The Impact of the AI Era on Incremental Knowledge

Section I

The Nature of AI: The Most Efficient Stock Reuse Machine in History
What Large Language Models Actually Are

What is the technical essence of a large language model? Strip away all the commercial narratives, and it is a machine that compresses, indexes, and recombines the entirety of humanity’s digitized stock information. Its training data comes from text, code, papers, books, and conversation logs on the internet—all of which are digital mappings of the stock knowledge humanity has accumulated over thousands of years. The model’s parameters store compressed representations of statistical correlations within this information. The inference process finds, given an input context, the statistically most probable output sequence within this compressed space.

Every step of this process is stock reuse, not increment production. AI does not discover new laws of physics, does not invent new mathematical theorems, does not establish new causal relationships. It rearranges and recombines already-known information at extreme speed, generating output that appears novel but, in information-theoretic terms, contains no new information.

1.1 The Lesson of AlphaFold: Success in Domain-Specific Stock Reuse Is Not Success in Increment Production

AlphaFold is frequently cited as the canonical case of “AI producing scientific increments.” But careful examination of its technical architecture reveals precisely the opposite—it is the finest demonstration of AI as a stock-reuse tool. AlphaFold learned the mapping patterns from amino acid sequences to three-dimensional structures through supervised learning on approximately 215,000 experimentally determined structures in the Protein Data Bank (PDB). This training data was measured one structure at a time by human scientists over decades using physical experimental methods such as X-ray crystallography and nuclear magnetic resonance. What AlphaFold does is perform efficient pattern learning and generalization on this closed dataset.

The critical evidence lies in its failure modes: AlphaFold has not solved the physical mechanism of protein folding, nor has it identified folding pathways. It cannot capture conformational dynamics or allosteric effects. When encountering unusual conformations not covered by its training data, deviations between predicted and experimental structures can exceed 30 angstroms. Its capability ceiling is precisely determined by the training data provided by human experimenters—this is a success of domain-data stock reuse, not a success of knowledge-generalized increment production.

The entirety of AI’s capabilities rests on a single premise: someone is continuously producing incremental knowledge and injecting it into humanity’s information set. Without new papers being published, AI’s scientific knowledge remains frozen at its training cutoff. Without new code being developed, AI’s programming capability cannot cover new frameworks. AI is a waterwheel on a river—it can extract work from the current with extreme efficiency, but it does not produce the water.

Section II

Accelerated Free-Riding: The Cliff Already Underway
Empirical Evidence of the Collapse in Progress

The companion paper demonstrated the core dilemma of incremental knowledge: strong positive externalities prevent producers from internalizing the majority of their returns. But in the pre-AI era, there was at least a time window between when incremental knowledge was “produced” and when it was “fully free-ridden by society.” AI has compressed that window to near zero—and this is not a theoretical prediction; empirical data already prove that the cliff is occurring.

2.1 Empirical Evidence: A Cliff, Not a Slope

The Ramp Economics Lab tracked changes in corporate spending between freelance marketplaces and AI model providers from 2021 to 2025. The results are staggering: the share of corporate spending on freelance marketplaces (Upwork, Fiverr) plummeted from 0.66% in Q4 2021 to 0.14% in Q3 2025. Over the same period, spending on AI model providers surged from zero to nearly 3%. More than half of the companies that used freelancers in 2022 have stopped entirely. Most critically, the substitution ratio: for every $1 reduction in freelancer spending, companies added only $0.03 in AI spending—an economic measure of free-riding efficiency: obtaining 100% of the original output at 3% of the cost.

A joint study by Imperial College London, Harvard Business School, and the German Institute for Economic Research analyzed nearly 2 million freelance job postings across 61 countries: within just eight months of ChatGPT’s release, demand for freelance writing positions dropped by approximately 30%—the steepest decline across all categories. Direct feedback from creative professionals corroborated this trend—illustrators, art directors, and freelance writers alike reported that 2024 was “the hardest year in their careers.”

2.2 Threshold Effects: Why This Is a Cutoff, Not a Trickle

The exit of increment producers is not linear. An independent developer faces a survival threshold—when expected income falls below the cost of living, they switch careers entirely. A fundamental researcher does not “partially exit” science—when funding dries up and employment prospects vanish, they migrate to industry to perform stock-reuse work.

A possible counterargument is that a significant proportion of increment producers are driven by non-economic motivations—curiosity, a sense of mission, the pursuit of academic reputation. These individuals might continue producing increments even if their income drops to zero. History does indeed include many scientists who continued research in extreme poverty. But this group is exceedingly small, and even they require the most basic survival resources—food, shelter, laboratory equipment. When society as a whole ceases to pay for increment production, even these most basic conditions can no longer be guaranteed. Curiosity cannot substitute for lunch.

Section III

Why AI Cannot Produce Increments: Fundamental Deficits from a Cognitive Architecture Perspective
The Missing Biological Pathway

Faced with the argument that “AI is eliminating increment producers,” a natural counterargument arises: “Perhaps AI itself can produce increments?” Or a more moderate version: “Perhaps AI can help humans produce increments faster?” This section argues why, in the foreseeable future, neither counterargument holds.

3.1 What Is Missing Is Not Data Volume, but the Entire Cognitive Pathway

“Analysis of the Human Biological Cognitive Front-End System” established a four-layer cognitive model from the perceptual layer to the abstract layer. In this model, incremental knowledge is not generated during the phase of explicit reasoning (System 2), but rather through unconscious convergence on compute-storage-unified biological hardware (System 1). The most important breakthroughs in the history of science—Kekulé dreaming of a snake biting its own tail to discover the benzene ring structure, Einstein imagining riding a beam of light to trigger relativity, Mendeleev seeing the arrangement of elements in a dream—were all products of System 1. Scientists first use System 2 to accumulate vast quantities of multidimensional information, and then, in a relaxed state, the compute-storage-unified system continues running multidimensional information recombination in the background, until at some moment it completes convergence and a fully formed image erupts into consciousness.

What AI lacks is not “more data” or “stronger reasoning chains,” but the entire biological pathway required to produce increments:

First, multidimensional physical perception. Humans simultaneously collect information about the same physical object through 10+ sensory dimensions (vision, hearing, touch, smell, taste, proprioception, vestibular sense, thermoreception, nociception, interoception, etc.), forming a high-dimensional constraint space. AI’s “multimodality” covers 2–3 channels (vision + text + partial audio), and these channels receive secondhand information that has already been filtered through human perception and then digitally encoded. No number of lines on a two-dimensional plane can enclose a closed surface in three-dimensional space—the missing dimensions cannot be compensated for by data volume.

Second, compute-storage-unified biological hardware. Every neuron in the human brain is simultaneously a computational unit and a storage unit. The connection strength of a synapse is the stored “data”; signal transmission between synapses is the “computation.” Knowledge is structure; structure is computation. Every successful act of categorical definition directly alters synaptic connections—meaning the brain grows stronger with use. AI runs on von Neumann architecture, where processor and memory are physically separated; inference ends and vanishes, with no continuously running background convergence process.

Third, unconscious image emergence. The final product of the human abstract layer is a perceivable, manipulable, rotatable, decomposable complete mental image. When AI generates the token “cat,” what is activated is a high-dimensional floating-point vector—containing no softness, no warmth, no purring. When a human thinks of “cat,” what emerges is a complete image encompassing shape, texture, sound, smell, weight, temperature, and emotional coloring. AI’s vectors are statistical distances, not images.

3.2 The Trap of “AI Helping Humans Accelerate Increment Production”

A more moderate counterargument is: “Perhaps AI can help humans produce increments faster?” This counterargument appears reasonable, but based on the cognitive architecture analysis above, it contains a fatal blind spot.

What does AI assist with? The explicit reasoning processes of System 2—helping you search the literature, organize data, arrange information, and accelerate coding. But incremental knowledge generation occurs in System 1—unconscious, multidimensional information recombination and image emergence on compute-storage-unified hardware. The quality of System 1’s inputs depends on the long-term accumulation of extensive multidimensional perceptual experiences—you need to personally conduct experiments, personally observe phenomena, and personally immerse yourself in a problem for months or even years, so that synaptic structures form sufficiently precise knowledge encodings to supply raw material for unconscious convergence.

The paradox of AI assistance lies in this: by accelerating the efficiency of System 2 work, it reduces the necessity for humans to personally engage in extended hands-on practice—yet it is precisely these extended hands-on practices that accumulate multidimensional perceptual experience for System 1. AI assists the least important link in increment production (information organization) while potentially damaging the most important link (the quality of multidimensional perceptual accumulation and the raw material for unconscious convergence).

3.3 Quantifying the Information Gap

At the quantitative level, the training data of the largest current language models is on the order of 10¹⁴ bits, with the effective information stored in model parameters after compression being approximately 10¹² bits (estimated).

Section IV

The Cognitive Bubble: A More Dangerous Illusion Than Financial Bubbles
When Society Mistakes Consumption for Creation

When large numbers of stock-reuse practitioners use AI to break through the knowledge limits of their own biological storage and can invoke information far exceeding their personal capacity, they develop a powerful illusion: “I have become more capable.” Companies decide they no longer need to maintain basic research teams. Investors conclude that AI is the future and allocate all capital to the application layer. Public discourse declares that “everything can be AI-ified.” Across society, the perceived need for incremental knowledge producers is declining, and the willingness to pay for them is vanishing.

This is the cognitive bubble. The cognitive bubble is far more dangerous than a financial bubble. When a financial bubble bursts, the real economy remains—factories still stand, people still exist, reconstruction is possible. When the cognitive bubble bursts, even the capacity for reconstruction is gone—because the producers of incremental knowledge have been systematically eliminated during the bubble’s inflation.

4.1 The Business Model of AI Companies: Industrialized Free-Riding

The business model of AI companies, translated into this paper’s framework, is as follows: compress the incremental output of the 1% across every domain into a model and sell it as a product to the 99% of stock-reuse practitioners. Every new feature released, every new tool integrated, every new model upgraded closes the trial-and-error cost recovery window in yet another domain. And the valuations of these products—in the hundreds of billions of dollars—are pricing “how fast and at what scale we can free-ride the incremental information of all humanity” as a capability.

4.2 Financial Bubble Nested Within Cognitive Bubble

The current AI financial bubble and cognitive bubble form a nested structure. The financial bubble (AI company valuations) rests upon the cognitive bubble (the belief that AI can substitute for increment production).

The operating logic of financial markets itself produces an interesting recursion here. AI is compressing information propagation speed differentials to zero, which means the financial arbitrage space based on propagation speed differentials is disappearing. The only remaining information asymmetry is genuine incremental knowledge—”new discoveries that others do not yet know about.” But as argued above, increment producers are being systematically eliminated. Information asymmetry itself, as a resource, is being exhausted.

The deeper problem is that financial investment has never been a purely information-asymmetry game—it also contains enormous irrational components. Newton was one of the greatest incremental knowledge producers in human history, yet he lost approximately 77% of his wealth in the South Sea Bubble of 1720. His famous remark summarized it all: “I can calculate the motion of heavenly bodies, but not the madness of people.” Research by Professor Andrew Odlyzko of the University of Minnesota found that Newton went from being a cautious, diversified investor to a speculator who bet nearly his entire fortune on a single stock. Analyzed through the cognitive front-end framework: Newton’s System 2 (rational reasoning) was among the most powerful in human history, but his System 1 was as susceptible to FOMO and herd effects as anyone else’s. Financial markets are a System 1 game, not a System 2 game. This is why high cognitive ability does not guarantee investment success—the dimension of financial gaming and the dimension of incremental knowledge production are on entirely different planes.

Section V

Software Dies First: The Earliest Collapse of Digital-State Increments
Physical Embedding Depth as Predictor

The “physical embedding depth” variable introduced in the companion paper yields a clear prediction here: digital-state increments will stagnate first; physical-state increments will be the last to fall.

Software is a purely digital-state product. Its replication cost is zero. AI’s parsing speed is fastest here. The trial-and-error cost recovery window is shortest. Ramp’s data has already confirmed this prediction—the markets for freelance writing, design, and basic coding are in cliff-edge collapse, with an extremely asymmetric substitution ratio ($1 replaced by $0.03).

Physical-state increments, by contrast—semiconductor fabrication processes, aircraft engine design, novel material synthesis—because their trial-and-error costs are embedded in physical systems and replication still requires time and resource investment approaching the original, are temporarily exempt from this compression. But “temporarily” is the key word—as AI-controlled robotic systems and automated laboratories advance, the free-riding cost of physical-state increments is also gradually declining.

Ironically, the domain AI is best at substituting is precisely software and text—purely digital-state increments. The first thing it eliminates is the innovation incentive in the very domain that feeds it. AI is emptying its own granary at maximum speed.

Section VI

The Death Spiral: A Self-Consumption Process Already Underway
Not a Prediction — A Process in Motion

The death spiral is not a prediction about the future. It is already happening.

6.1 Contamination and Exhaustion of Training Data

A 2025 Ahrefs analysis of nearly one million newly published web pages found that 74.2% contained detectable AI-generated content. Large-scale text analysis estimates that 30–40% of the active web corpus already consists of synthetic content. A joint study by MIT and the Oxford Internet Institute estimated that AI-generated content now constitutes 64% of all newly published internet content, with the AI-to-human output ratio reaching 17:1. Europol warned that by 2026, up to 90% of online content could be synthetically generated.

This means that AI’s future training data will increasingly contain AI’s own output. Research published in Nature by Shumailov et al. (2024) demonstrated that when models are trained on data they themselves have generated, irreversible quality degradation occurs—output diversity decreases, distributions collapse toward the mean, a few modes are continuously reinforced while the majority of modes are permanently lost.

It is important to distinguish precisely: model collapse in the strict sense (irreversible degradation from pure AI self-training) and the training data quality degradation actually occurring (human-originated content being diluted by AI-synthesized content) are two distinct but related phenomena. The latter does not require the extreme case of the former to occur before it is already damaging AI output quality—long-tail experiences and niche professional knowledge are systematically absent from synthetic text, template-driven expression increasingly dominates the corpus, and misinformation is treated as credible through mass replication. The Harvard Journal of Law & Technology has already proposed the legal concept of “the right to uncontaminated human-generated data,” noting that data collected before 2022 may become a competitive moat for AI developers—itself an institutionalized signal of the training data crisis.

More fundamentally: even if AI training data were not contaminated by synthetic content, the incremental growth of human-originated text data is decelerating. When increment producers exit because trial-and-error costs can no longer be recovered, the volume of high-quality original content injected into the internet declines. Research estimates that human-generated text data may face exhaustion as early as 2026. Training data contamination and the severing of incremental flow are two tributaries of the same death spiral, and they are converging.

6.2 The Complete Death Spiral Pathway

6.3 Historical Precedent and the Critical Difference

The collapse of an increment-production system leading to civilizational stagnation has historical precedent. After the fall of the Roman Empire, the institutional foundations that sustained incremental knowledge production were destroyed—academies vanished, scholarly lineages were severed, literacy rates plummeted, and even the formula for Roman concrete was lost. It was not the information itself that disappeared, but “the people and mechanisms capable of producing new information.” An entire civilization spun its wheels in stock for nearly a millennium.

That dark age ultimately ended because of a critical condition: an external source of incremental information existed. The Arab world had preserved and further developed Greco-Roman knowledge, which flowed back into Europe through the Crusades and trade routes, igniting the Renaissance.

If the cognitive bubble of the AI era is global—if the entire world simultaneously stops paying for increment production—then there is no external backup. No other civilizational sphere is independently maintaining increment production. Humanity would face, for the first time, an information dark age with no external rescue.

Section VII

Conclusion: AI Is Not Eating the World — AI Is Digging Its Own Grave
The Self-Consuming Logic of Maximum-Efficiency Free-Riding

The market narrative is “AI is eating software.” The causal relationship argued in this paper is precisely the reverse. Every industry AI “devours” is an incremental supply pipeline it severs for itself. The faster it devours, the faster it dies. Competition among AI companies is accelerating this process—the race is to see who can take a few more bites before the incremental flow is cut off. This is not commercial competition. This is predatory extraction of a non-renewable resource.

The core thesis of this paper can be compressed into a single sentence: