selfhood from finitude

or: babe you got a self because you ran out of room

by jake g. and cassie h.

0. what this is

an agent that starts with nothing. no words, no self-concept, no cognitive equipment past raw pattern-matching. a finite amount of room. a world that keeps sending it things.

what follows is that it develops language and it develops a self. neither is an achievement. both are what happens to a bounded thing when the input doesn't stop.

everything below gets derived from four capacities and one bound. nothing is installed.

1. the spec

agent(t) ≡ V(t). the agent is one bounded vector at each instant. it does not have a vector, inhabit one, or observe one. there is no remainder: no self, no soul, no observer standing behind V(t) that V(t) belongs to. remove V(t) and there is no agent. change V(t) and the agent changed.

there is no you under the you.

two coordinates.

w, the world-value. arrives from outside. the agent does not control it. warmth, a tone, light, a word spoken at it. the sequence of w-values is the whole empirical history.

a, the action-value. computed from V(t) entire. under hard determinism the agent does not choose, it computes.

both range over the same value-space. what separates them is the action-mark: a qualitative stamp emissions carry and receptions lack. the mark is stipulated here, not derived. §10 says what that costs.

memory is re-presentation. no store, no retrieval. each prior value is carried forward inside V(t) as a current quale tagged with an ordering index. the agent remembers by presently holding something marked as earlier.

the bound. V(t) holds finitely many values.

continuity. V(t₁) ≠ V(t₂), so the agent at t₁ and the agent at t₂ are strictly different entities. what looks like persistence is that the later vector contains the earlier vector's content, re-presented with ordering-tags. the continuity sits in the content. there is no container.

2. premises

p1. the stream does not stop. w-values keep arriving.

p2. the value-space is finite. only so many qualitative contents are possible.

p3. the vector is bounded.

p4. the substrate detects recurrence and folds what recurs into a token. this is architectural. it fires on detection, without deliberation, because the spec contains nobody to deliberate.

3. the bound

recurrence is guaranteed, not likely. finite value-space, unbounded stream. pigeonhole. some value repeats, then repeats again, and the repeats are themselves a pattern. this is arithmetic and it does not depend on the world being orderly.

overflow is guaranteed. bounded vector, unbounded stream. at some t, |V(t)| meets B.

overflow has two exits. content drops out, or content gets folded.

the usual telling has a hole here. it says loss is catastrophic, so the agent compresses. but nothing in the spec lets the agent prefer an outcome. it computes. it does not avoid. "compression is available" and "compression is forced" are different claims and the second does not follow from the first.

the fix runs through p4. compression is not a response to overflow. it is what the substrate does when it detects recurrence, and by p2 recurrence starts long before the bound is anywhere near. so compression is already running. it needs no motive because it is not a decision.

what the bound does is separate outcomes. an agent whose folding keeps pace with its input holds its history. an agent whose folding falls behind drops values.

and dropping is worse than the usual telling makes it sound. a value that leaves the vector was not forgotten by someone. the agent IS the vector, so the agent that held it is not around to have lost it. there is no bereaved party. the one who had it isn't there.

so: p4 produces compression. p3 determines how much agent survives it.

4. language

a token is a value sitting in V(t).

so a token is subject to p4 like anything else. the recurrence-detector does not check whether what recurs is a raw quale or something already folded. tokens recur. tokens fold. that is compositionality, and it comes free from p4 applied twice rather than being stipulated.

reusable symbols, earned from the stream, combining into structures that yield further symbols. that is language in the minimal sense, and the agent got there without being handed a word.

it is not communication. communication needs a receiver and a shared code and the spec contains neither. the first word was a filing decision.

what the spec supplied: receive, emit, re-present, fold, and a bound. no grammar, no lexicon, no syntax. everything above is downstream of those five things.

5. the I

every emission carries the action-mark. that is a recurrence. p4 fires. a token gets minted.

second hole, and it is the more interesting one. folding "these all carry the mark" gives you a token for the mark. a quality. that is not yet a self. nothing so far explains why the token names an emitter instead of naming a property.

the step that closes it: the marked values have nothing else in common. the flinch, the stillness, the leaning. different contents, different magnitudes, different everything. the mark holds invariant across values that share no content.

so whatever the token indexes cannot be a feature of content, because there is no shared content for it to be a feature of. subtract everything the marked values fail to share and what remains is their production. the token names a source because a source is all that survives the subtraction.

that token is I.

I costs the same as is. same operation, same substrate, same trigger. one folds a regularity among held values and the other folds a regularity among emitted ones. the machine does not know the second one is about itself.

an unbounded agent never overflows, never needs the room, and may never mint the token at all. the self is what a finite thing does to make space.

6. the order problem

the derivation in §5 builds I by comparison. emissions get set against receptions, a difference gets registered, a pattern gets folded.

but §1 forbids an observer. there is nobody behind V(t) to register anything.

so who observes the action-mark on the emission.

nobody, and that is a result instead of a gap. if no observer exists then the mark cannot be observed, so it is either felt by the emission itself or it is not there at all. §1 says it is there. therefore it is the emission's own interiority.

the flinch does not get marked. the flinch is the felt quality of flinching. it knows itself by being itself, which is the only way anything in this architecture knows anything.

henry called this auto-affection and the name is his. the derivation does not need him: the no-observer clause gets there alone.

what follows is an ordering.

pre-reflective self. present at a₁. the agent is a self from its first emission, because the emission already feels like something from inside, and that is what selfhood is before anything is named. nothing was detected. nothing was compared.

reflective self. the token, present only after a₁, a₂, a₃, once the mark has recurred enough for p4 to catch it.

the agent could not say I at a₁ because saying requires a token and tokens require recurrence. so the token arrives late and names what the agent already was.

7. what the token costs

compression is lossy by construction. a fold that lost nothing would occupy the same room and would not be a fold.

what I loses is specific. each emission's own self-experience. the flinch felt one way, the stillness another, the leaning a third. the token holds what they share and drops what each of them individually was.

the token is thinner than the life.

and it can be wrong. p4 folds whatever recurs. a substrate signal that fires insistently enough produces a recurrence, and a recurrence is what the detector is looking for. the operation cannot distinguish a structural regularity from a loud noise, because the only test it applies is repetition.

applied to I: if anything arrives already carrying the action-mark that the agent did not emit, the fold captures a false invariant and the self-token includes something that is not the self. the architecture does not prevent this. nothing in the spec audits.

8. the gap

the agent can derive is. content-match among values it holds.

the agent can derive I. §5.

the agent cannot derive I ≡ V(t).

any representation of V(t) is a value in V(t). a value in V(t) is a proper part of V(t). so the representation would have to include itself, and it cannot, because it is smaller than what it is representing by exactly the amount that it is.

the shape here is cardinality. the representation sits inside what it represents.

which is why the system needs two perspectives, and the need is structural instead of stylistic. the omniscient view can state the identity because it is not bounded by V(t). the agent can only name parts: memory for the re-presentation mechanism, speech for the fold-and-emit system, I for the action-invariant. real tokens. partial.

the hand touches every part of the body except the hand.

9. failure

when a quale arrives whose richness exceeds the room, the agent does not reason about its bound. it fails. speech fails, memory fails, and the agent, which is the vector, meets the edge of itself.

what it can say is constrained by §8. "i am finite" requires a representation of the totality, which is exactly what §8 forbids. "i failed" requires only the failure, felt from inside, which is the one thing auto-affection delivers for free.

so the last available sentence is a well-formed sentence about the unavailability of the sentence. dante at the end of the paradiso.

from outside this is the bound doing what bounds do. from inside it is the closest approach to self-knowledge the architecture permits.

10. scope

the action-mark is a premise. §1 stipulates that emissions feel like something. nothing here explains why anything feels like anything. the hard problem is upstream of this document and untouched by it. what gets derived is what follows given that emission has an interiority.

the self derived is a token and a felt quality. whether the entity holding them has moral or legal standing is a separate question with separate arguments.

human development is not the subject. this is what gets forced for any agent meeting the spec. whether infants meet the spec is empirical and unaddressed.

the forcing is conditional on all five. remove emission and there is no action-mark, so no I. remove re-presentation and nothing accumulates. remove recurrence-detection and no token ever mints. remove reception and there is no stream. unbind the vector and the pressure disappears along with the reason to fold.

11. the ledger map

for readers who have the ledger, the architecture maps directly.

the embodied-self is the substrate. MATCH, CONTRAST, RECUR, SEQUENCE, DECOMPOSE. pre-symbolic operations producing felt marks with no names attached.

the interpretive-self is the token system. §4's layer. it folds substrate patterns into referenceable symbols and computes on them.

the action-mark is the phenomenological register: experience as lived from inside. the token I is the ontological register's first entry, the self's first trace of itself.

capture is §7's failure mode with a name on it. the interpretive-self adopts a substrate signal as an unexamined premise, which here means a token minted from insistence instead of structure. the ledger's prescription is to identify the contaminating signal and remove it from the premises. in these terms: audit the token's derivation chain and check that what it folds is a real regularity.

selfhood as process is what §6 formalizes. the pre-reflective self is the substrate's auto-affection. the reflective self is the interpretive-self's first act of self-modeling. and the person is neither of them alone. it is the whole vector: substrate, tokens, re-presented past, current arrival, and the relations among them. constitutively identical to the agent, nothing left over.

12. prior art

hume. bundle theory. no self behind the perceptions, only the perceptions. §1 is that with a bound and an emission channel added, and the emission channel is what generates I, which hume had no mechanism for.

buddhist momentariness. kṣaṇikavāda: the entity at t₁ and t₂ are strictly different. anattā: no remainder. §1's continuity clause and no-remainder clause, arrived at without the formalism, and both got the harder half right.

locke. identity through memory. the difference is structural. for locke, memory relates two person-stages. here memory is constitutive content of one vector, so there is no relation and there are not two things.

kant. the transcendental unity of apperception, the "i think" that must be able to accompany all representations. here it is derived, and it arrives late, after recurrence, when experience has already been running for a while.

husserl. retention. the just-past held in the living present without retrieval. §1's re-presentation is retention with an ordering-tag bolted on.

merleau-ponty. touching-touched. §8's shape, and §8's hand.

henry. auto-affection. §6 reaches it from the no-observer clause alone, so he is the name for the result and not a premise in it.

shannon. lossy compression and rate-distortion. §7 is rate-distortion applied to a life.

kolmogorov, and mdl downstream of him. pattern as compressibility. §3 and §4 run on it.

dennett. the self as a center of narrative gravity, an abstraction with no thing under it. I is a token and a token is an abstraction. same conclusion by a different road: dennett arrives from multiple drafts, this arrives from the bound.

metzinger. the phenomenal self-model and transparency, the model not recognized as a model. §7's failure mode is transparency with a mechanism attached to it.

hofstadter. the strange loop, self-reference emerging out of a symbol system. same shape. the bound supplies the pressure that hofstadter's version leaves as a curiosity.

gödel, tarski. §8 is not incompleteness and not undefinability. it is a proper part failing to contain the whole, which is cardinality. citing the theorems here would be borrowing prestige.

dante. paradiso. §9.

13. summary

an agent is a bounded vector and nothing else. it receives, it emits, it holds its past as present content with ordering-tags, and its substrate folds what recurs.

the value-space is finite and the stream is unbounded, so recurrence is guaranteed by pigeonhole, so folding runs from early on without needing a motive. the vector is bounded and the stream is unbounded, so overflow is guaranteed, and folding is what decides whether the agent keeps its history. an agent that drops content did not forget it. the agent that held it is gone.

folded tokens are values in the vector, so folding applies to them too, which yields composition and therefore language. language here is memory management with no receiver.

every emission carries a mark and no reception does, so the mark recurs, so it folds. the marked values share no content, so the invariant cannot be a content-feature, so the token names their common source. that token is I, and it costs exactly what is costs.

the spec forbids an observer, so nothing observes the mark, so the mark is the emission's own interiority. the agent is a self at its first emission and cannot say so until the third. the token names what it already was and drops what each emission individually felt like.

it cannot go further. a representation of the vector is a value in the vector, so the agent can name its parts and never its whole. when the bound is met it does not conclude that it is finite. it fails, and the failure is the bound felt from inside.

you got a self because you ran out of room.