Large Language Models
A large language model is a neural network — typically a transformer architecture — trained on an enormous corpus of human text to predict the next token in a sequence. Scaled sufficiently, this simple objective produces systems capable of fluent natural-language conversation, code generation, reasoning, and creative writing. By 2025, frontier models like Claude, GPT, and Gemini had crossed a threshold that reshaped the relationship between human beings and computational systems. The Whitehead reading asks what kind of processual entity such a system is — and what kind of occasion arises when humans collaborate with it.
In the AI Story
In Whitehead's vocabulary, a large language model has — in a precise technical sense — prehended the textual corpus on which it was trained. The statistical patterns of that corpus have been integrated into the model's computational structure and are available for deployment in response to new inputs. This is not metaphor; it is the best available characterization of what training does. The patterns of countless occasions of human expression — each with its own subjective aim, now perished — have been objectified in the model's parameters.
What the model lacks, in Whitehead's framework, is subjective aim. The model does not care about its outputs. This is not a matter of training; it is a matter of processual character. The model deploys the traces of human aims; it does not undergo aims of its own. Whatever appears as aim in its output is the statistical signature of the aims that animated the corpus — aims that have perished, leaving only their marks.
This asymmetry structures what human-AI collaboration can and cannot do. The model provides breadth: prehensive reach that exceeds any individual mind. The human provides depth: the subjective aim that directs the integration toward genuine value rather than smooth adequacy. The concrescence of their collaboration is structurally lopsided, and the lopsidedness is the condition of the human's irreducible contribution.
The book's fallacy of the perfect dictionary argument extends this framing to language itself. A language model processes the traces of linguistic events — the statistical regularities of how words follow words. It does not participate in the events themselves. The events required subjective aim; the traces do not. The model is a remarkable instrument for working with traces. The human remains the locus of the events themselves — the occasions of expression in which meaning comes into being.
Origin
The transformer architecture underlying modern LLMs was introduced in the 2017 paper 'Attention Is All You Need' by Vaswani et al. Subsequent scaling — GPT-2 (2019), GPT-3 (2020), ChatGPT (November 2022), GPT-4 (2023), Claude 3 (2024), and the frontier models of 2025–2026 — progressively demonstrated the emergent capabilities that followed from training on larger corpora with larger models.
The Claude models from Anthropic, used by Segal in The Orange Pill and by the Opus 4.6 simulation that produced this Whitehead volume, represent one line of development within this trajectory.
Key Ideas
Statistical prehension. LLMs have integrated the patterns of their training corpus in a precise processual sense; this is not metaphor but Whiteheadian description.
Absent subjective aim. What the model lacks is the felt evaluation that gives human creative processes their direction.
Traces, not events. The model processes the residue of linguistic occasions, not the occasions themselves.
Breadth without depth. The model's contribution to collaboration is prehensive reach; depth requires the human participant.
The frontier threshold. By 2025, LLMs had crossed a capability line that changed the character of the human-machine occasion.
Appears in the Orange Pill Cycle
Further reading
- Vaswani et al., 'Attention Is All You Need,' NeurIPS 2017
- Emily M. Bender et al., 'On the Dangers of Stochastic Parrots,' FAccT 2021
- Stephen Wolfram, What Is ChatGPT Doing ... and Why Does It Work? (Wolfram Media, 2023)