AI Has Hit a Wall
November, 2026
The diminishing returns of scaling, the limits of silicon, the exhaustion of data, and the growing disconnect between artificial intelligence and economic reality.
For more than a decade, the technology industry has been selling the world a remarkably simple proposition: artificial intelligence will keep getting smarter if we keep making it bigger. More parameters, more training data, more graphics processors, more electricity, more data centers, and more money. Scale was not merely an engineering strategy. It became an article of faith.
The formula appeared irresistible. If a billion parameters produced impressive results, a trillion would produce something extraordinary. If a few thousand GPUs could train a powerful model, a million GPUs might deliver artificial general intelligence. If today's models could write software, tomorrow's would replace entire engineering departments. If they could answer questions, perhaps they would soon discover new laws of physics, cure diseases, and solve humanity's most difficult problems.
The industry convinced investors, governments, corporations, and much of the public that intelligence was essentially a commodity that could be manufactured by pouring computational resources into increasingly enormous neural networks.
But there is a problem with this narrative.
The evidence is accumulating that artificial intelligence, at least in its prevailing architectural and economic form, has hit a wall.
Not one wall, but several. A scaling wall. A memory wall. A data wall. A reasoning wall. An energy wall. And perhaps most consequentially, an economic wall.
Each is formidable on its own. Together, they threaten the fundamental assumption on which the modern AI industry has built its future: that progress can continue indefinitely through brute-force scaling.
The remarkable thing is not that these limits exist. Every technology eventually encounters physical, mathematical, and economic constraints. The remarkable thing is how much money has been committed on the assumption that artificial intelligence would somehow be exempt.
The End of the Scaling Miracle
The modern AI revolution was built on the observation that increasing model size, training data, and computational resources often produces predictable improvements in performance.
OpenAI's 2020 paper, Scaling Laws for Neural Language Models, helped establish the relationship between computational investment and model performance. DeepMind's 2022 Chinchilla research subsequently demonstrated that model size and training data needed to be balanced for efficient training.
These studies helped create one of the most expensive engineering philosophies in history: scale first, and intelligence will follow.
The early results were spectacular. The transition from GPT-2 to GPT-3, and subsequently to GPT-4, produced dramatic improvements in language generation, coding, reasoning benchmarks, and general-purpose usefulness.
But early breakthroughs are not proof of unlimited future returns.
The distinction between exponential expenditure and diminishing improvements is becoming impossible to ignore. Increasingly large investments are required to extract improvements that are less transformative than the breakthroughs that originally justified the spending.
This is the familiar pattern of diminishing marginal returns.
A company might double its computational expenditure without doubling the useful intelligence of its models. It might increase training costs severalfold to achieve improvements that are noticeable on benchmarks but barely perceptible in ordinary business applications.
The underlying mathematics of scaling laws never guaranteed exponential growth in intelligence. Many of the relationships describe power-law improvements in losses, meaning that extracting successive increments of performance can require disproportionately larger computational investments.
This is a crucial distinction. The industry has frequently treated scaling laws as though they were laws of unlimited technological progress. They are not.
Even OpenAI's former chief scientist Ilya Sutskever signaled the changing landscape in a widely reported 2024 interview with Reuters, saying:
“Pre-training as we know it will unquestionably end.”
The significance of that statement is difficult to overstate. Pre-training at enormous scale was the engine that powered the modern large-language-model revolution.
When the architects of that revolution begin acknowledging that its central mechanism faces limits, the industry should pay attention.
The response has been to pursue new forms of scaling: longer reasoning traces, inference-time computation, reinforcement learning, synthetic data, specialized architectures, and enormous systems of interacting AI agents.
But shifting the expenditure from training to inference does not eliminate diminishing returns. It changes where the money is spent.
A system that requires substantially more computational effort to answer increasingly difficult questions is not necessarily escaping the scaling wall. It may simply be encountering that wall from another direction.
The Memory Wall: When Moving Data Costs More Than Computing
One of the least appreciated constraints on artificial intelligence is also among the most fundamental: modern processors can perform arithmetic much faster than memory systems can deliver the data required for that arithmetic.
The technology industry loves to advertise floating-point operations per second. GPU specifications prominently feature extraordinary numbers of teraFLOPS and petaFLOPS.
But theoretical computational capacity is not the same as useful computational throughput.
A processor cannot calculate with data it has not received.
As AI models grow, their parameters, intermediate activations, attention caches, and working memory place enormous demands on memory bandwidth and capacity. Moving this information between high-bandwidth memory, processor cores, accelerators, and interconnected servers becomes an increasingly important determinant of performance.
This is the memory wall.
The problem is not new. Researchers have discussed the growing disparity between processor performance and memory performance for decades. What is new is the extraordinary scale at which AI workloads are exposing it.
The classic 2009 paper The Landscape of Parallel Computing Research: A View from Berkeley emphasized the importance of communication and data movement in modern computing. Subsequent research into energy-efficient computing has repeatedly shown that moving data can consume far more energy than performing simple arithmetic operations.
Mark Horowitz's influential 2014 paper, Computing's Energy Problem (and What We Can Do About It), quantified the enormous energy differences between arithmetic and memory access.
This leads to a deeply uncomfortable conclusion for the AI industry.
Buying more arithmetic capability does not automatically produce proportional improvements in AI performance.
Adding GPUs can increase available compute, but it also increases pressure on memory systems, interconnects, synchronization mechanisms, and power infrastructure.
Large models must frequently be partitioned across multiple accelerators. This introduces communication overhead, latency, and increasingly complex scheduling requirements.
The bigger the system becomes, the more engineering effort is devoted to coordinating the hardware rather than performing the useful computation.
The situation resembles a factory that keeps purchasing faster manufacturing machines while failing to expand the roads, warehouses, and loading docks that supply them.
Eventually, the machines spend more time waiting for materials than manufacturing products.
AI infrastructure is increasingly confronting the computational equivalent.
More GPUs cannot indefinitely compensate for architectural inefficiency. At some point, the fundamental economics of moving information become the limiting factor.
The Data Wall: The Internet Is Not Infinite
The next constraint is even more fundamental.
Modern large language models depend on enormous quantities of human-generated information: books, websites, scientific papers, software repositories, encyclopedias, discussions, and other forms of digital content.
The extraordinary performance of these models is inseparable from the accumulated intellectual output of humanity.
But humanity has not produced an infinite amount of high-quality digital information.
In their research paper Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data, Pablo Villalobos and colleagues examined the availability of training data and projected that the supply of usable public human-generated text could become a significant constraint during the 2020s.
This challenges the assumption that training datasets can expand indefinitely alongside model size.
There is a fundamental difference between increasing the number of processors and increasing the amount of useful knowledge available for training.
A semiconductor manufacturer can produce more chips. A data center operator can purchase more servers. An investor can finance additional infrastructure.
But none of these activities automatically creates new scientific discoveries, new mathematical insights, or new high-quality human reasoning.
The industry's response has increasingly been synthetic data: using AI models to generate training examples for other AI models.
At first glance, the approach appears ingenious. If human-generated data is scarce, let machines generate unlimited additional data.
But this introduces a dangerous possibility: models learning from the outputs of other models, reproducing their errors, amplifying their biases, and progressively losing contact with the diversity and complexity of the original information.
A 2024 paper published in Nature, AI Models Collapse When Trained on Recursively Generated Data, by Ilia Shumailov and colleagues, demonstrated how recursively training generative models on generated content can degrade the resulting distributions.
The phenomenon is commonly called model collapse.
The implication is profound.
A photocopy of a photocopy does not create new information. Repeating the process indefinitely does not improve the original document.
Similarly, generating enormous quantities of synthetic language does not automatically generate new knowledge.
Carefully designed synthetic datasets can be useful. But the underlying problem remains: the system must obtain reliable new information from somewhere.
An AI model cannot manufacture an unlimited supply of verified scientific truth merely by generating more text.
The internet provided the first generation of large language models with an extraordinary inheritance: centuries of accumulated human knowledge.
That inheritance cannot simply be multiplied forever.
The Reasoning Wall: When Artificial Intelligence Stops Thinking
Perhaps the most damaging evidence against the prevailing AI narrative comes from the very capability that supposedly distinguishes intelligence from sophisticated pattern recognition: reasoning.
Modern AI systems can solve mathematical problems, generate elaborate chains of thought, write convincing explanations, and produce sophisticated computer programs.
But what happens when the underlying problems become sufficiently complex?
In June 2025, Apple researchers published The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity.
The researchers systematically evaluated large reasoning models using controlled puzzles in which the complexity could be increased while preserving the underlying logical structure.
Their findings were striking.
At low complexity, conventional language models could outperform specialized reasoning models. At intermediate complexity, additional reasoning computation offered advantages. But beyond certain complexity thresholds, the models experienced what the researchers described as a complete accuracy collapse.
Even more revealing, the models sometimes reduced their reasoning effort as problems became more difficult, despite having sufficient token budgets available.
The researchers also observed failures to apply explicit algorithms consistently, even when the relevant algorithms were straightforward to describe.
Apple Machine Learning Research
Consider the implications.
A calculator does not suddenly forget arithmetic because the numbers become larger. A correctly implemented algorithm does not abandon its logical rules because the problem requires additional steps. It may exhaust available memory or computational resources, but its underlying mathematical operations remain well-defined.
A reasoning model, by contrast, can generate pages of apparently thoughtful analysis while failing to preserve the logical consistency required to solve a difficult problem.
This distinction matters enormously.
Generating a convincing explanation is not the same as executing a correct reasoning procedure.
Producing intermediate steps that resemble mathematical deduction is not proof that a system has acquired reliable abstract reasoning.
A language model can reproduce the appearance of logic without possessing the algorithmic guarantees that make formal computation dependable.
The Apple research gives scientific substance to a concern that has been apparent to many users: AI can be astonishingly capable on familiar problems and surprisingly fragile when asked to solve unfamiliar combinations of those same problems.
The industry's answer has increasingly been to allocate more inference-time computation, allowing models to generate longer reasoning traces and explore more possibilities.
But more thinking tokens do not necessarily repair a defective reasoning mechanism.
If the system lacks a dependable way to preserve logical invariants, verify intermediate deductions, or systematically explore a problem space, additional computation may simply generate more elaborate mistakes.
The fundamental limitation is not necessarily how long AI thinks. It is whether the underlying architecture can reliably perform the kind of thinking the problem requires.
And that is an architectural question, not merely a question of scale.
The Enterprise Wall: Billions Invested, Little to Show
For years, AI companies have justified their extraordinary valuations by promising an equally extraordinary transformation of business.
Artificial intelligence would eliminate repetitive work, replace expensive software processes, reduce headcount, accelerate development, automate customer service, transform operations, and generate unprecedented productivity gains.
The economic promise was enormous.
The realized returns have been far less impressive.
A 2025 report from MIT's NANDA initiative, The GenAI Divide: State of AI in Business 2025, examined enterprise AI adoption through interviews, surveys, and analysis of hundreds of implementations.
Its headline finding was extraordinary: approximately 95% of integrated enterprise generative AI pilots were failing to produce measurable profit-and-loss impact. Only a small minority were generating substantial financial value.
The report described a sharp divide between widespread experimentation and meaningful business transformation.
This is not a minor implementation detail. It strikes directly at the economic justification for the AI investment boom.
There is an enormous difference between demonstrating that AI can perform a task and demonstrating that AI can profitably transform an organization.
A chatbot that produces an impressive demonstration is not necessarily a commercially viable enterprise system.
A model that summarizes documents is not necessarily capable of navigating a company's contractual obligations, operational procedures, data permissions, regulatory requirements, and exceptions.
An AI agent that successfully completes a benchmark workflow is not necessarily capable of operating reliably inside a business where mistakes have financial consequences.
The fundamental obstacle is the messy reality of enterprise information.
Corporate data is rarely clean, complete, consistent, or conveniently organized. It exists across databases, spreadsheets, emails, legacy applications, document repositories, internal systems, and the institutional knowledge of employees.
Different departments frequently maintain conflicting versions of the same information.
Access permissions are inconsistent. Business rules are undocumented. Exceptions are handled informally. Critical decisions depend on context that may never have been recorded.
This is the enterprise data chaos that AI demonstrations routinely overlook.
Imagine deploying an AI procurement agent into a multinational corporation.
The agent must understand supplier contracts, negotiated discounts, delivery schedules, regulatory restrictions, inventory levels, approval hierarchies, tax implications, and changing business priorities.
Some information may be stored in an ERP system. Some may exist in email correspondence. Some may be known only to individual employees.
An impressive language model cannot magically transform this fragmented environment into a coherent, trustworthy operational system.
The company must first undertake the difficult and expensive work of integration, governance, process redesign, security, and verification.
And after completing that work, the organization may discover that the AI model itself was the least expensive component of the entire project.
This is the great irony of enterprise AI.
The industry promised that intelligence would eliminate complexity. Instead, deploying AI frequently reveals how much complexity organizations have accumulated.
The result is a growing gap between technological demonstrations and economic outcomes.
Companies can purchase AI capabilities almost instantly. They cannot purchase organizational transformation with the same ease.
The Productivity Illusion: Faster Output Is Not Faster Work
One of the most persuasive arguments for AI has been its apparent ability to increase individual productivity.
A developer can generate hundreds of lines of code in seconds. A marketing professional can draft a campaign in minutes. A business analyst can produce a report almost instantly.
But productivity is not measured by how quickly intermediate artifacts are produced.
It is measured by how much useful, correct, completed work is delivered.
In July 2025, the research organization METR published a randomized controlled study examining the effect of AI coding tools on experienced open-source developers.
The researchers studied 16 developers completing 246 tasks in established software projects.
Before beginning, developers predicted that AI assistance would make them approximately 24% faster.
After completing the tasks, they believed AI had made them approximately 20% faster.
The measured result was the opposite.
Developers using AI tools took approximately 19% longer to complete their tasks than developers working without them.
This finding illustrates an important distinction between perceived productivity and actual productivity.
AI can make work feel faster because it produces visible output almost immediately.
But generating code is only one part of software engineering.
The code must be understood, reviewed, tested, debugged, integrated, secured, documented, and maintained.
AI-generated code may introduce subtle defects, incompatible assumptions, unnecessary dependencies, security vulnerabilities, or architectural inconsistencies.
A developer who saves twenty minutes generating code but spends an additional hour understanding and correcting it has not gained productivity.
The same principle applies beyond software engineering.
An AI-generated legal document still requires legal review. An AI-generated financial analysis still requires verification. An AI-generated medical recommendation still requires professional judgment. An AI-generated engineering design still requires validation against physical and safety constraints.
The more consequential the application, the greater the importance of verification.
This creates a structural productivity bottleneck.
AI accelerates the production of candidate outputs, but it does not automatically accelerate every downstream activity required to turn those outputs into dependable results.
In many organizations, the bottleneck simply moves.
Instead of spending time creating the first draft, employees spend time evaluating, correcting, and integrating machine-generated work.
The apparent productivity gain can disappear into a growing burden of supervision.
There is also a psychological dimension.
Employees may feel compelled to produce more because AI makes generating material easier. Managers may expect faster delivery without accounting for verification. Teams may become overwhelmed by an increasing volume of documents, code changes, proposals, and automated requests.
The organization becomes busier without necessarily becoming more productive.
AI can dramatically increase the quantity of work produced while delivering far smaller improvements in the quantity of work actually completed.
That distinction may prove to be one of the defining economic lessons of the AI era.
The Energy Wall: Intelligence Cannot Escape Physics
The AI industry frequently describes computational progress in abstract terms: parameters, tokens, FLOPS, context windows, and benchmark scores.
But beneath these abstractions lies an industrial reality.
Artificial intelligence requires electricity, cooling, water, semiconductor manufacturing, transmission infrastructure, land, and enormous capital expenditure.
The physical footprint of AI is becoming increasingly difficult to ignore.
The International Energy Agency's 2025 report Energy and AI estimated that data centers consumed approximately 415 terawatt-hours of electricity in 2024, representing around 1.5% of global electricity consumption.
Its baseline projection placed global data-center electricity demand near 945 terawatt-hours by 2030, with AI among the principal drivers of growth.
To appreciate the scale, this projected demand approaches the annual electricity consumption of a major industrialized nation.
And electricity is only one part of the infrastructure challenge.
Modern AI facilities require substantial cooling capacity. Advanced accelerators generate extraordinary heat densities. Power distribution equipment, transformers, substations, transmission lines, and cooling systems must all scale alongside the computational hardware.
Unlike software, these physical systems cannot be expanded instantly.
A new AI model can be released overnight. A new power station, transmission corridor, or major data-center campus can require years of planning, permitting, construction, and commissioning.
The AI industry operates on software timelines while depending on infrastructure that operates on industrial timelines.
This mismatch creates increasingly serious economic consequences.
Data centers compete with other industries and communities for electricity, land, and water. Local governments must evaluate whether the economic benefits justify the infrastructure requirements.
Communities are becoming more vocal about the consequences of massive data-center developments, particularly where electricity prices, water availability, environmental impacts, and land use are already contentious.
And there is an even more fundamental problem.
If increasingly large computational investments produce increasingly small improvements in useful intelligence, the energy required for each additional unit of capability may become economically unattractive.
A model that costs ten times as much to train but delivers only modest additional commercial value represents a deteriorating economic proposition.
The same applies to inference.
A reasoning system that consumes dramatically more computational resources to solve marginally more difficult problems may be technically impressive but commercially unsustainable at scale.
There is no law of nature requiring the economics of AI to improve forever.
The laws of thermodynamics, semiconductor physics, and electrical engineering are indifferent to investor expectations.
The Capital Wall: When Expenditure Becomes the Business Model
The final wall may be the one that matters most to investors.
The modern AI industry has become extraordinarily capital-intensive.
Training frontier models requires vast computational clusters. Operating them requires continuing expenditure on electricity, cooling, networking, maintenance, and hardware replacement.
Serving millions of users creates substantial inference costs. Increasing reasoning capabilities can make individual queries more computationally expensive.
Meanwhile, competition places downward pressure on the prices customers are willing to pay.
This creates a difficult economic equation.
The cost of producing ever more advanced models can rise dramatically while the market price of intelligence continues to fall.
A capability that commands a premium today may become widely available tomorrow.
As competing models converge in their practical usefulness, customers gain the ability to switch between providers. The underlying models become increasingly interchangeable for many routine applications.
This undermines the assumption that every additional dollar invested in model development will produce durable pricing power.
The semiconductor industry has seen variations of this pattern before. So have telecommunications, cloud infrastructure, and numerous capital-intensive technologies.
An industry can experience spectacular technological progress while producing disappointing returns for many of its investors.
The historical analogy with telecommunications is particularly instructive.
During the late 1990s, investors financed enormous expansions of fiber-optic networks on the expectation that internet traffic would grow without limit.
The demand for internet connectivity did indeed grow dramatically.
But that did not prevent the collapse of numerous telecommunications investments when infrastructure expansion ran ahead of commercially realizable demand.
The technology was valuable. The financial assumptions were unsustainable.
The same distinction is essential in AI.
Artificial intelligence can remain useful, widely adopted, and economically important while the companies financing its infrastructure experience disappointing returns.
The value of a technology and the value of the investments made in anticipation of that technology are not the same thing.
An AI system might save a customer a few dollars per task while requiring an infrastructure provider to commit billions of dollars to deliver that capability at scale.
The customer benefits. The infrastructure provider may struggle to earn an adequate return.
This is particularly concerning when enormous capital commitments are justified by expectations of future applications that have not yet demonstrated corresponding profitability.
The AI industry is increasingly asking investors to finance the infrastructure first and trust that the economic demand will eventually materialize.
History offers no guarantee that it will.
The Architecture Wall: More of the Same Is Not a Breakthrough
All these constraints point toward a deeper problem.
The AI industry has become extraordinarily skilled at scaling a relatively narrow family of computational architectures.
Transformers, attention mechanisms, gradient-based optimization, and large-scale statistical learning have delivered remarkable capabilities.
But scaling an architecture is not the same as discovering a fundamentally better architecture.
The history of computing is full of examples in which progress required conceptual breakthroughs rather than simply larger implementations of existing techniques.
The transistor did not emerge because engineers built increasingly enormous vacuum tubes.
Integrated circuits did not emerge because manufacturers found ways to connect infinitely more discrete components.
Modern computing did not become possible merely by making existing mechanical calculators larger.
At critical moments, technological progress required new ideas.
The same principle applies to artificial intelligence.
If the prevailing architectures struggle with exact reasoning, memory efficiency, data dependence, and computational cost, then building larger versions of those architectures cannot be assumed to eliminate their fundamental limitations.
A neural network that has not learned a reliable algorithm does not necessarily acquire that algorithm simply because its parameter count increases.
A model that lacks access to new information cannot create unlimited verified knowledge merely by generating synthetic examples.
A computational system constrained by memory bandwidth does not escape that constraint merely because additional processors are installed.
A business process that requires human accountability does not become autonomous simply because an AI model can generate convincing instructions.
The industry's instinct has been to answer architectural limitations with additional scale.
More parameters. More GPUs. More data centers. More tokens. More agents.
But at some point, adding resources to an inefficient approach becomes an increasingly expensive substitute for genuine innovation.
The next major advances in artificial intelligence may require fundamentally different approaches to computation, memory, reasoning, learning, and knowledge representation.
They may require architectures that integrate explicit algorithms, symbolic reasoning, formal verification, structured memory, causal models, or entirely new computational paradigms.
Whatever those breakthroughs turn out to be, they cannot be assumed to emerge automatically from larger clusters of the same hardware running larger versions of the same models.
The Great AI Reckoning
The central question facing the AI industry is no longer whether artificial intelligence can do impressive things.
It clearly can.
The question is whether the prevailing approach can continue delivering economically meaningful improvements at a rate sufficient to justify the extraordinary resources being committed to it.
The evidence supporting the wall thesis is increasingly difficult to dismiss.
Scaling requires greater expenditure for incremental gains. Memory systems struggle to keep pace with computational capacity. High-quality human-generated training data is finite. Synthetic data introduces its own limitations. Reasoning models exhibit striking failures at higher complexity. Enterprise deployments frequently struggle to deliver measurable financial returns. AI-assisted workflows can transfer effort from creation to verification rather than eliminating it. Energy and infrastructure demands are growing rapidly.
These are not isolated inconveniences.
They are interconnected constraints on the same technological and economic model.
The industry has spent years celebrating what artificial intelligence can generate. It is now being forced to confront what artificial intelligence can reliably deliver.
And there is a profound difference between the two.
A convincing answer is not necessarily a correct answer. A successful demonstration is not necessarily a viable product. A benchmark improvement is not necessarily an economic breakthrough. A faster processor is not necessarily a faster system. A larger model is not necessarily a more intelligent model.
Most importantly, a larger investment is not necessarily a better investment.
The history of technology suggests that the most dangerous moment in an industry's development often arrives when its early successes are mistaken for proof that its underlying growth model can continue indefinitely.
The first generation of large language models delivered extraordinary breakthroughs. Those breakthroughs created enormous expectations. Those expectations attracted unprecedented capital. And that capital is now financing an infrastructure expansion whose economic justification depends on the assumption that the breakthroughs will keep coming.
But the scientific and engineering constraints are becoming more visible just as the financial commitments become more enormous.
This is how technological optimism turns into economic vulnerability.
The AI wall does not mean that artificial intelligence will disappear. It does not mean that existing systems have no value. It means that the industry's dominant formula for progress is encountering limits that cannot simply be purchased away.
The next revolution in artificial intelligence, if it comes, will have to be earned through genuine scientific and architectural breakthroughs rather than assumed to follow automatically from exponential expenditure.
For years, the industry has behaved as though the road to unlimited intelligence was a straight highway, and the only question was how much money could be spent accelerating along it.
The road is no longer straight.
The limits of data, computation, memory, reasoning, energy, and economics are converging.
Artificial intelligence has hit a wall. And the most expensive mistake the industry can make is to assume that building a bigger machine will make the wall disappear.
References
The article is supported by the following research and reporting:
- Apple — The Illusion of Thinking (2025). Experimental evidence of reasoning-model accuracy collapse beyond certain complexity thresholds.
- METR — Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. Randomized trial documenting a 19% slowdown in the studied developer tasks.
- MIT NANDA — The GenAI Divide: State of AI in Business 2025. Research on the limited measurable financial returns from integrated enterprise AI pilots.
- International Energy Agency — Energy and AI (2025). Analysis of electricity consumption, infrastructure requirements, and projected data-center growth.
- Reuters — Ilya Sutskever on the Limits of Pre-training (2024). Reporting on the limitations of the pre-training paradigm.
Additional foundational research discussed includes Scaling Laws for Neural Language Models (Kaplan et al., 2020), Training Compute-Optimal Large Language Models (Hoffmann et al., 2022), Will We Run Out of Data? (Villalobos et al., 2024), AI Models Collapse When Trained on Recursively Generated Data (Shumailov et al., 2024), and Computing's Energy Problem (Horowitz, 2014).