% ============================================================================
% Part II — The Evolution of Computer Science
% Written by Agent L (Academic Writer), WP44, 2026-11-03.
% Bibliography: references_part02.bib (every key verified; see references_verification_part02.csv).
% Corpus quantities only through the macros of tables/numbers.tex; \pending{} marks missing macros.
% ============================================================================

\section{The disciplinary structure of computer science}
\label{sec:p2-structure}

Any claim that computer science has been transformed since late 2022 presupposes a map of the discipline on which the transformation can be located. That map is not given by nature. Computer science is organised simultaneously by its objects (processors, programs, data, networks, users), by its methods (proof, construction, measurement, learning), and by the communities and venues through which results circulate; the three organisations overlap only partially, and every bibliometric classification is a compromise among them. This section describes the compromise adopted in the present study and the alternatives against which it must be read. It serves two purposes for the remainder of the manuscript: it fixes the vocabulary of \emph{areas} and \emph{categories} used in Parts~V--IX, and it makes explicit the sense in which a ``shift of research toward artificial intelligence'' is a measurable proposition rather than a slogan.

\subsection{Classification schemes and what they encode}
\label{sec:p2-schemes}

Four classification systems are relevant to a corpus of the kind assembled here, and they encode the structure of the discipline in different ways.

The \emph{ACM Computing Classification System} (CCS) is the only one produced by a professional society as an explicit ontology. Its 2012 revision replaced the 1998 four-level scheme with a poly-hierarchy of roughly two thousand concepts organised under thirteen top-level nodes, among them ``Theory of computation'', ``Computing methodologies'' (which houses artificial intelligence, machine learning, computer graphics and simulation side by side), ``Computer systems organization'', ``Software and its engineering'', ``Security and privacy'', ``Human-centered computing'' and ``Networks'' \citep{rous2012ccs}. The CCS is a statement of how the ACM understood the field in 2012. It is applied by authors at submission time, is used almost exclusively in ACM venues, and is therefore sparse in any corpus that spans IEEE, Springer, Elsevier, open-access publishers and preprint servers. Its value for this study is conceptual rather than operational: it provides an independent, expert-designed reference against which the forty categories of CLAUDE.md~\S4 can be positioned, and it documents a pre-transformer view in which large language models, foundation models, AI agents and AI safety do not exist as concepts at all.

The \emph{arXiv subject classes} are the second system. arXiv was created in 1991 for high-energy physics and its taxonomy reflects that origin; the computer science archive (\texttt{cs}) is organised into forty subject classes largely inherited from the ACM scheme of the 1990s \citep{ginsparg2011arxiv}. Authors choose a primary class and may add cross-lists, so the classification is self-reported, multi-label, and coarse. Its distinctive property is temporal: a preprint is classified at the moment of submission, before peer review, so the arXiv class records how authors positioned their work at the time, not how a later indexer reconstructed it. For the questions of this paper that property matters: when a paper on, say, program verification with language models is submitted to \texttt{cs.PL} with a cross-list to \texttt{cs.LG}, the pair of classes is a small but genuine piece of evidence about disciplinary boundaries as practitioners perceive them.

The \emph{DBLP computer science bibliography} is the third. DBLP does not classify papers at all; it classifies \emph{venues}. Begun in 1993 as a database and logic programming bibliography and extended to the whole of computer science, it records the tables of contents of conferences and journals with curated author identities \citep{ley2002dblp}. Membership in DBLP is itself a judgement that a venue belongs to computer science, and the venue stream to which a paper belongs is a strong signal of community: a paper in the proceedings of OSDI is an operating-systems paper in the sociological sense whatever its subject matter. DBLP therefore encodes the discipline as a set of communities rather than as a set of topics, and the two encodings disagree precisely where this paper expects change, namely where machine-learning methods enter venues that were not built for them.

The fourth system is the family of \emph{algorithmic, citation-based classifications} that underlie modern open bibliographic indexes. OpenAlex, which succeeded the Microsoft Academic Graph after the latter was discontinued at the end of 2021 \citep{wang2020mag,priem2022openalex}, assigns to every work a ranked list of \emph{topics} from a fixed vocabulary of about 4{,}500 topics grouped into subfields, fields and domains; the topics were obtained by clustering the citation network and labelling the clusters with a language model, and are assigned to new works by a classifier over title, abstract and venue \citep{openalex2025topics}. The approach descends from the publication-level clustering methodology of \citet{waltman2012classification}, refined by the Leiden community-detection algorithm \citep{traag2019leiden}, and it inherits the known properties of that family: clusters are well connected and reproducible, but their granularity is a free parameter with no natural value \citep{sjogarde2018granularity}, and citation-based taxonomies, while more accurate than text-based or journal-based ones on average \citep{klavans2017taxonomy}, lag behind the literature by construction because a cluster must accumulate citations before it can exist. The Semantic Scholar academic graph offers a comparable but differently constructed resource \citep{kinney2023s2}. Older, journal-based classifications of the kind proposed by \citet{glanzel2003classification} remain in use in evaluative scientometrics but are too coarse, and too journal-centric, for a conference-dominated discipline.

Each of these systems is a model of the discipline with its own inductive bias. The CCS encodes expert judgement frozen in 2012; arXiv classes encode author self-positioning at submission; DBLP streams encode venue communities; OpenAlex topics encode citation neighbourhoods as of a recent snapshot. None of them is ground truth, and this study treats their disagreements as data rather than as noise.

\subsection{The forty-category ontology of this study}
\label{sec:p2-ontology}

The classification used throughout the manuscript is the hierarchical ontology specified in CLAUDE.md~\S4 and built in Phase~2 (Appendix~A). It comprises forty leaf categories grouped into eight areas: artificial intelligence and machine learning (nine leaves, from large language models to AI safety); perception, language, robotics and interaction (six); software, programming languages and formal methods (four); theory of computation (two); data, information and distributed systems (six); security, privacy and trust (five); hardware, architecture and emerging computing (five); and computational science and discovery (three). A small number of \emph{specialisation} edges record that one leaf refines another: large language models and multimodal AI refine foundation models; compilers refine programming languages; cloud and edge computing refine distributed computing; federated learning refines privacy-preserving computation; AI accelerators refine computer architecture. The areas are a reporting device; classification operates on leaves.

The assignment rule is deliberately simple so that it can be audited. Let $\mathcal{W}$ be the set of works and $\mathcal{C}$ the set of forty leaves. For each leaf $c$ the ontology lists weighted signals of five types $k \in \{\text{topic},\ \text{subfield},\ \text{arXiv},\ \text{stream},\ \text{keyword}\}$, with signal set $S_k(c)$ and weights $\omega_{kj} \in (0,1]$ for $j \in S_k(c)$. A work $w$ matches a signal $j$ if the corresponding attribute of $w$ (an OpenAlex topic identifier, a subfield, an arXiv class, a DBLP stream, or a regular expression over title and abstract) contains it. The score of $w$ for $c$ is
\begin{equation}
s(w,c) \;=\; \sum_{k} \min\Bigl\{1,\ \sum_{j \in S_k(c)} \omega_{kj}\, \mathbf{1}[w \text{ matches } j]\Bigr\} \;-\; \infty \cdot \mathbf{1}[w \text{ matches an exclusion of } c],
\label{eq:p2-score}
\end{equation}
so that no single signal type can contribute more than one unit, a weight of one means ``this signal alone suffices'', and an exclusion pattern vetoes the leaf outright. The work belongs to every leaf with $s(w,c) \geq 1$; its \emph{primary} leaf is $\arg\max_c s(w,c)$, with ties broken in favour of a child over its parent and then by category number. Multi-label membership is not an accident of the rule but its purpose: an instruction-tuning paper is expected to score in both large language models and natural language processing, and the share of works with more than one leaf is itself one of the quantities tracked in Part~IX.

Two design choices deserve comment because they shape every downstream result. First, the ontology is not a relabelling of OpenAlex topics. Six AI categories that the citation-based topics cannot separate (large language models, foundation models, multimodal AI, generative AI, AI agents, AI safety) and seven small categories with no dedicated topic or venue (federated learning, sustainable computing, autonomous scientific discovery, privacy-preserving computation, blockchain, quantum computing, edge computing) rely on keyword patterns of weight one; the build fails if such a pattern is absent. The reason is historical and is the subject of Section~\ref{sec:p2-waves}: the citation clusters that OpenAlex labels were formed by a literature in which ``large language model'' was not yet a stable term, and a topic such as ``Topic Modeling'' in fact hosts most BERT-era and retrieval-augmented work. Second, the ontology is positioned relative to the CCS rather than derived from it. The AI area corresponds to the CCS branch \emph{Computing methodologies $\to$ Artificial intelligence / Machine learning}, but the leaves inside it are drawn from the post-2017 literature; theory, systems, networks, security and human-centred computing map onto the corresponding CCS top-level nodes almost one-to-one; computational biology, scientific computing and sustainable computing sit in CCS under \emph{Applied computing} and \emph{Social and professional topics}, which the present ontology promotes to first-class areas because the questions of this paper concern precisely the traffic between computing and the sciences it serves.

The ontology's limitations are stated in Appendix~A and are not repeated here, except for one that bears on interpretation. A category defined partly by keywords is defined partly by vocabulary, and vocabulary changes faster than research practice. If the phrase ``large language model'' spreads into papers whose methods are unchanged, a keyword-driven category will register growth that is partly nominal. Section~\ref{sec:p2-claims} returns to this problem under the heading of LLM-assisted writing; Part~IV describes the embedding-based checks designed to detect it.

\subsection{Where the classifications disagree}
\label{sec:p2-disagree}

The disagreements among the four systems are systematic, and they fall into three kinds.

The first is \emph{coverage}. Comparisons of bibliographic databases consistently show that no source indexes the whole of a discipline and that the gaps are not random: conference proceedings, preprints, and non-English venues are the most unevenly covered \citep{visser2021comparison,martinmartin2021coverage}. OpenAlex, the backbone of this corpus, inherits the broad but uneven coverage of Microsoft Academic and Crossref; its reference lists are less complete than those of the commercial indexes for part of the period, which affects citation-based measures more than counts \citep{culbert2025openalex}. The linking phase of this study illustrates the scale of the problem at the level of record identity. The OpenAlex snapshot contains \LinkOaCsRecords{} works classified in computer science between 2018 and 2026; DBLP contributes \LinkDblpTwoZeroOneEightTwoZeroTwoSix{} records over the same years, of which \LinkDblpLinkedDistinct{} could be matched to an OpenAlex work by DOI or by title and year in the first linking pass and \LinkDblpUnlinked{} could not; the arXiv \texttt{cs} harvest contributes \LinkArxivCsRecords{} preprints, of which \LinkArxivLinkedDistinct{} were matched, overwhelmingly through the arXiv identifier that OpenAlex stores. These are linkage statistics, not findings about research, but they establish that a classification applied to one source cannot simply be transferred to another: a non-trivial fraction of the venue-defined (DBLP) universe has no citation-defined (OpenAlex) counterpart, and vice versa. Part~III describes the second linking pass and the treatment of unmatched records.

The second kind of disagreement is \emph{granularity and boundary}. The arXiv class \texttt{cs.LG} is one node among forty; in OpenAlex, machine learning spreads across dozens of topics in several fields, including statistics, engineering and medicine, because the citation network does not respect the boundary of the computer-science field. Conversely, OpenAlex's field~17 (``Computer Science'') contains clusters on educational technology, media studies and physics that no computer scientist would claim, and the ontology of this study deliberately leaves most of them unmapped (Appendix~A). The practical consequence is that the apparent size of ``AI'' depends on which boundary is used: a venue-based definition (DBLP streams of NeurIPS, ICML, ICLR, AAAI and the vision and language conferences) gives a different set from a topic-based definition, which gives a different set again from an arXiv-class definition, and the three sets diverge most in the applied sciences, where machine-learning methods are used without the paper belonging to any machine-learning community. The multi-label design of the ontology, and the use of all three signal types with capped weights in equation~\eqref{eq:p2-score}, is the response to this divergence: a work that is an AI paper by every criterion scores high on all of them, while a work that is an AI paper by vocabulary alone is held to a higher evidentiary standard.

The third kind is \emph{temporal}. The CCS was fixed in 2012 and the arXiv classes are older still; OpenAlex topics were generated from a citation network observed in the early 2020s and are reassigned when the snapshot is rebuilt; DBLP streams are stable but new venues appear and old ones are renamed. A classification that does not change cannot register the emergence of a field except as the growth of an existing node; a classification that is regenerated cannot be compared across snapshots without care. Neither property is a defect, but they pull in opposite directions, and a study of structural change must choose. This study freezes all classifications at a single snapshot (Part~III), so that any growth it reports is growth of a fixed category, and it treats the emergence of \emph{new} categories as a separate detection problem (Part~IV, Section~\ref{sec:p2-claims}).

A final, sociological disagreement should be named explicitly because it affects what counts as a ``paper''. Computer science is unusual among the sciences in treating selective conferences as the primary venue of record, a practice that has been defended on grounds of speed and criticised on grounds of reviewing quality and the fragmentation of the literature \citep{vardi2009conferences,fortnow2009growup}. The convention has two consequences for bibliometrics. Conference papers are indexed less uniformly than journal articles, so coverage disagreements are concentrated in exactly the venues that the discipline values most; and the conference calendar imposes a strong seasonality on publication dates, so that monthly series must be deseasonalised before any break can be read off them. The preprint culture adds a third consequence. Computer scientists came to arXiv late relative to physicists and mathematicians, and a large share of computer-science preprints are never matched to a published version, either because they are not published or because the match fails \citep{lin2020preprints}. A corpus that counts preprints and published versions as separate works will over-count the preprint-heavy subfields; a corpus that links them will under-count whenever linkage fails; and the subfields where linkage fails most often are, again, the fast-moving ones.

\subsection{Why the growth of a field is hard to measure}
\label{sec:p2-growth}

The questions of this paper are ultimately about rates: of publication, of citation, of movement between categories. Rates presuppose denominators, and the denominators of science are themselves contested quantities.

The literature on the growth of science offers two sobering lessons. The first is that measured growth depends heavily on the database. Using cited references as a proxy for the literature that scientists actually use, \citet{bornmann2015growth} estimated the long-run growth of science at roughly three per cent per year with a doubling time of about a quarter of a century, while estimates based on the publication counts of individual databases vary widely with the database's coverage policy and its own expansion. A later analysis with segmented growth curves across several databases found that the growth rate has itself changed over time, with faster growth in recent decades, but that the location of the segments depends on the source \citep{bornmann2021growth}. For the present study the lesson is that a change in the volume of computer-science publications between 2018 and 2026 is the sum of a change in research activity and a change in indexing, and that only the first is of scientific interest. OpenAlex grew substantially during the period as it absorbed new sources and improved its matching; the indexing lag (the delay between a work's publication and its appearance in the snapshot) is longest for the most recent months, so that any series ending in 2026 slopes downward at its end for reasons that have nothing to do with research. These are among the pre-period diagnostics and sensitivity checks required by CLAUDE.md~\S6.5 and specified in Part~IV.

The second lesson concerns the ``science of science'' as a field in its own right. The programme described by \citet{fortunato2018sos} and developed at book length by \citet{wang2021sos} has established robust regularities: citation distributions are heavy-tailed within every field; team size has grown steadily; novelty in the sense of atypical combinations of prior work is associated with higher impact at the top of the distribution; and the productivity of individual scientists is approximately stable over careers. Those regularities are the background against which any field-specific claim must be read. A rise in the citation inequality of AI papers after 2022, for instance, is interesting only if it exceeds the inequality that heavy-tailed citation distributions produce in every growing field; the hypotheses H1--H8 of the preregistered protocol (Part~III) are formulated as comparisons between the post-2022 period and the baseline for that reason.

A further difficulty is specific to computer science and concerns what the discipline takes itself to be. \citet{denning2005science} argued that computer science is a science of information processes and not merely an engineering of artefacts, and \citet{wing2006computational} that ``computational thinking'' is a general intellectual skill rather than a professional one; both arguments were written at a moment when the discipline was defending its boundaries. The post-2022 period raises the converse question: whether the discipline's methods have diffused so widely that the boundary is no longer where the classifications place it. Papers that apply learned models to protein structure, materials, weather or law are classified as computer science by some systems and not others; a transformation that consists partly of such diffusion is partly invisible to a classification that keeps the boundary fixed. The ontology's eighth area, computational science and discovery, and the ``penumbra'' set defined in Part~III (works with a computer-science topic but without venue or arXiv support) exist to make that diffusion observable without inflating the core.

\section{Three prior transformations and how they propagated}
\label{sec:p2-waves}

The post-2022 period is not the first time computer science has been reorganised around a new technology. Three earlier transformations are directly relevant because they established the publication practices, the infrastructure, and the conceptual vocabulary within which the present one unfolds: the internet and web era, which created the systems and the industrial research laboratories; the deep-learning wave of 2012--2017, which created the benchmark culture and the compute economy; and the transformer and pretraining era of 2017--2022, which created the models themselves. For each, the account below is deliberately narrow. It does not attempt a history of computing; it asks what changed in how research was done and published, and how the change propagated from the subfield in which it began to the others. The answers suggest a way of reading the post-2022 data that the empirical parts of this paper then test.

\subsection{The internet and web era}
\label{sec:p2-web}

The World-Wide Web was presented to the computing community in 1994 as an information system built on three modest conventions, a naming scheme, a transfer protocol and a markup language, that happened to compose \citep{bernerslee1994www}. Within five years the problem of finding things on the web had produced two results that still structure information retrieval and network science: the use of the hyperlink graph as a source of authority, through PageRank in the Google prototype \citep{brin1998anatomy} and through the hubs-and-authorities formulation of \citet{kleinberg1999hubs}. Within another five, the engineering problem of running such systems at scale had produced the Google File System and MapReduce \citep{ghemawat2003gfs,dean2008mapreduce}, and within a decade the commoditisation of that infrastructure as ``cloud computing'' had been analysed as a shift in the economics of computing rather than in its technology \citep{armbrust2010cloud}. Apache Spark, which generalised MapReduce into a unified engine for iterative and interactive workloads, marks the maturation of the era \citep{zaharia2016spark}.

Three features of this transformation matter for what followed. First, the decisive results came from industrial laboratories or from academic groups that immediately founded companies, and they were published in the discipline's selective conferences (SOSP, OSDI, WWW, SIGCOMM) rather than in journals; the practice by which an industrial systems paper describes a deployed artefact, with measurements but without a released implementation, dates from this period and persists in the large-model reports discussed below. Second, the web made open-source software the ordinary form of a systems research artefact: Hadoop, the open re-implementation of MapReduce, propagated the method more effectively than the paper did, and the later machine-learning frameworks follow the same pattern. Third, the era created the data. Web-scale text, images, logs and graphs were the precondition for statistical methods in language, vision and recommendation, and the subfields that had access to such data (information retrieval, speech, machine translation) were the first to adopt statistical learning, a decade before the deep-learning wave made it general.

The propagation pattern was \emph{infrastructural}. The web changed the objects of study in networks, databases, distributed systems and security, and it created new subfields (web search, social computing, cloud computing), but it did not change the methods of theory, programming languages, graphics or human-computer interaction except by supplying them with new problems. The discipline's map was extended rather than redrawn: the ACM classification of 2012 could still accommodate every web-era subfield under its existing top-level nodes \citep{rous2012ccs}.

\subsection{The deep-learning wave, 2012--2017}
\label{sec:p2-dl}

The second transformation is usually dated to 2012, when a convolutional network trained on two GPUs reduced the top-5 error on the ImageNet Large Scale Visual Recognition Challenge by a margin that no incremental improvement in hand-designed features could match \citep{krizhevsky2017imagenet,deng2009imagenet,russakovsky2015imagenet}, and when deep networks replaced Gaussian mixtures in the acoustic models of the major speech laboratories at roughly the same time \citep{hinton2012speech}. The methods were not new: the long short-term memory cell dates from 1997 \citep{hochreiter1997lstm}, convolutional networks from the 1980s, and the retrospective of \citet{lecun2015deep} presents the decade's results as the delayed payoff of a thirty-year programme. What was new was the conjunction of labelled data at the scale of ImageNet, general-purpose graphics processors, and a benchmark with a public leaderboard.

The subsequent five years produced the architectural vocabulary that the transformer era would inherit. Distributed word representations trained by simple predictive objectives \citep{mikolov2013word2vec} made it natural to treat language as a sequence of dense vectors; the variational autoencoder \citep{kingma2014vae} and the generative adversarial network \citep{goodfellow2014gan} established deep generative modelling as a research programme; sequence-to-sequence learning with recurrent encoders and decoders \citep{sutskever2014seq2seq} and, crucially, the attention mechanism that let a decoder consult every encoder state \citep{bahdanau2015attention} reframed machine translation and then every conditional generation task; residual connections \citep{he2016resnet} removed the practical ceiling on depth. In reinforcement learning, deep Q-networks learned Atari games from pixels \citep{mnih2015dqn}, and the AlphaGo lineage combined deep networks with tree search to defeat professional Go players, first with human games as training data and then without them \citep{silver2016alphago,silver2017alphagozero}, generalising to chess and shogi with a single algorithm \citep{silver2018alphazero} and, in a different formulation, to real-time strategy games \citep{vinyals2019alphastar}.

The changes in publication practice were as consequential as the methods. Four deserve to be stated because the post-2022 data will be read against them.

The first is the \emph{benchmark as the unit of progress}. ImageNet established the pattern in which a public dataset, a fixed metric and a leaderboard define a subfield's agenda; the pattern was reproduced in language understanding, question answering and reinforcement learning, and it made progress legible to outsiders at the cost of concentrating effort on whatever the benchmark measured. The critiques of that culture, which argue that benchmarks are chosen for convenience rather than validity and that leaderboard competition rewards narrow optimisation \citep{dehghani2021lottery,raji2021benchmark}, were written in the transformer era but describe a practice born in this one.

The second is the \emph{preprint-first, conference-second} publication cycle. Machine-learning results began to appear on arXiv months before, or instead of, their conference versions; the practice spread from a minority to a norm within the decade, and it is the reason the arXiv \texttt{cs} harvest is indispensable to this corpus. The accompanying costs were catalogued at the end of the period: speculation presented as explanation, mathematics as decoration, suggestive language and the conflation of explanation with description \citep{lipton2019troubling}. Concerns about reproducibility were raised at the same time, both in audits of what published papers actually report \citep{gundersen2018reproducibility} and in the organised reproducibility programmes that the major conferences introduced in response \citep{pineau2021reproducibility}.

The third is the \emph{open-source framework as a shared instrument}. TensorFlow and PyTorch \citep{abadi2016tensorflow,paszke2019pytorch}, released by Google and Facebook respectively, lowered the cost of entry to deep learning to the point where a method published on arXiv could be reproduced and extended within days; they also standardised the computational substrate on NVIDIA GPUs and, later, on tensor processing units \citep{jouppi2017tpu}, which made compute a measurable input to research. The retrospective analysis of \citet{sevilla2022compute} identifies this period as the start of a ``deep learning era'' in which the training compute of notable systems grew with a doubling time of about six months, far faster than before 2010; \citet{thompson2020limits} argued that such growth is economically and environmentally unsustainable, a theme to which the sustainable-computing category of this study returns.

The fourth is the \emph{entry of industry as a producer rather than a consumer of research}. The laboratories of Google, DeepMind, Facebook, Microsoft and Baidu published the decisive results of the period in the academic venues and on arXiv, and their access to data and compute began to separate them from universities. \citet{ahmed2020dedemocratization} documented the growing presence of large technology firms and elite universities at the major AI conferences and the relative decline of other institutions; the later and broader analysis of \citet{ahmed2023industry} found that industry produced an increasing share of the largest models, the most-cited papers and the leading benchmarks. Whether that concentration intensified after 2022 is hypothesis H6 of the protocol.

The propagation pattern of this wave was \emph{methodological} and sequential. Deep learning entered computer vision and speech first, natural language processing second (roughly 2014--2016, with the sequence-to-sequence results), reinforcement learning and robotics third, and then, more slowly and with more resistance, the subfields whose objects are not perceptual: program analysis \citep{allamanis2018survey}, systems (learned index structures and learned scheduling), databases, and scientific computing \citep{raissi2019pinn}. In each case the receiving subfield first treated learning as a tool for a pre-existing problem, and only later, if at all, reorganised its problems around it. That two-stage pattern, tool adoption followed by problem redefinition, is the template against which Part~IX reads the post-2022 movement of methods across the forty categories.

\subsection{The transformer and pretraining era, 2017--2022}
\label{sec:p2-transformer}

The third transformation began with an architecture and ended with a change in what a model is for. The Transformer of \citet{vaswani2017attention} replaced recurrence with self-attention in the encoder--decoder framework of neural machine translation; its immediate merit was parallelism in training, and its lasting one was that the same block could be stacked to any depth and applied to any sequence of tokens. Within eighteen months it had been combined with the idea, already present in ELMo's contextual word vectors \citep{peters2018elmo}, that a language model trained on unlabelled text could be \emph{pretrained} once and \emph{fine-tuned} on many tasks: BERT \citep{devlin2019bert}, with its masked-language objective, set new results on every task of the GLUE benchmark \citep{wang2018glue} and on SQuAD \citep{rajpurkar2016squad}, and its successors showed that most of its design choices were less important than the amount of data and training \citep{liu2019roberta,raffel2020t5}. The generative branch of the family, the GPT series of decoder-only models, reached its third member in 2020 with 175 billion parameters and the demonstration that a sufficiently large language model could perform new tasks from a few examples placed in its prompt, without any gradient update \citep{brown2020gpt3}. In-context learning, more than any architectural novelty, is what separates the transformer era from the deep-learning wave: the unit of research shifted from the task-specific model to the general model and the prompt.

The era had a quantitative theory of its own growth. \citet{kaplan2020scaling} reported that the test loss of autoregressive language models falls as a power law in parameters, data and compute over several orders of magnitude, with the implication that larger models are more sample-efficient; \citet{hoffmann2022chinchilla} corrected the allocation, showing that models of the GPT-3 generation were undertrained relative to their size and that compute-optimal training requires data and parameters to scale together. Between the two papers, laboratories built models of increasing size on the strength of the first law (Gopher, PaLM, OPT and the open multilingual BLOOM \citep{rae2021gopher,chowdhery2022palm,zhang2022opt,bigscience2022bloom}), and the systems research needed to train them, model parallelism, memory-partitioned optimisers and IO-aware attention kernels \citep{shoeybi2019megatron,rajbhandari2020zero,dao2022flashattention}, became a research area in its own right, as did sparse mixtures of experts that decouple parameter count from per-token compute \citep{fedus2022switch}. The claim that certain abilities appear discontinuously with scale \citep{wei2022emergent} became one of the era's most cited and most contested propositions; the companion observation that prompting a model to write out intermediate reasoning improves its answers on arithmetic and symbolic tasks \citep{wei2022cot} established prompting as a method. New benchmarks were built at the scale of the models: a fifty-seven-subject examination of knowledge \citep{hendrycks2021mmlu}, a harder successor to GLUE \citep{wang2019superglue}, and a collaborative suite of more than two hundred tasks explicitly designed to resist saturation \citep{srivastava2023bigbench}.

The transformer did not remain a language architecture. Vision transformers showed that image patches could be treated as tokens and that, with enough data, the inductive bias of convolution was dispensable \citep{dosovitskiy2021vit}; self-supervised objectives, contrastive \citep{chen2020simclr}, self-distilled \citep{caron2021dino} and masked \citep{he2022mae}, removed the dependence on labels. Contrastive language--image pretraining aligned the two modalities in a shared embedding space and produced zero-shot classifiers from natural-language prompts \citep{radford2021clip}; in the same months, autoregressive text-to-image generation \citep{ramesh2021dalle} and then diffusion models, a decade-old idea \citep{sohl2015diffusion} made practical by denoising objectives \citep{ho2020ddpm}, score-based formulations \citep{song2021score} and architectural refinements that overtook adversarial models \citep{dhariwal2021beats}, converged on text-conditioned image synthesis at photographic quality \citep{ramesh2022dalle2,saharia2022imagen}. The latent-diffusion formulation \citep{rombach2022ldm}, released with open weights in the summer of 2022 under the name Stable Diffusion, is the generative-AI counterpart of ChatGPT: a system whose public availability, more than its technical novelty, changed who could use it. Speech followed the self-supervised path \citep{baevski2020wav2vec}; few-shot visual language models \citep{alayrac2022flamingo} and a single network trained across text, images, games and robot control \citep{reed2022gato} prefigured the multimodal and agentic systems of the following period. In structural biology, a transformer-based system solved, for most single-chain proteins, a problem that had resisted fifty years of effort \citep{jumper2021alphafold}, and was quickly joined by an independent three-track network \citep{baek2021rosettafold}.

Two further lines of work prepared the specific object that appeared in November 2022. Models trained on public code could synthesise functions from docstrings \citep{chen2021codex}, which turned software engineering into an early adopter of large models and created the first commercial product built on one. And the problem of making a language model do what a user asks, rather than what the pretraining distribution predicts, was addressed by reinforcement learning from human preferences, a method developed for control tasks \citep{christiano2017rlhf}, applied to summarisation \citep{stiennon2020summarize}, and then used to fine-tune GPT-3 into InstructGPT \citep{ouyang2022instructgpt}, whose 1.3-billion-parameter version was preferred by raters to the 175-billion-parameter base model. The same method, framed around helpfulness and harmlessness, was being developed in parallel elsewhere \citep{bai2022hh}. Retrieval-augmented generation, which conditions a generator on documents retrieved from an external index, was proposed in the same period \citep{lewis2020rag} and would become the standard answer to the models' factual unreliability.

The era also produced its own critique. The report that coined the term ``foundation model'' argued that the homogenisation of research around a few pretrained models was a change in the structure of the field, with risks that were social and institutional as much as technical \citep{bommasani2021foundation}; the ``stochastic parrots'' paper questioned whether ever-larger language models were a sound research direction at all, on environmental, documentary and epistemic grounds \citep{bender2021parrots}. The environmental argument had been quantified for NLP models by \citet{strubell2019energy}, contested and refined for the largest models by \citet{patterson2021carbon}, and turned into a research agenda by \citet{schwartz2020green}; the carbon footprint of a 176-billion-parameter model was measured end to end by \citet{luccioni2023bloom}.

What changed in publication practice during this era can be summarised in three observations. The leading models were described in \emph{technical reports} rather than in peer-reviewed papers, often with hundreds of authors, sometimes without the training data, the code, or the weights; the preprint became the version of record for the most consequential work in the discipline. The infrastructure for sharing models became an institution in its own right: the Transformers library \citep{wolf2020transformers} and the model hub built around it made the pretrained model, not the paper, the thing that propagated. And the compute required to participate at the frontier grew beyond the reach of universities, so that the concentration documented by \citet{ahmed2023industry} is, in this era, a concentration of capability rather than merely of attention.

\subsection{A comparative reading of the three waves}
\label{sec:p2-compare}

Table~\ref{tab:p2-waves} summarises the three transformations along the dimensions that matter for the empirical design of this study. The comparison is qualitative; the corpus-level counterpart is the subject of Part~IX.

\begin{table}[htbp]
\centering
\caption{Three prior transformations of computer science compared on the dimensions relevant to the post-2022 analysis. Qualitative characterisation by the authors from the works cited in Sections~\ref{sec:p2-web}--\ref{sec:p2-transformer}; no corpus quantities.}
\label{tab:p2-waves}
\small
\begin{tabularx}{\textwidth}{@{}lXXX@{}}
\toprule
 & Internet and web (c.~1994--2010) & Deep learning (2012--2017) & Transformer and pretraining (2017--2022) \\
\midrule
Originating subfields & Networks, information retrieval, distributed systems & Computer vision, speech & Natural language processing \\
Core artefact & Protocols, search engines, warehouse-scale systems & Trained task-specific networks; frameworks & Pretrained general models; prompts; weights \\
Locus of production & Industry and start-ups; academic systems groups & Academic groups with industry laboratories entering & Industry laboratories, with academic participation declining at the frontier \\
Publication practice & Selective conferences; deployed-artefact papers; open-source re-implementations & arXiv-first; leaderboards; shared frameworks; reproducibility programmes & Technical reports; model hubs; very large author lists; preprint as version of record \\
Propagation pattern & Infrastructural: new objects and subfields, methods of other areas unchanged & Methodological and sequential: tool adoption then problem redefinition, perceptual fields first & Substitutive: one architecture and one pretraining recipe across modalities; general models replace task-specific ones \\
Visibility in classifications & Accommodated by existing top-level nodes & New venues and arXiv classes gain weight; topics split & Vocabulary outruns citation clusters; keyword-defined categories required \\
\bottomrule
\end{tabularx}
\end{table}

Three conclusions follow from the comparison and shape the hypotheses tested later. First, the mechanisms of propagation differ in kind. The web changed the problems of other subfields; deep learning changed their methods; the transformer era changed the \emph{models} they use, which can be adopted without changing either problems or methods. A subfield can therefore register the third transformation at low cost, by using a pretrained model as a component, and the resulting growth in the vocabulary of the subfield need not indicate a reorganisation of its research. Distinguishing component adoption from reorganisation is the purpose of the citation-flow and embedding analyses of Parts~IV and~VIII.

Second, each wave shortened the publication cycle and shifted the version of record toward the preprint. The consequence for this study is that the arXiv harvest and the OpenAlex index measure different things: the former records when a result was first made public, the latter when it entered the citation system. A break detected in one series and not in the other is informative about the mechanism, not merely about measurement.

Third, the three waves were cumulative in infrastructure and in institutions. The industrial laboratories that produced the web systems trained the generation that produced deep learning, and the frameworks and compute of the deep-learning wave made the transformer era possible. The post-2022 period inherited all of it: the laboratories, the frameworks, the model hubs, the benchmark culture, and the preprint norm. Whatever is new about the period must be identified against that inheritance rather than attributed to a single product release, and the baseline window of 2018--2022 was chosen so that the inheritance is observed directly rather than assumed.

\section{The non-AI areas at the baseline, 2018--2022}
\label{sec:p2-baseline}

The hypotheses of this study concern not only the growth of AI but the condition of everything else. H2 predicts that at least one non-AI area declines in relative share while growing in absolute terms; H3 predicts that formal verification, computational complexity, operating systems and computer architecture show \emph{no} structural break in the window around November 2022; H4 predicts that citations crossing from those areas into AI increase. Such predictions can be evaluated only if the state of each area at the baseline is described independently of the AI narrative. This section provides that description for the thirty-one non-AI leaves of the ontology, grouped by area, in a deliberately compressed form: for each, the problems that defined the 2018--2022 literature, the landmark results that the corpus's citation data should recover, and the points of contact with machine learning that existed before 2022. Where a result is known to the authors only from a report rather than from a peer-reviewed paper, that is stated.

\subsection{Systems, cloud and distributed computing}
\label{sec:p2-systems}

The systems literature of the baseline period was organised around the consequences of the warehouse-scale computer \citep{barroso2018datacenter}. Three agendas dominated. The first was the \emph{serverless} abstraction, in which the unit of deployment is a stateless function and the provider manages everything else; the Berkeley view presented it as the next step in the simplification of cloud programming \citep{jonas2019serverless}, while a contemporaneous critique from the database community argued that its restrictions on state, communication and hardware access made it a step backward for data-intensive work \citep{hellerstein2019serverless}. The disagreement was productive: much of the 2019--2022 systems literature consists of attempts to give serverless platforms the state and the locality they lacked. The second agenda was \emph{resource disaggregation}, the separation of compute, memory and storage into independently scaled pools connected by fast networks, an idea with roots in blade-server memory expansion \citep{lim2009disaggregated} that became a central topic of operating-systems research once network latency fell within an order of magnitude of memory latency. The third was \emph{tail latency}: in a service composed of thousands of components, rare slow responses in any of them dominate the user-visible distribution, and the techniques of hedged requests, replication and tied requests described by \citet{dean2013tail} became the standard vocabulary of large-scale service design.

Operating systems research in the period was marked by the consequences of the Spectre and Meltdown disclosures, discussed under security below, which forced kernel designers to reconsider the cost of isolation, and by the growth of in-kernel programmability. Distributed computing as a theory remained focused on consensus, fault tolerance and the coordination of replicated state, with blockchain systems supplying new adversarial models (Section~\ref{sec:p2-security}). The contact with machine learning before 2022 was of two kinds: ``systems for ML'', the training and serving infrastructure of Section~\ref{sec:p2-transformer}, which systems venues absorbed readily, and ``ML for systems'', the use of learned models for scheduling, configuration and resource management, which was an active but minority topic. The ontology treats cloud and edge computing as children of distributed computing precisely because the venue structure does: OSDI, SOSP, EuroSys and NSDI publish all three.

\subsection{Hardware and computer architecture}
\label{sec:p2-hardware}

The baseline period in computer architecture was dominated by the end of two scaling regimes. Dennard scaling, the observation that transistor power density stays constant as transistors shrink \citep{dennard1974scaling}, had ended in the mid-2000s; its consequence, that a growing fraction of a chip must be left dark at any time, was quantified by \citet{esmaeilzadeh2011dark} and set the agenda for the following decade. The slowing of Moore's law completed the picture. The Turing lecture of \citet{hennessy2019golden} drew the conclusion that general-purpose performance gains were over and that the future lay in domain-specific architectures, open instruction sets and agile hardware design; \citet{leiserson2020room} argued in parallel that performance would henceforth come from software, algorithms and hardware specialisation rather than from the ``bottom'' of the stack. The tensor processing unit \citep{jouppi2017tpu} is the canonical example of the domain-specific architecture, and its publication in a mainstream architecture venue is a measure of how completely machine learning had become the architecture community's defining workload by 2017. AI accelerators are a child of computer architecture in the ontology for that reason, and the hypothesis that the architecture category shows no post-2022 break is, in effect, the hypothesis that this absorption was complete before ChatGPT.

Compilers underwent a parallel reorganisation around machine-learning workloads. The multi-level intermediate representation MLIR \citep{lattner2021mlir}, designed to let one compiler infrastructure serve domain-specific languages, tensor programs and hardware targets alike, and the end-to-end tensor compiler TVM \citep{chen2018tvm} represent the state of the art at the baseline. Both treat the deep-learning graph as a first-class compilation target; neither depends on large language models.

\subsection{Theory of computation}
\label{sec:p2-theory}

Two developments characterise theoretical computer science at the baseline, one internal and one at the boundary with machine learning.

The internal development is \emph{fine-grained complexity}: the programme of explaining why polynomial-time problems such as edit distance, all-pairs shortest paths and orthogonal vectors have resisted improvement for decades, by proving that faster algorithms for one would contradict a small number of hardness conjectures, above all the strong exponential-time hypothesis \citep{vassilevska2015hardness,williams2018finegrained}. The programme reorganised algorithms research around conditional lower bounds in the same way that NP-completeness had organised it around unconditional intractability, and its vocabulary is a reliable marker of the complexity category in this corpus.

The boundary development is the \emph{theory of deep learning}. The empirical observation that networks with enough parameters to memorise random labels nevertheless generalise \citep{zhang2021rethinking} undermined the classical uniform-convergence account of generalisation and prompted new theory: the double-descent curve, which reconciles the classical bias--variance trade-off with the behaviour of overparameterised models \citep{belkin2019reconciling}, and the neural tangent kernel, which characterises the training dynamics of infinitely wide networks \citep{jacot2018ntk}. The machine-learning-theory leaf of the ontology houses this literature, which belongs sociologically to the learning conferences and intellectually to the theory community. Its presence is a reminder that the theory area was in contact with AI long before 2022, through learning theory, optimisation and the complexity of sampling and inference.

\subsection{Programming languages, compilers and formal verification}
\label{sec:p2-pl}

The programming-languages literature of the baseline is marked by the success of a language. Rust, whose ownership type system gives memory safety without garbage collection \citep{matsakis2014rust}, became the first systems language to be adopted at scale on the strength of a safety guarantee; the RustBelt project supplied the semantic foundation by proving, in a separation-logic framework mechanised in Coq, that the safe fragment is sound and that the unsafe library code on which it rests respects the invariants \citep{jung2017rustbelt}. The result is characteristic of the period's formal-methods culture: a verification effort that follows an engineering artefact rather than preceding it, and that uses a proof assistant as its medium.

Proof assistants themselves reached a new stage. The verified C compiler CompCert \citep{leroy2009compcert} and the verified microkernel seL4 \citep{klein2009sel4} had shown a decade earlier that production software could be proved correct; the Kepler conjecture was formally proved in 2017 after a twelve-year effort \citep{hales2017kepler}; and the Lean theorem prover \citep{demoura2015lean}, re-implemented as a programming language in its fourth version \citep{demoura2021lean4}, acquired a community-maintained mathematical library whose scale and organisation were new in formalised mathematics \citep{mathlib2020}. The relevance to the post-2022 period is direct: Lean and its library became the environment in which language models were later asked to prove theorems, so that formal verification, one of the categories predicted \emph{not} to break in H3, is also one where the contact with AI is most visible in the recent literature. The hypothesis is therefore a sharp one.

Software engineering had already begun to absorb learned models before 2022. The survey of \citet{allamanis2018survey} organised a decade of ``big code'' research around the naturalness hypothesis, that source code is as statistically regular as natural language; the CodeXGLUE benchmark collection \citep{lu2021codexglue} standardised evaluation; and the Codex models \citep{chen2021codex} made program synthesis from natural language a product. The software-engineering category is thus expected to show an early and strong AI signal, and its baseline is not an AI-free one.

\subsection{Security and cryptography}
\label{sec:p2-security}

Three events organise the security and cryptography literature of the baseline. The first is the disclosure, in January 2018, of Spectre and Meltdown, which showed that speculative execution in essentially all high-performance processors leaks data across isolation boundaries \citep{kocher2019spectre,lipp2020meltdown}. The disclosures created a sustained literature on microarchitectural side channels and their mitigations that cuts across security, architecture and operating systems, and they illustrate the kind of cross-area citation flow that H4 measures.

The second is the standardisation of \emph{post-quantum cryptography}. Shor's algorithm \citep{shor1997algorithms} had made the eventual obsolescence of RSA and elliptic-curve schemes a matter of engineering timelines, and the case for migrating before large quantum computers exist was made to a general scientific audience by \citet{bernstein2017pqc}. The process run by the United States National Institute of Standards and Technology from 2016 selected lattice-based and hash-based schemes, published as federal standards in August 2024 \citep{nist2024fips203,nist2024fips204,nist2024fips205}. The applied-cryptography category of this corpus therefore contains a large and growing body of work whose dynamics are driven by a standards calendar entirely independent of language models, which makes it a useful placebo field. Fully homomorphic encryption, shown to be possible by \citet{gentry2009fhe}, remained the subject of a steady literature on efficiency.

The third is the maturation of \emph{adversarial machine learning}. The observation that imperceptible perturbations can change a network's output \citep{szegedy2014intriguing} and the attacks and evaluation standards that followed \citep{carlini2017robustness} created, by 2018, a field with ten years of history and a recognised research agenda \citep{biggio2018wild}. The ontology scores adversarial work in both cybersecurity and AI safety, which reflects the baseline state: the problem belongs to both communities and neither owns it. Blockchain systems, finally, entered the period with a systematised research agenda \citep{bonneau2015sok} and left it as a mature if contested area whose publication dynamics follow cryptocurrency markets more than academic calendars.