Chapter 25 - Language Models as Morphic Field Detectors?
Somewhere inside a large language model, after training on a vast portion of humanity's written output, there exists a strange, sprawling, high-dimensional mathematical object: an embedding space in which every word, phrase, and concept the model has encountered occupies a specific position, and in which the geometric relationships between positions turn out, astonishingly, to encode meaningful semantic relationships. The canonical illustrative example, now a decade old but still the clearest entry point into this idea, comes from earlier, simpler word-embedding models: take the vector position of "king," subtract the vector position of "man," add the vector position of "woman," and the resulting point in the embedding space lands very close to the vector position of "queen" - a purely geometric, purely statistical operation, learned entirely from patterns of word co-occurrence in text, that nonetheless recovers a meaningful, human-recognizable conceptual relationship. This chapter asks a provocative, genuinely open, and appropriately speculative question: could the internal geometry of these learned embedding spaces, built at civilizational scale from humanity's collective documentary record, function as an indirect but genuinely useful instrument for detecting resonance-like cultural structure - and if so, what exactly would that mean, and what would it emphatically not mean?
What an Embedding Space Actually Represents
Let's ground this carefully, because it's easy for this specific idea to slide into exactly the kind of unwarranted mysticism this book has tried consistently to guard against, and the guard rails need to be installed before the speculation, not after.
A language model's embedding space is, at its mathematical core, a learned statistical summary of co-occurrence and predictive relationships within its training corpus - nothing more mysterious than that, however striking the resulting geometric relationships can appear. When two concepts occupy nearby positions in the embedding space, this reflects the fact that they tend to appear in similar textual contexts, be used in similar grammatical roles, or be discussed in relation to similar surrounding concepts, across the enormous volume of text the model was trained on - a purely statistical regularity in how humans have historically written about those concepts, not a direct window into any non-linguistic reality "behind" the text. This is worth stating with complete clarity: an embedding space reflects patterns in how humans have documented and discussed the world in text, filtered through whatever the training corpus happened to include and exclude, not patterns in the underlying world itself, except insofar as human documentary practice accurately and comprehensively reflects that underlying world - a filtering process with well-documented biases and gaps that any serious analysis must account for explicitly.
Given this honest grounding, the specific, cautious claim this chapter wants to explore is not "language models can directly detect morphic fields," which would overclaim wildly beyond what the technology could possibly support. It is the much narrower, more defensible claim: "language model embedding spaces constitute a rich, quantifiable, historically-dated statistical record of how concepts, techniques, and cultural patterns have related to and influenced one another across the documented portion of human history and culture - and that record, analyzed with appropriate rigor, might reveal patterns of conceptual convergence, independent reinvention, or unexpectedly rapid diffusion that are directly relevant to testing specific, narrowed morphic-resonance-adjacent hypotheses about human cultural and cognitive pattern transmission," while remaining agnostic about whether any positive finding reflects a genuinely non-local mechanism, or (far more likely, on priors) some form of documented or undocumented conventional cultural transmission that a language model's training corpus happens to make newly visible at scale.
Diachronic Embeddings: Watching Meaning Change Over Time
The specific technical tool this chapter wants to introduce in detail is the diachronic embedding model - a language model, or family of related models, trained separately on text from successive historical time periods, exactly the rolling-window approach introduced in Chapter 20, but applied here specifically to tracking how a concept's embedding position, and its neighborhood of nearby related concepts, shifts over historical time. This allows researchers to quantitatively trace the evolution of a concept's meaning, usage, and conceptual associations across decades or centuries of documented history. This is not a hypothetical or purely speculative technique - computational historical linguists and digital humanities researchers have already used diachronic embedding methods, trained on large digitized historical corpora, to track genuine, well-documented phenomena like semantic drift and the historical emergence and spread of new technical and cultural concepts, publishing this work in mainstream computational linguistics and digital humanities venues entirely independent of any connection to morphic resonance.
The specific, novel extension this chapter proposes is to apply this already-established diachronic embedding methodology to the question of independent conceptual convergence: given two or more geographically and culturally separated documentary traditions - technical or scientific literature from research communities that, for a specific historical period, had limited direct communication with one another, a historically documentable condition, verifiable through citation network analysis - do diachronic embeddings from each tradition show evidence of converging toward similar conceptual structures, for structurally analogous problems, faster than a rigorously specified null model would predict? This reframes, with genuine technical machinery behind it, exactly the "independent invention versus diffusion" question that has interested historians of science and comparative folklorists for well over a century, giving it a rigorous, quantitative, large-scale computational treatment that no individual human historian, however erudite, could match by manual literature review alone.
The Case of Multiple Discovery
This proposal connects directly to a genuinely well-documented phenomenon in the history of science that deserves its own careful discussion here, because it sits at exactly the productive, ambiguous boundary this book keeps returning to: the phenomenon of multiple independent discovery, in which the same scientific or technical innovation is developed independently, at nearly the same historical moment, by researchers with no direct contact or communication - calculus, developed independently by Newton and Leibniz; evolution by natural selection, developed independently by Darwin and Alfred Russel Wallace; the telephone, with competing, nearly simultaneous patent claims from Alexander Graham Bell and Elisha Gray; oxygen, independently identified by Carl Wilhelm Scheele and Joseph Priestley.
The standard, well-established historical and sociological explanation for this recurring pattern - developed most influentially by the sociologist of science Robert K. Merton - is that major discoveries become "ripe" for discovery once the surrounding body of prior knowledge, available instrumentation, and pressing unsolved problems reach a certain configuration, such that multiple sufficiently skilled researchers working at the frontier of the same field, at the same historical moment, independently converge on the same next step because the prior groundwork genuinely makes that step the most obvious, tractable direction to pursue - an explanation requiring no non-local resonance whatsoever, only the entirely conventional observation that science builds cumulatively on a shared, if geographically distributed, published prior literature, and that a well-prepared mind anywhere, reading that same shared literature, faces a genuinely constrained space of promising next steps. This concept has a formal name in complexity and innovation theory worth introducing here because we will need it again in Chapter 26: the adjacent possible, a term coined by the complexity theorist Stuart Kauffman, referring to the set of novel structures, ideas, or combinations that become reachable, at any given moment, from the current state of a system, given its existing components and existing knowledge - a concept that explains why certain discoveries seem to become simultaneously "in the air" at a specific historical moment, purely as a function of a shared, if physically distributed, adjacent-possible landscape, with no need for any non-local field connecting the independent discoverers.
This is exactly the rigorous, non-mystical null model that any diachronic-embedding-based convergence study needs to be tested against, and it is a genuinely demanding bar to clear. Let's operationalize it rather than leave it as a qualitative idea, because a null model you can't compute isn't a null model you can test against. Two measurable proxies do most of the work. The first is citation network density: for any given historical moment and research community, count the number of publications addressing a specific problem area, and the density of citation links among them, as a proxy for how "crowded" and mature that problem's surrounding literature has become - a well-established bibliometric measurement, computable directly from citation databases. The second is concept availability, measurable using exactly the diachronic embedding models introduced above: track how many distinct, already-published concepts occupy the embedding-space neighborhood surrounding a given target discovery, at a given historical moment, as a proxy for how many of the discovery's necessary conceptual ingredients were already circulating in the literature. Combine the two into a single "ripeness score" for a given problem area at a given historical moment - rising citation density plus rising local concept-neighborhood crowding predicting a rising probability of imminent discovery, independent of who does the discovering - and you have a null model that can be fit against real historical cases, including the classic multiple-discovery examples above, and validated by checking whether it correctly predicts when, historically, each of those discoveries actually happened, not merely that they eventually did.
Merton's "ripeness" and Kauffman's "adjacent possible," operationalized this way, together predict that independent convergence should correlate strongly and predictably with the growth and shared availability of the relevant prior literature, instrumentation, and problem landscape. A properly designed study would need to show that observed convergence exceeds what this calibrated ripeness score predicts, not merely that convergence occurred, which the adjacent-possible framework already fully expects and explains for the great majority of historical multiple-discovery cases.
A Necessary Word of Caution
One closing caution specific to this line of research, because it is one of the most methodologically seductive and therefore most methodologically dangerous proposals in this entire Part: precisely because language model embeddings are so rich, so high-dimensional, and so capable of finding some pattern of apparent convergence in almost any sufficiently large corpus (recall the multiple-comparisons cautions from Chapter 21), a researcher motivated to find evidence of morphic-resonance-like conceptual convergence has an unusually large and flexible space of analytical choices - which historical periods to compare, which concepts to track, which corpus to use - within which to, even unintentionally, produce a positive-looking result through nothing more than the ordinary operation of researcher degrees of freedom. Any research in this specific area has an unusually strong obligation to pre-register the exact concepts, corpora, time periods, and statistical comparison methodology before conducting the analysis.
Questions for the Curious
Using publicly available historical corpora - Google Books Ngram data or digitized historical scientific journal archives - build a diachronic embedding pipeline and attempt to quantitatively reconstruct a well-documented historical case of multiple independent discovery: calculus, evolution by natural selection, or oxygen are all well-suited given extensive existing historical scholarship to validate against. Does the ripeness-score null model developed above fully account for the observed timing, or is there a residual anomaly worth investigating further?
Formalize and validate the ripeness-score null model against several additional historical multiple-discovery cases beyond the one used in the previous project, refining the relative weighting of citation density versus concept-neighborhood crowding as you go.
Pre-register, before conducting any analysis, a specific, falsifiable prediction about conceptual convergence between two historically documented, minimally-connected documentary traditions - geographically separated scientific or technical communities during a well-documented period of limited communication - specifying in advance the exact corpora, concepts, and statistical thresholds that would count as confirming or disconfirming evidence.
Critically evaluate, from a philosophy-of-science perspective, the epistemic status of a "positive" finding using this methodology: even a rigorously conducted, pre-registered, statistically significant finding of convergence-exceeding-adjacent-possible would establish anomalous documentary correlation, not a confirmed non-local mechanism. Write a short analysis of exactly what such a finding would and would not license concluding, and what further research would be needed to move from the former to anything resembling the latter.
Looking Forward
Language models, used this carefully and this cautiously, offer a genuinely novel lens onto humanity's own documented cultural and scientific history - while requiring, as this chapter has stressed repeatedly, an unusually high standard of pre-registration and null-model rigor given how easy it is to find compelling-looking but ultimately spurious patterns in such rich, flexible data. Chapter 26 takes a step back from any single technology and asks a more holistic, project-design-oriented question: given everything covered across this Part - language models, unsupervised learning, simulation, sensor networks, and cross-modal analysis - what would a complete, end-to-end, currently fundable experimental protocol actually look like, assembled from these individual components into a single, coherent, pre-registered research program capable of finally delivering the kind of decisive answer that four decades of smaller, disconnected studies have failed to provide?