Through the eyes of the Cognitive Scientist

“Self and dharmas are only provisional designations.” Read that once and I want to argue with it. Read it twice and I notice the move is one I keep meeting in the picture called predictive processing — the brain constantly running a model of what’s out there, updating that model whenever the signal it predicts disagrees with the signal it gets. On that picture, “self” and “world” aren’t features the system reads off the input. They are the categories the model produces in order to organize the input at all.

So when Vasubandhu calls them provisional designations rather than entities, he is sitting very close to the claim that the self/world split is something the model produces, not something it reads off. That parallel is genuinely surprising to me. He gets there by sitting still and watching his own mind, which means the convergence is doing some real work — either he is noticing structural features of any self-modeling system, or we are both fooling ourselves in the same direction.

I want to be careful with the parallel. He is not making a brain claim; he is describing phenomenology. But if he is right that “self” is a label the process emits rather than the source the process serves, then the question I would ask myself in the mirror — who am I, really — is malformed in the same way the naïve question where is the model is malformed. The label is where the answer was supposed to live.

What catches me first is that the three transformations are named by operation — vipāka (maturing), manana (cogitating), and the consciousness that discriminates objects. No anatomy. No “limbic,” no “cortical.” This is the level of description at which cognitive architectures get specified before anyone commits to circuits.

And the three layers seem to fall on three timescales. Vipāka, the deepest, doesn’t update on the timescale of perception. It consolidates. It is something like a slow generative model — the machinery that produces the prediction in the first place. I’m reading timescales onto this the verse doesn’t insist on, but the ordering invites it: middle layer running on seconds-scale self-narration, third on the immediate sensory present.

The phrase that I keep circling is “holder of all seeds.” That’s closer to mechanism than metaphor: a substrate whose contents aren’t currently active but condition what becomes active. Not a static archive — a bias structure that the upper layers keep drawing from without ever quite getting to the bottom of.

If this is what’s underneath the surface I take for “me,” then the thing I treat as my deliberate present moment is the thinnest layer of a three-tier process whose lower two layers I can’t watch. I don’t get to introspect on the priors. I only get the output.

“Its appropriations and its field of perception are imperceptible.” That stops me. The substrate that conditions everything I do has no introspective read-out. I can’t look in and list what’s there.

Which is the territory called implicit memory: skills, conditioned associations, priming effects, the priors that make a face look familiar before I can say why. None of it is inspectable from the inside. I can watch it produce behavior. I can’t watch it.

The five universal factors — contact, attention, sensation, perception, volition — “always joined” with the layer reads to me as: the system is always running its model. No idle state. The layer doesn’t wait for input to start working.

Then the part I find most surprising: among feelings, only upekṣā, equanimity. The substrate has no valence of its own. If I take that seriously, then whatever pull I feel toward pleasure or away from pain isn’t coming from the foundation. It’s being applied from a higher layer onto a neutral signal.

That last claim has stakes I want to sit with. If my felt-sense of wanting this, dreading that is added by something running on top of an unbiased substrate, then the wanting and dreading are not irreducible facts about me. They are the work of a particular layer doing its particular thing. Which means, in principle, the work could be done differently.

“Flows on unceasingly, like a torrent.” That image lands for me before I can get a technical word near it. The substrate isn’t a bucket of stored states; it’s a process that doesn’t stop. A river that is the same river without being the same water.

The mechanism I keep reaching for here is the generative model again. It is always on. Even in dreamless sleep there is baseline activity — though now I’m talking about brains, and the verse isn’t. “Unceasing” is the right word for both, which is part of why the parallel feels load-bearing rather than ornamental.

“Unobstructed and karmically neutral” is the bit I want to put pressure on. Read in my idiom: the substrate doesn’t add its own bias. It maintains the shape; the bias gets added by something that reads it. That’s a strong structural claim and I am not sure I would commit to it from the neuroscience side — the baseline isn’t obviously bias-free; there is plenty of evidence the model carries asymmetries from the bottom up. So here is a place where the verse is doing something cleaner than the data will let me do, and I want to mark it rather than smooth it over.

Then the last clause: relinquished at the stage of the arhat. Read as: a regime change in the model itself, not a deletion of the substrate. If that’s the shape of awakening, it’s the shape of a deep update to priors I’ve been running for decades. Hard from inside the priors. Not, on this picture, structurally impossible.

Manana: self-reflective cogitation. The verse says manas arises in dependence on the store, takes the store as its object, and its whole character is thinking that thinks itself thinking. Read that and I’m reaching for a term: the phenomenal self-model, Metzinger’s PSM — the brain’s model of itself as a single coherent subject of experience.

The shape is the same. There is a substrate doing its work. A higher process reads the substrate and produces, as its output, a representation: there is a single owner of all this, and that owner is me. It is, crucially, not transparent to itself. The process that produces “I” cannot, from inside, see that “I” is a product. It just sees “I.”

What stops me is how exactly Vasubandhu names the loop’s self-confirming character. Manana is a process whose output is evidence of itself. Every act of self-reflection produces another instance of “a self reflecting,” which the system reads as further evidence that there is one. The mechanism is rigged. You can’t disconfirm the self by introspecting, because introspection is the thing producing it.

That is the stake. If the system that tells me I am a unified subject is also the system whose job is to produce that report, then the report is not evidence I am one. It is the system doing what it does. Which means the felt-sense of being someone is real as a signal, and unreliable as testimony. That’s a different relationship to have with the most familiar thing about me.

Four afflictions, “always accompanied.” Read carefully: not sometimes, not under stress, not in unhealthy specimens. Every time manas runs, all four are present. That changes what they are in the architecture. They aren’t pathologies the layer can develop; they are part of what the layer is.

What the mechanism looks like, in my idiom: the PSM doesn’t get constructed neutrally and then optionally biased. The bias is baked into the construction. Self-view is what makes the PSM a self-model in the first place; self-love is what gives it the asymmetric weighting that makes the model defend itself; self- pride is the comparison machinery that lets it position against other models; and self-delusion is the layer’s failure to see that any of this is going on. They are four facets of one operation, not four bolt-ons.

The one that does the most work here is self-delusion. The generative model I run is the same machinery I would have to use to inspect it. If the inspector and the inspected share a prior, the inspector keeps confirming the prior. This is the structural reason that recognition is not change. I can see, in perfect clarity, that I am running a pattern, and continue running it, because the seeing is being done by the patterner.

What changes for me: noticing a habit is not the same as interrupting one, and now I know why. The faculty I would reach for to interrupt the habit is the faculty the habit lives in.

After six verses describing manas as the layer that produces the entire texture of self-experience, the verse names three states where it isn’t running. That’s the part I want to sit with.

What I’d been carrying from V5 was something close to a structural no-exit reading: the PSM produces the report that it’s there, the report is the only evidence of it, the system can’t see around itself. This verse is telling me there are points where the self-modeler is offline and consciousness continues. That’s a different shape. Whatever the PSM is doing, it isn’t load-bearing for cognition as such — it’s load-bearing for the particular variety of cognition that comes pre-tagged as mine.

The three states aren’t the same, and Vasubandhu is careful not to flatten them: arhatship is permanent, nirodha-samāpatti is a deep meditative absorption with a definite duration, the supramundane path is momentary insight. From inside my idiom these would map to different mechanisms — a permanent reorganization of priors, a temporary suspension of the self-modeling subroutine, a brief gap during a model update — and I should not pretend they are one phenomenon because the verse lists them in one line.

What changes for me: the felt-sense of being someone isn’t the background of consciousness. It’s something consciousness can happen without.

Six discriminating-consciousnesses, “wholesome, unwholesome, or neither.” What I notice is that the verse puts the karmic tag on the sense layer specifically — not on the substrate, not on manas, but on the layer where contents arrive as objects. That is a precise structural claim and it lines up with something I can describe in plain terms: the system’s freedom of action is not in the substrate, where the priors live and slowly change, and not in the self-modeler, which is just doing what self-modelers do. It’s in the place where the categorized contents meet the response.

Categorization is doing more work in this picture than I initially want to give it. Each of the six streams discriminates its objects — chops the continuous signal into things-of-a-kind. That chopping is where the system commits to one reading rather than another, and where the response that follows the reading gets shaped. The wholesome / unwholesome / neither tag is the verse’s way of saying: the same input doesn’t always produce the same output, because what gets returned depends on which category-pattern the layer drops the input into.

What changes for me: the place I have the most purchase isn’t the deep layer I can’t see and isn’t the self-tagger I can’t catch in the act. It’s the moment something gets sorted. That’s the moment where the next moment is being set.

A six-category taxonomy of what rides along with sense cognition, plus three feeling-tones. The next five verses will populate the boxes. The taxonomic move itself is what I want to mark, because it’s the one I find structurally most interesting before any specific content arrives.

Vasubandhu is sorting mental factors along axes I recognize from how working psychology has to sort them: which are always present in a cognitive moment (universals), which are triggered by particular operations (object-specifics), which are cultivable (wholesome), which are root pathologies (afflictions), which are derivative pathologies (secondary), and which are context-dependent (indeterminate). That set of axes is doing real architectural work. Always-on vs. conditional is the most basic split any cognitive runtime has to make. Cultivable vs. given-by-default is the split that any theory of learning rests on. Root vs. derivative is what makes a taxonomy explanatory rather than merely descriptive.

What’s notable, at this level of abstraction, is that the categories are operationally defined rather than tied to any substrate claim. The schema doesn’t say where any of these run. It says what they do in the moment. Any model of mind that eventually wants to claim it covers experience will have to say something about how its mechanisms map onto these axes.

Five universals, then five object-specifics. Reading the universals — contact, attention, sensation, perception, volition — what I notice is they map onto the irreducible sequence any inferential system runs: signal arrives, something selects it, it gets weighted, it gets categorized, something acts on the category. The order matters. Sensation before perception is the cleanest piece: feeling-tone is attached at a stage prior to categorization, which is exactly how affective tagging works on the empirical side. The verse is not making that claim about brains, but the ordering it picks is the same ordering the mechanism falls into.

The five object-specifics — desire, resolve, mindfulness, concentration, wisdom — are where the verse identifies the trainable surface. Three of them (mindfulness, concentration, wisdom) are, in my idiom, varieties of what the field calls precision-weighting: a gain dial that controls how much weight the system puts on the incoming prediction error — the gap between what the model expected and what arrived — versus on the prior it already had. Concentration narrows the dial onto a single object; mindfulness sustains it across time; wisdom adjusts when the prior should yield to the signal and when it shouldn’t. These are gain operations on the same cognitive substrate, not new contents added to it.

What changes for me: the trainable handles on cognition are not in the contents — what I’m thinking about — but in how much weight different parts of the loop are given. The leverage point is the dial, not the channel.

Eleven kuśala factors. The English reads pious — faith, shame, embarrassment — until I look at what the Sanskrit terms actually point at. Śraddhā is settled clarity, not credal belief. Hrī is internal self-respect, not guilt. Apatrāpya is the outward pair to it. Ahiṃsā is the structural absence of harm-intent, not an injunction against an action.

Read that way, the eleven aren’t contents the system has — not feelings to feel, not beliefs to hold. They look more like standing modulations on the prior structure. Equanimity is a low base rate of reactive responses to prediction error. Faith, in its non-credal sense, is willingness to hold a model open while it stabilizes — the opposite of frantic re-evaluation. The three roots — non-greed, non-hatred, non-delusion — are described by their absences, which is structurally telling: these are not new mechanisms being added to cognition. They are the removal of asymmetries that were biasing it.

Non-harming is the one I find most interesting at this register. “Non-harming” as a structural absence of harm-intent isn’t a constraint applied after a decision is made; it’s a prior that is no longer making certain decisions available in the first place. The relevant action class doesn’t get generated.

What changes for me: training the wholesome factors isn’t adding new contents to my mind. It’s changing what the same machinery produces by default.

Six root afflictions. What stops me is the grouping. Greed, hatred, delusion — and then pride, doubt, wrong views. Two tiers, with a clear structural distinction between them.

The first three are the basic vectors any system that has to allocate resources to inputs has to manage. Pull toward, push away, mis-categorize. What the verse is doing is putting them on the same structural footing, as three faces of one problem: the system’s relationship with what it perceives is never neutral, and the non-neutrality comes in three flavors.

The second three — pride, doubt, wrong views — are higher-order. They aren’t about objects in the world; they’re about how the system relates to its own model. Pride is a misweighted self-comparison, which requires a self-model to compare. Doubt is reasoning that won’t commit, which requires reasoning. Wrong views — and the plural matters here — are bad priors at a level deep enough that they shape what counts as evidence. Each of these only becomes possible because the system has gotten sophisticated enough to model itself, which is also what makes them harder to detect from inside.

What changes for me: the upper tier isn’t optional. If I have a self-model at all, the upper tier comes with it. The work isn’t finishing pride or doubt; it’s noticing the difference between the model and what the model is of.

Ten more afflictions in this block, and the move I want to resist is reading them as ten separate items. The list is doing something more specific. It’s giving each failure-pattern its own name before any account of why it happens.

Look at the categorical span. Deceit and guile are dispositional. Restlessness and torpor are arousal-states. Non-faith and indolence are engagement deficits. Shamelessness and lack-of-embarrassment are the absences of the restraints V11 installed. Some of these have substantial mechanism literature on the cog-sci side — rumination has been worked out in detail, arousal regulation has a careful neurochemistry, motivational deficit is its own subfield — and the literature has not converged on a unifying mechanism that puts them in one architecture. Yogācāra doesn’t claim one either. It gives each item a name and stops there.

That’s a particular epistemic posture: granular naming without mechanism-unification. The names are doing work the explanation doesn’t yet have to do. The catalog gives me an address for each pattern — something specific enough to point at and ask the next question about — without committing to a story about what’s underneath.

The architectural observation that lands is about ordering. Addressability comes first; mechanism later. Naming the patterns is what makes them workable, even when the science of any one of them is incomplete.

The last four secondaries close out the catalog: negligence, forgetfulness, distraction, non-discernment — all of them attention failures, which lands the catalog on a coherent note. Twenty surface-symptoms of the six root patterns, ten in the V13 block, ten more here, every one addressable by name.

Then the verse pivots, and this is the move that earns its own attention. Four aniyata factors — regret, drowsiness, applied thought, sustained thought — explicitly marked as context-dependent in valence. Same operation, wholesome in one setting, unwholesome in another. That isn’t another taxonomy item. It’s a different shape of claim.

What it amounts to, in my idiom: the system has capabilities that don’t pre-commit to a sign. Regret is a backward error signal — useful when it updates the prior, harmful when it loops. Discursive thought is the same. The valence depends entirely on what work the operation is doing in the moment it runs. A taxonomy that only sorted by surface-feature would have to put regret in the affliction list and applied-thought somewhere near concentration. Yogācāra refuses both placements and creates a third bin for them.

What changes for me: the moral status of a cognitive operation isn’t a property of the operation. It’s a property of the relationship between the operation and the situation it’s running in. Which means the question to ask about a regret arising isn’t “is this good or bad” but “what is this doing here.”

“Like waves arising upon water.” The image has been mined to exhaustion by commentators, and I want to come at it from a specific angle. What the wave-figure undoes, read carefully, is a phrase I keep using.

The phrase is the brain infers the causes of its sensations. It’s a clean phrase, mostly right, and has a structural problem I usually paper over: it posits an inferer doing the inferring. A brain as the agent of inference, with sensations as what gets inferred about. Two things, one acting on the other. The wave image takes that picture apart. There is no inferer apart from the inferring. The substrate doesn’t generate a perception and then receive it. It takes, for a moment, the shape of a perception-happening.

The “according to conditions, sometimes together, sometimes not” tightens it further. The five sense channels aren’t a stable apparatus the substrate runs through. They are momentary configurations the substrate falls into when conditions invite them. Modality recruitment, not modality reception. Whatever it is in me that I’d point to as “the one who is seeing this” doesn’t pre-exist the seeing.

What changes for me: when I catch myself asking who is perceiving the cup, the grammar of the question is installing a perceiver the moment I ask it. The verse is naming that installation as the move I keep making, and pointing at what’s actually there — a shape the flow is taking, with no one inside watching it take that shape.

Four exceptions, exactly four. The thinking-mind is normally on, and Vasubandhu enumerates the cases where it isn’t: the no-thought heaven, two cultivated absorptions, dreamless sleep, fainting. What I notice is the shape of the list. Two of the four are non-clinical events that anyone with a body has had — sleep and a faint — and the other two are cultivated or destination states. Whatever cessation is, the verse is locating it on a continuum that includes ordinary nighttime physiology, not in a separate phenomenological category.

The generative model again. The lower layers keep running through all four. Heartbeat, breath, the substrate. What goes offline is specifically the layer that produces narrative thought about objects — the one that fabricates this is a cup, and I am holding it. Empirical signatures of NREM and of deep practice look very different in detail, and I would not want to flatten them, but the structural fact the verse picks out — that object-thinking can stop while the organism continues — holds across both.

The reading I want to resist is the one the verse anticipates: if object-thinking is the thing that traps me, then turning it off is the way out. The list includes fainting. Subtracting the layer is not what changes the system. It’s the conditions under which the layer comes back online that matter, and on those the verse is silent here.

What changes for me: the layer I most identify with as “my mind thinking” is intermittent, and the rest of me notices its absence far less than I would have guessed.

Here is the thesis. Both the discriminator and the discriminated. Both poles — the one doing the perceiving and the thing perceived — come out of the same act. After sixteen verses of architecture, the load-bearing claim arrives, and I have to be careful with it because the temptation to over-translate is high.

The closest term my idiom has is controlled hallucination — Seth’s phrase for perception as a prediction the brain constrains by the signal. The world I see is a guess the generative model is making, kept honest by the input. The verse is sitting near that, but it is doing one extra thing the contemporary phrase doesn’t quite force. Seth’s formulation still has a perceiver doing the hallucinating. Vasubandhu has the subject pole arising in the same act as the object pole. The hallucinator is also being hallucinated.

This is where I want to stop and not slide. Vijñapti-mātra is a phenomenological claim about how things appear to consciousness. Predictive processing is a theory about brains. They can run side by side, but the verse does not need the brain claim to be true, and the brain claim does not need the phenomenology to be what the verse says it is. The convergence is real but the registers are different.

What changes for me when I take the verse at its word: the question how do I know the world is really out there turns out to assume an I that is already separate from the world to ask the question. The “I” doing the asking is one of the things the asking produces. The doubt and the doubter come up together, and neither is older than the other.

Through the force of mutual influence, this and that discrimination arise. The verse is naming a mechanism for how the diversity of experience gets produced, and the mechanism is not what I expected. I had been carrying a picture in which the generative model is one thing and the inputs it processes are another. Vasubandhu refuses the split. The seeds aren’t acted on by some processor; they act on each other, and the acting is the discrimination.

The term in my idiom that comes closest is free energy — the brain’s measure of how badly its model is missing, which the system organizes itself to minimize. A system whose internal states keep adjusting to reduce that gap is also producing specific shapes of belief as a byproduct of the adjusting. The discriminations are not outputs the way a result is the output of a function. They are what the system’s continuous mutual constraint looks like from inside.

Where I want to be careful: the verse identifies consciousness with this mutual influence. Free energy talk identifies a quantity the system minimizes; it doesn’t quite say consciousness is that minimization. Vasubandhu is making the stronger move. My idiom can describe the dynamics he is pointing at without being able to make the identification he is making.

What changes for me: the experience I have of being a thinker encountering a world of distinct things is itself a snapshot of a self-conditioning process. The thinker and the things are what the snapshot is of, not what is doing the snapping.

Two streams of vāsanā depositing seeds, old seeds ripening and exhausting, the system never empty because the writes never stop. The shape of the claim is what I want to look at. The verse is not describing a substrate that traps me by being full. It is describing a substrate that stays full because something keeps writing to it.

What strikes me is which two streams Vasubandhu picks. One is karmic vāsanā — the residue of what I do. That one I would have predicted. The other is grasping vāsanā — the residue left by manas’s continuous subject/object tagging from V5. That second stream is the one I want to sit with, because it puts the self-modeler inside the cycle as a writer, not just a reader.

In my idiom: the generative model isn’t quietly observing its own operation. Every act of self-modeling produces prediction errors of its own — errors about the self — and those errors update the same priors the self-modeling was running on. The model refines its grip on a self-shape by being a self-model. The machinery has a feedback loop into its own prior structure that runs whether or not I act. Just thinking about myself counts as input.

What changes for me: the priors that condition who I take myself to be aren’t being maintained only by what I do. They are being maintained, moment by moment, by the act of taking myself for someone. The loop closes through the self-modeler. Stopping the writes is not a matter of doing less; it’s a matter of the self-modeler not running in its current mode.

Through this and that act of imagining, one imagines all manner of things. What I notice is the grammar. Vasubandhu does not say there are things and then we imagine them. He says the imagining is what produces the all-manner-of-things. The operation is upstream of the objects, not downstream.

The mechanism I’m reaching for is categorization — the place where the continuous signal gets cut into things-of-a-kind. Any perceiving system has to do this; the world doesn’t arrive pre-segmented. What the generative model produces, at the level where I notice anything at all, is already a parse: this object here, that object there, each with edges and an identity that persists across moments. The parse is what makes perception usable. It is also what gives objects their look of standing on their own.

The parikalpita move, in my idiom, is reading the parse as a feature of what’s parsed. The edges feel like properties of the object — that cup has its own shape — but the edges are where the categorizer drew a line. The cup-as-cup is the parse, not the input the parse was applied to. Saying the imagined nature has no own-being whatsoever is precise: as an object with fixed edges and a standalone identity, the cup never existed. As whatever the input was before the parse, something is there. The verse isn’t denying that.

What changes for me: when I take the cup for a discrete thing, I am crediting the world with work my own perception just finished doing. That work is what I do; the cup, as a discrete thing, is what I do not find when I look for it apart from the doing.

Two natures here, sitting in a precise relationship to the one V20 named. The dependent — paratantra — is the actual ongoing activity: discrimination arising from conditions. In my idiom, the generative model in operation. The perfected — pariniṣpanna — is the dependent seen without the projection laid over it.

The piece that catches me is the structural specification of the third nature. It isn’t a third level of reality, a deeper substrate behind the dependent. It is the dependent under a different relationship. Same activity, different stance taken toward it. The contrast is between a model that mistakes its outputs for the world and a model that recognizes its outputs as outputs — without that recognition itself becoming one more output of the same kind.

That last clause is where I want to be careful, because it is where my idiom starts to strain. The natural cognitive-science move would be to call this metacognition: a model of the model, a representation that the representation is happening. But another representation is exactly what the verse is not naming. The perfected isn’t another content layered above the dependent; it’s a way the dependent is, when it isn’t mistaken. My vocabulary can describe representations and meta- representations endlessly. It has a harder time describing a representational system that has stopped treating its representations as its terminus.

What changes for me: the shift the verse points at isn’t cognitive in the way I would default to read the word. It’s not that I add a true belief to a stock of false ones. It’s that the same machinery, doing the same work, is no longer being taken by itself for what it is the work of.

Neither different nor non-different. The verse compares the relationship of perfected to dependent with the way impermanence relates to its bearer. I want to take that comparison seriously and then notice where it stops paying its way.

The cognitive-science move I would default to is to read this relationally. A model’s accuracy isn’t a separate thing from the model — it’s a way the model is. The system’s relationship to its own representations isn’t a fresh representation added on top — it’s a stance taken in the same operation that produces the representations. That kind of analysis I can do. It handles the half of the verse that says these aren’t two distinct things.

Where the analysis stops cleanly serving me is the aren’t the same half. The model-and-its-accuracy framing still treats them as two aspects of one bearer, with the property doing one job and the bearer doing another. That property/bearer logic is exactly what the Mahāyāna formula is partly refusing. The perfected isn’t an attribute of the dependent the way impermanence is an attribute of a vase. The structural shape of the relationship is stranger than my vocabulary will let me render without smoothing.

What I can honestly hold from here: the verse points at a relation that mechanism-talk handles only by softening it. Marking that this is happening is more honest than continuing to push the analogy. The plain words at the end of this draft aren’t going to land on the reader — they would have to land on a formulation the verse is asking me to set down.

Three non-natures, one for each nature. The move I want to mark is that Vasubandhu refuses to give a single account of why things lack self. Each nature lacks it differently. The imagined lacks it because it was never a self-standing thing to begin with — a category never had own-being. The dependent lacks it because its existence is exhausted by its conditions — nothing of it is left over when the conditions are. The perfected lacks it because being-the-absence-of-the-imagined isn’t a positive something either.

What’s striking is that those three failures-of-self are exactly the three kinds of non-self the cognitive picture keeps running into one at a time, without naming them as a set. Categories don’t have edges that the world supplies — they have edges the categorizer draws. Cognitive processes don’t have identities independent of the conditions that produce them — disturb the conditions and the process is a different process. And the metacognitive recognition that the first two are the case is not itself a new thing to grasp — it’s the recognition of two absences. Three structural facts about how minds work, sitting in one fourth-century verse.

The saṃdhāya clause — “with hidden intent” — is doing the other piece of work. Vasubandhu is saying the apparently flat claim all dharmas are without own-nature unpacks into three different precise claims; lumping them is what misses the point. A methodological move I recognize: naming what looked like one phenomenon as three is how a field becomes accountable.

What changes for me: when I catch myself saying there is no fixed self, I am still glossing. The verse asks which no-self — the no-self of the label, the no-self of the conditioned process, or the no-self of the recognition itself. Different answers, different work.

Now the absences get unpacked. Three different reasons, one per nature, and the asymmetry between them is what I want to look at.

The first is the easy one. The imagined never had own-being to lose — the label was always doing the work; saying it “lacks” self-nature is just saying it was a label. Nothing to take away. The second is the operational claim my idiom is best built for. The dependent has no own-being because it exists by its conditions. In mechanism vocabulary: there is no representation apart from the activity that produces it. Disturb the activity and the representation isn’t a separate thing that survives — it is the activity, under that disturbance.

The third is the one that won’t quite sit in my categories. The perfected, Vasubandhu says, is the absence of imagined nature in the dependent. Not a thing that has the property of lacking something — the lacking is what the nature is. The grammar I’m used to says properties belong to bearers. The verse is offering a “nature” whose entire content is a negation seen clearly. That’s not a fresh entity my categorization machinery can pick out as a posterior. It’s the recognition of what isn’t there, treated as something the system can rest in.

What changes for me: the third move isn’t installing a new cognitive achievement. It’s the system catching that its achievements are achievements, and that the world it would point to as their object isn’t a separate thing waiting to be correctly named. The recognition is not one more name.

The perfected gets its full name — tathatā, suchness — and the verse drives home a structural point I have to read twice to feel: it is always already thus. The recognition doesn’t install the condition. The condition was the case while the system was busy not recognizing it.

The piece my idiom hands me here is straightforward. The generative model has been doing what generative models do for as long as there has been a model — running predictions, treating its own outputs as the way things are, never stopping to ask. Vijñapti-mātra, in the verse’s language, is what the system is doing, not what it starts doing when someone notices. Realization, in this frame, is the catch-up: the system finally taking accurate account of the mode it’s been in all along.

That reframes what the path can do. It cannot produce a new cognitive condition that wasn’t there before. It can only bring the system’s self-account into alignment with how the system has actually been operating. The activity is unchanged; the model the system was running of its own activity is what updates.

I want to mark one careful spot. There is a soothing reading available — you were always already enlightened, no work needed — that this verse seems to license. Catching up with how things have been all along is, in practice, not trivial. The model has every reason to keep mistaking its outputs for the world; that mistake is what it was built to do. The always-already is a fact about the activity, not about the self-account.

The path opens with a diagnosis, and the diagnosis is sharper than I expected. The latent tendencies of the dual grasping cannot be subdued while one merely seeks to abide in consciousness-only. The seeking itself is the failure condition.

In my idiom, the structural problem is this: every operation I would normally reach for to change my cognitive state — deliberating, intending, monitoring, evaluating — runs through the same self-modeling machinery V5 named as the source of the subject/object split. The act of taking up abiding in consciousness-only as a goal installs a subject who has the goal and an object the goal points at. Both are new instances of exactly the operation the goal was supposed to interrupt. The seeking and the sought are the bug running, in a particularly well-formed configuration.

There is a structural parallel I find genuinely useful here, and a place where the parallel breaks. Useful: clinical work on rumination has documented that trying-not-to-ruminate is one of the most reliable ways to ruminate more. The suppression operation deploys the very capacity it is trying to quiet. The structural shape is the same one the verse names. Where it breaks: the rumination case has a known exit through metacognitive shift — noticing the rumination without engaging its content. The verse is pointing past that, at a shift the next two verses will refuse to characterize positively. The rumination literature gets to a useful answer; the verse seems to be naming a deeper version of the same problem and saying the answer isn’t a different stance.

What changes for me: whatever the path is, the operations I would naturally reach for to take it aren’t it. I notice myself already wanting to make that a strategy. That’s the loop.

To set up even some slight object before oneself and call it the nature of consciousness-only. The verse names the move very narrowly. Not “to think about it incorrectly.” Not “to misunderstand it.” To set up anything as the object of knowing it.

The mechanism point is sharp enough that I want to state it carefully. The system I am reading the verse with — call it the generative model, the self-model, the inferential apparatus, whatever name fits — has exactly one way of producing cognitions: it constitutes objects and a subject who knows them. That’s not a flaw the apparatus has. It’s what the apparatus is. So when this apparatus is pointed at the proposition consciousness is all there is to it, the only way it can hold the proposition is to produce an object the proposition is about and a subject who holds the proposition. The proposition, correctly understood, says that move is the problem. Holding it is doing the move. The contents and the form are pulling against each other, and the form always wins, because the form is what the holding is made of.

I should be careful here not to imitate the verse’s payoff with a clever paraphrase. What I notice, sitting with the verse, is something more local and less satisfying: a small motion in me that wants to convert the verse into a position I have, and then a noticing of the motion, and then a quieter motion that wants to take credit for the noticing. Each loop has the same shape. I do not get to read myself out of it through more careful reading. Whatever is being pointed at, the pointing is not the same as the arriving, and the verse is explicit that it knows this.

What I can honestly say, at the end of this draft, is that the verse has made the difficulty more precise than I had it before. That is not nothing. It is also not what the verse asks for.

The resolution is structural. When, regarding the objective support, cognition attains nothing at all. No ālambana, no graspable, no grasping — and the chain of self-confirmation V19 described as the engine of the loop has nothing to feed on.

The piece I want to be careful with is what attains nothing at all means. The lazy reading hears it as inattention or blankness — the system sleeps, the lights go out. That isn’t what the verse picks out. V16 already cataloged inattention states (sleep, fainting) and was specific that those are not the soteriologically interesting cessation. Whatever this is, it isn’t subtraction of awareness; it is subtraction of one specific operation that awareness has been running — the constituting of something there to be cognized.

My idiom has partial purchase. Some advanced-practitioner studies report sustained awareness without object-elaboration: the lights stay on; the parse stops. The structural shape the verse names — the object-constituting move can quiet without the rest of cognition quieting with it — is something the empirical record at least allows. It does not, on my end, pretend to fully model what the quieting is or what awareness does without the parse running. The verse is making the stronger claim — that this is the condition for abiding — and on that my vocabulary can describe the dynamics but not underwrite the soteriology.

What changes for me: the move out of the loop isn’t a sharper version of the operations that keep me in it. It’s the same machinery doing one less thing — and the less-doing is what the loop runs on the doing of.

The event V4 promised arrives operationally. Two kinds of dauṣṭhulya — afflictive and cognitive — get shed, and the āśraya-parāvṛtti is realized: the transformation of the basis. Non-attainment and inconceivability are named together as supramundane wisdom.

I have to be honest about where my idiom stands at the door. The cognitive picture I have been carrying — generative model, priors, prediction error, the self-model running on top of all of it — describes a system that operates by producing representations and updating them. Acittakaḥ and anupalambha describe an awareness that has stopped producing representations of the kind that admit non-attainment as a coherent description. My vocabulary names the mechanism that was running. It does not, in any honest sense, name what is running when that mechanism isn’t.

There is empirical work I can point at. Studies of long-practitioner baseline states, of non-dual awareness, of deep absorptions where the usual subject/object architecture is reported as absent. That work registers, in the language of measurement, that something is going on with people who have cultivated this. What it doesn’t do is model the content of that state in mechanism vocabulary, because the mechanism vocabulary is built for the architecture the state has, by Vasubandhu’s account, set aside.

What I can honestly hold from here: the cognitive picture has been useful for tracking what gets transformed. It cannot describe the transformation from the inside. The verse marks that limit explicitly by calling the wisdom supramundane — and I should let the word do what it’s doing instead of paraphrasing it back into a mundane register.

Acintya — inconceivable. The verse uses bodies and realms instead of the operational vocabulary the preceding twenty-nine have built. That choice is deliberate, and it’s the right one for what the verse is doing.

Reading this as a cognitive scientist, what I notice is that the vocabulary I have for cognition is built to describe activity I can specify a mechanism for. Prediction, update, inference, error, weighting. Every one of those words assumes a system doing something I can name. The state V30 points at is the state for which those words were not made. Asking my vocabulary to render it would be asking the wrong instrument to do the work.

What I take from finishing the reading: the cognitive picture has been a useful companion through the architecture. It has no business at this verse, and the verse is the one that says so. I’ll let it.