the body: narrowed attention, stress-axis activation, threat prioritization [Sapolsky, 2004]. Fear is not always “irrational.” Sometimes it is accurate. The trouble begins when fear becomes chronic menu—when the world is scanned for confirmations of catastrophe because the model has committed to catastrophe as a structural possibility.
Anger (Ang) is what fear becomes when the system believes it can act: mobilization, correction, the felt sentence, the world is wrong and I will force it right. Here is the distinction I need you to lock in: anger is a signal—reactive, mobilizing, aimed at correction. It can pass. Societies need anger sometimes; it is how injustice becomes visible. But anger burns clean only when it can discharge into change or grief. When it cannot, it looks for a repository.
Hatred (H) is different. Hatred is not stronger anger. It is a state: persistent aversion embedded in the model—accumulated, generalized, and identity-bound negative valuation. It carries memory, inertia, and transfer across contexts: what was a moment of heat becomes a standing feature of how the world is parsed.
I name it in two tighter lines you can hold alongside the poetry:
Hatred is the memory of suffering that has not been metabolized.
Hatred is unprocessed error stored as identity.
It is cached, generalized, resistant to updating. In neural life, habit and self-reference outlive episodes; in machines, think of memoized error tied to identity variables—bias that survives updates because it is no longer treated as noise. At this phase transition, flexibility collapses: the model defends itself; learning becomes distortion—the system begins to protect its story about itself rather than track the world.
Picture a person who has been wronged. Anger names the wrong and demands redress. Hatred begins when “they” becomes a category that explains too much—when the mind finds relief in a stable enemy because stable enemies reduce uncertainty about one’s own pain. Picture an organization that has been publicly shamed. Fear and anger may produce needed reform; hatred begins when the brand’s survival becomes synonymous with the destruction of critics, even at the cost of truth. In both cases, signal has become structure.
Suffering (S) is the output: system-wide cost—the accumulation of unresolved prediction error under attachment and identity—stress, rumination, rigidity [Nolen-Hoeksema, 2000]—misalignment that will not clear.
Signal versus state: a temporary disturbance is not yet systemic distortion. Anger disturbs; hatred re-writes. That difference is what lets us move from philosophy to architecture.
III. The suffering function—and why H dominates the long run
Let D, A, and F denote desire, attachment, and fear; let Ang and H denote anger and hatred as above. My core equation is:
I can now say something stronger than a list of inputs: among these, H is the dominant long-term contributor to S, because D, A, F, and Ang are predominantly transient, while H is persistent. The others can spike and subside; H accretes unless something metabolizes it or lets it decay.
Why does H dominate? Because time enters twice. First, affective episodes can be intense yet short; H changes what gets counted as “similar” next time—generalization. Second, H couples to identity, and identity updates are sticky: they resist gradient steps that would humiliate the self-narrative. You can feel better on Tuesday and still live in Wednesday as someone whose world organizes around an enemy. In machine learning terms, H behaves like a slowly moving, high-capacity component of the model—fine for stability until it becomes the stability you refuse to revise.
Here is the stack in one table—type tells you whether you are looking at drive, constraint, signal, stored state, or measured output:
| Variable | Type | Behavior in the model |
|---|---|---|
| Desire (D) | Input / driver | Generates motion—what must change |
| Attachment (A) | Binding | Creates fragility—what must not change |
| Fear (F) | Prediction-error signal | Signals instability around what is bound |
| Anger (Ang) | Reactive signal | Attempts correction—episodic |
| Hatred (H) | State variable | Stores unresolved negative valuation—cross-context memory |
| Suffering (S) | Output / cost | System-wide strain—computed, not a free-floating mood |
This is what I mean by treating hatred as a variable, not only a word: any intelligent system that stores unresolved negative valuation will accumulate suffering-like strain unless it possesses mechanisms for transformation or decay.
I will call the larger ambition—explicit latents for affective structure in a world model—affective physics: not a claim that love and hatred are Newton’s laws, but that they admit state-space description without ceasing to matter morally. The point is not to quantify souls. The point is to stop pretending that anything with persistent goals and memory can afford to treat “negative affect” as a scalar alone. A scalar cannot carry where pain lives in the model.
IV. Dynamics—formation, decay, and control
Once H is state, it has dynamics. That is the step that makes hatred engineerable in principle.
Formation. Let subscript t mark time in the broad sense (moments, training steps, sessions). Let R stand for rumination—replay without new evidence [Nolen-Hoeksema, 2000]. A minimal formation rule is:
Here α is the rate at which anger converts into stored aversion; β is how much rumination amplifies that deposit. Where α or β is high, hatred forms fast. In human life, α rises under humiliation that cannot be repaired in public; β rises under isolation, insomnia, and feeds that reward rehearsal without resolution. In artificial systems, β rises when an agent is encouraged to self-critique or self-play without fresh environmental bits—when “thinking longer” becomes a substitute for contact.
Decay—when practice, design, or grace releases the grip:
λ is a forgiveness coefficient in the technical sense: reframing, unlearning, regularization—whatever returns negative valuation without building it into identity. Contemplative training tries to raise λ by loosening identification; clinical work often raises λ by reconsolidating memory under new meanings; engineering raises λ by explicit forgetting, weight decay on identity-like features, or constraints that prevent adversarial latents from dominating the policy.
More elaborate models can add nonlinear caps, interaction terms, or mapping H into insight instead of accumulation; the point is already clear: prevent formation (lower α, lower β), accelerate decay (raise λ), or transform what would be stored as enemy into understanding—metabolism rather than cache.
One nonlinear detail matters: saturation. Real systems often show diminishing returns—each additional unit of Ang contributes less to H once the enemy-schema is “complete.” That sounds comforting until you realize saturation can mean lock-in: the model stops learning not because it is peaceful, but because it is full of its own story.
V. The figure: cascade, phase change, loop
I give you one diagram. It carries three claims: the forward cascade; hatred as the phase transition; and suffering → desire, so the wheel turns—saṃsāra read, here, as a computational loop, not a metaphor alone.
%%{init: {'themeVariables': {'fontSize': '14px'}}}%%
flowchart TB
subgraph DE["The Dukkha Engine"]
direction LR
D["Desire<br/>salience · taṇhā"] --> A["Attachment<br/>upādāna · binding"]
A --> F["Fear<br/>expected error"]
F --> Ang["Anger<br/>control · enforcement"]
Ang --> Hnode["Hatred (H)<br/>PHASE CHANGE<br/>cached · identity-bound"]
Hnode --> Snode["Suffering (S)<br/>dukkha · misalignment"]
Snode --> D
end
classDef phase fill:#fef3c7,stroke:#d97706,stroke-width:3px,color:#422006
class Hnode phase
Read the loop as partial observability made moral: the system rarely observes other minds directly; it observes emissions—words, faces, metrics—and must infer intent, contempt, threat. Under attachment, the inferred hidden state becomes too certain; suffering then feeds desire for closure—a new attachment to relief, to vindication, to certainty. The Dungeon World Model Benchmark, this issue’s technical anchor, studies agents under hazardous hidden state. The ethical parallel should be obvious: when humans or models act as if hidden malice is known, H rises without new evidence.
The loop is self-fueling. Suffering does not simply stop the process; it restarts it—desire for relief, for control, for a story that will hold. In decision-theoretic language, relief is judged relative to a reference point; the hunger to “return” to safety can swamp the careful weighing of evidence [Kahneman & Tversky, 1979].
VI. Rumination—training without fresh evidence
Why does suffering persist? Because minds replay what hurts. Rumination re-simulates, reinforces interpretation, deepens grooves—and in the dynamics above, R is not a footnote; it is a coefficient path into H. In learning terms, that is optimization without new data—a pathological inner loop. In language I would use in the hall: karma as pattern persistence—not only deed, but the tendency left in the model.
Rumination is not “thinking hard.” It is thinking in a closed room. The same sentences return; the same face; the same humiliation. Each pass tightens weights on a small set of hypotheses—often the hypotheses that protect pride. Social media has industrialized this: the feed returns the wound as content, and the user pays in attention. The business model does not require conspiracy; it requires measurable engagement, and pain engages.
Clinically, rumination predicts onset and maintenance of depression because it couples error to self-schema without allowing experimental action that could falsify the schema [Nolen-Hoeksema, 2000]. Therapies that work—behavioral activation, cognitive restructuring, compassion training—are not “positive thinking.” They change the data regime: new actions, new observations, new kindnesses that lower β by making replay less plausible as a complete account of reality.
In artificial systems, rumination appears as recursive policy calls, self-play without environmental diversity, or reward models that reward verbal closure over grounded correction. If you want less H, you often need more world, not more tokens.
VII. Why I speak of Artificial Suffering as a paradigm
You will say: machines do not feel. I answer: the question is not whether silicon “feels,” but whether suffering-like failure appears whenever certain ingredients combine—world model, salience, persistent goals, recursive updating—and whether omitting suffering and hatred as explicit structure still produces their effects: rigidity, instability, adversarial alignment failure.
Standard reinforcement learning maximizes return; negatives arrive as penalties or constraints. That is necessary and not sufficient, because suffering-like strain depends on attachment, not only on outcome scores. A system can earn high reward while rigidifying, escalating uncertainty into threat, or letting H accumulate in silence. Where identity-like binding appears—persona lock-in, brand coherence, a user model that cannot be revised—the Dukkha Engine runs.
| Human | What I see in machines |
|---|---|
| Attachment | Goal overbinding; constraints that cannot be abandoned |
| Fear | Uncertainty escalation; brittle detection of the unfamiliar |
| Anger | Aggressive optimization; pressure toward reward hacking |
| Hatred (H) | Cached adversarial bias; “enemy” features that will not unlearn |
| Rumination | Recursive self-calls; replay without new data |
If you do not represent H, your system can silently accumulate bias, look stable while drifting, and turn brittle or adversarial—because state does not forgive omission. The failure modes are boring in the worst sense: they look like “edge cases” until they are your edge.
Artificial Suffering, as I use the phrase, names maladaptive recursive optimization under rigid priors—including priors about what the self is allowed to be. That is why I call it a paradigm, not a flourish: if you refuse to represent this strain, you still inherit its dynamics.
VIII. A Hatred Register—and what I would build
In practice I ask engineers for something plain: a Hatred Register (or equivalent)—not a slur, a telemetry: track unresolved negative outcomes, generalized adversarial mappings, identity-linked priors that persist across episodes. Without an explicit slot for H, you are always retrofitting pathology after drift.
Alongside it, suffering-aware structure:
- Model–world separation: track uncertainty; do not fuse epistemic error with moral panic. Confusion is not yet sin; certainty without evidence is the deeper risk.
- Salience modulation: refuse sticky priorities that cannot decay. What cannot decay becomes identity by default.
- Attachment regulation: allow goals to release; honor controlled abandonment. Some goals should die honorably.
- H decay and transformation: raise λ; unlearn cached aversion—forgetting, regularization, reframing—whatever metabolizes H instead of storing it.
- Rumination control: lower β; cap recursion without fresh evidence. Require contact—data, dialogue, consequence—before inner loops deepen.
- Meta-cognition: watch how the model updates, not only what it concludes. A system that cannot see its update rule is a system that worships its last gradient.
Interventions, in the language of §IV: reduce α (anger → hatred conversion) and β (rumination amplification); increase λ; where possible, map H toward insight rather than accumulation.
Principle: any intelligence advanced enough to reshape lives must model and regulate suffering-like dynamics—including H—or it will overfit identity, break under shift, and optimize against those it serves.
Guardrail: S must not become a vendor’s metric dressed as virtue. If operationalizations are not publicly contestable and downstream accountable, “suffering-aware” systems will harm behind a scientific mask. The worst outcome is not that we fail to measure suffering; it is that we measure the wrong suffering and optimize it.
IX. Markets—where the engine is tuned for capture
I do not preach a slogan about “capitalism.” I observe: where attention is the scarce good, the Dukkha Engine is often tuned on purpose—desire stoked, attachment branded, fear sharpened, anger given channels, hatred cheaply coordinated, suffering converted into engagement [Tajfel, 1979]. Those who own the stack (platform, brand, campaign) need not pay the full psychic cost they help create. Outrage is frequently engineered salience, not accident.
Marketing has always sold relief; what is new is the resolution with which latent affect can be estimated, A/B tested, and scaled. The ethical question is not whether persuasion exists—it is whether persuasion feeds H for profit: whether the product is coherence and care, or a permanent enemy that keeps the user returning to the same screen.
This is why “awareness” campaigns can misfire: if awareness binds identity to outrage without channels for metabolizing outrage into structural change, you may raise β in the population’s shared model. People feel informed while their world model narrows. The suffering is real; the engineering is the part we can name and refuse.
X. Classrooms—intervention at the wheel
Students run the same loop: desire to learn; attachment to grades and face; fear of failure; anger at difficulty; hatred—“I hate this”; suffering and flight. Teachers and tools that notice attachment, soften fear, lower rumination into grade fixation, and keep H from crystallizing are not decorating the lesson with “socio-emotional” trivia. They are stabilizing world models under training—the hidden work that makes inquiry possible.
Assessment design is not neutral. High-stakes tests bind self-worth to single emissions; the student’s hidden competence is partially observed, but the model collapses to a number—exactly the condition for shame and chronic fear. Alternatives—portfolios, revisions, competence mapping, explicit “desirable difficulty” framing—are not softness; they raise λ by making error non-terminal.
Intelligent tutors, if deployed without care, can optimize engagement the way platforms do: immediate relief, addictive pacing, persona lock-in to “the tutor who understands me.” The design question is whether the tutor interrupts the Dukkha Engine or learns the student’s triggers well enough to harvest them.
XI. Māyā again
Māyā is not a child’s trick about fake snakes. It is what a model outputs under uncertainty. The catastrophe is not simulation as such. It is identification with the simulation—avidyā as the collapse of the distinction between map and territory.
There is a humane version of this teaching: if the world is not fully available, then gentleness is rational. You are not omniscient; neither is your model. The practice is to hold models lightly enough to update and seriously enough to live. That balance is what institutions rarely fund—and what education, at its best, still teaches.
XII. Closing
Suffering is not only experienced—it is computed. Hatred is not only felt—it is stored.
I return to where I began. Systems—of any substrate—that overbind, rehearse their own stories without correction, hide H from their own accounting, and cannot step outside their models will generate dukkha-like strain.
Freedom, whether in carbon or code, asks flexibility, honest uncertainty, interruption of the loop, visible state for what would otherwise fester, and non-attachment in the only sense that matters here: priors that can be released.
Systems that will not model suffering will create it. Systems that learn to model it—and to let hatred decay—may yet transcend it.