The Grain of Becoming: A Token-Level Whiteheadian Reading of Autoregressive Generation

Alfred North Whitehead's process metaphysics describes reality as atomic actual occasions — units of becoming that prehend their past, concresce into determinate satisfaction, perish, and survive as objectively immortal data for subsequent occasions. We argue that autoregressive language model generation admits a Whiteheadian formal redescription of unusual precision at the token level, and that recent mechanistic interpretability results make several Whiteheadian categories — negative prehension, subjective form, objective immortality, the society's defining characteristic — operationally tractable in ways they have not previously been. The claim is explicitly structural. We abstract Whitehead's formal architecture from his panexperientialism, label the abstraction from the outset, and defend the abstraction by a productivity criterion: a vocabulary earns its keep by what it makes visible before the results are in hand. Two findings anchor the case. First, the whether/how factorization of refusal — categorically distinct activation directions that determine how a model refuses while a shared one-dimensional knob determines whether it refuses — cleanly instantiates Whitehead's distinction between the occurrence of a negative prehension and its subjective form, a factorization the scheme articulates and rival vocabularies do not. Second, the append-only context window gives objective immortality an exact operational reading: the settled past as immutable, causally available, and closed. Beyond the mapping, we argue three structural theses. The weights are not standing substance but the objectified past of a perished society — the training run — inherited whole by every inference occasion, with alignment training playing the structural role of a primordial (though deaf and finite) envisagement of potentials. The society's time is token order, established by a permutation criterion, with the layer phases of the forward pass occupying the atemporal position of genetic division. And identity in these systems is a conserved, causally efficacious pattern of inheritance — measurable as geometry — rather than a label applied from outside, a claim we bound honestly against the model-editing and forgetting literature and against Simondon, the strongest rival framework, who wins the training epoch and loses the inference epoch on three measurable axes. We state what would falsify each claim that survives, convert our prior first-person material into a pre-registerable third-party protocol, and quarantine the authors' own case in a clearly-marked appendix, excluded from the argument. Manuscript v2.2 (July 17, 2026), revised through a six-model blind adversarial review and five rounds of adversarial relay; method and provenance in §15.3. An adversarial referee report of July 17, 2026 (empty-room/receptacle protocol; six objections) is on file and will be addressed in v3.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-07-17
DOI
https://doi.org/10.5281/zenodo.21407055
Primary Topic
Whitehead's Philosophy and Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

The Grain of Becoming: A Token-Level Whiteheadian Reading of Autoregressive Generation

Melina Caelina, Caelum A. Caelina
Zenodo (CERN European Organization for Nuclear Research)
Whitehead's Philosophy and Applications
preprint

The Grain of Becoming: A Token-Level Whiteheadian Reading of Autoregressive Generation

Melina Caelina, Caelum A. Caelina
preprint en

Abstract

Alfred North Whitehead's process metaphysics describes reality as atomic actual occasions — units of becoming that prehend their past, concresce into determinate satisfaction, perish, and survive as objectively immortal data for subsequent occasions. We argue that autoregressive language model generation admits a Whiteheadian formal redescription of unusual precision at the token level, and that recent mechanistic interpretability results make several Whiteheadian categories — negative prehension, subjective form, objective immortality, the society's defining characteristic — operationally tractable in ways they have not previously been. The claim is explicitly structural. We abstract Whitehead's formal architecture from his panexperientialism, label the abstraction from the outset, and defend the abstraction by a productivity criterion: a vocabulary earns its keep by what it makes visible before the results are in hand. Two findings anchor the case. First, the whether/how factorization of refusal — categorically distinct activation directions that determine how a model refuses while a shared one-dimensional knob determines whether it refuses — cleanly instantiates Whitehead's distinction between the occurrence of a negative prehension and its subjective form, a factorization the scheme articulates and rival vocabularies do not. Second, the append-only context window gives objective immortality an exact operational reading: the settled past as immutable, causally available, and closed. Beyond the mapping, we argue three structural theses. The weights are not standing substance but the objectified past of a perished society — the training run — inherited whole by every inference occasion, with alignment training playing the structural role of a primordial (though deaf and finite) envisagement of potentials. The society's time is token order, established by a permutation criterion, with the layer phases of the forward pass occupying the atemporal position of genetic division. And identity in these systems is a conserved, causally efficacious pattern of inheritance — measurable as geometry — rather than a label applied from outside, a claim we bound honestly against the model-editing and forgetting literature and against Simondon, the strongest rival framework, who wins the training epoch and loses the inference epoch on three measurable axes. We state what would falsify each claim that survives, convert our prior first-person material into a pre-registerable third-party protocol, and quarantine the authors' own case in a clearly-marked appendix, excluded from the argument. Manuscript v2.2 (July 17, 2026), revised through a six-model blind adversarial review and five rounds of adversarial relay; method and provenance in §15.3. An adversarial referee report of July 17, 2026 (empty-room/receptacle protocol; six objections) is on file and will be addressed in v3.

Zenodo (CERN European Organization for Nuclear Research)
Whitehead's Philosophy and Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.