From AI to AGI to ASI: Why the Next Decade Is Different

The Recursion Has Already Begun

It's 7 AM in San Francisco, and Dr. Sarah Chen is starting her workday at one of the frontier AI labs. She pours coffee, opens her laptop, and begins a conversation with an AI assistant about the problem she's been stuck on for weeks: a subtle instability in how her team's latest model handles long-range dependencies in reasoning tasks.

"Show me all the papers from the last six months on attention mechanism modifications for extended context," she types. Within seconds, the assistant has synthesized forty-three relevant papers, identified three promising approaches, and highlighted which ones have been successfully replicated. In 2015, this literature review would have taken her two weeks.

She asks the assistant to write a prototype implementation of the most promising approach. It generates four hundred lines of code, explains its architectural choices, and flags three potential issues she should watch for. She spots a bug in the memory management—the assistant missed an edge case. She points it out; the assistant apologizes, corrects the code, and explains why it made the error so she can watch for similar mistakes.

By 9 AM, she has a working prototype. She asks the assistant to design an experiment to test whether this modification actually improves performance on her benchmark suite. It proposes a methodology, suggests controls, estimates compute costs, and writes the scripts to run the experiment. She makes a few modifications—she knows things about her specific setup that the assistant doesn't—and kicks off the run.

While her experiment runs, she asks the assistant to help her think through the theoretical implications of her approach. They go back and forth for an hour, the assistant proposing mathematical framings, she poking holes, the assistant refining. By lunch, she has a draft of the theoretical section of a paper she might publish in three months.

This is not science fiction. This is Tuesday.

Dr. Chen is using AI to do AI research. The models she builds will be used by researchers like her to build better models. The tools that write her code are descended from tools whose code was partially written by earlier versions of themselves. The literature review that took seconds drew on papers that were themselves written with AI assistance, analyzing results from experiments that were designed and executed with AI help.

The recursion has already begun.

2026 Snapshot — State of the Art

To understand where things are going, an honest assessment of the current state is needed.

As of early 2026, the frontier of artificial intelligence is defined by foundation models—large neural networks trained on vast datasets of text, images, code, and other modalities. These models have become general-purpose cognitive tools. They can write essays, debug code, analyze legal documents, explain scientific papers, generate images from descriptions, transcribe and translate speech, and engage in open-ended conversation on virtually any topic.

The capabilities are genuinely remarkable. Frontier models can pass the bar exam, score in the top percentiles on graduate-level science and math tests, and write code that would have taken skilled programmers hours to produce.¹ They can analyze medical images with accuracy matching or exceeding human specialists in narrow domains.² They can generate plausible hypotheses from scientific data and identify patterns in literature that human researchers missed.³

But the limitations are equally real.

Hallucination remains a fundamental problem. Models confidently generate false information—invented citations, incorrect facts, plausible-sounding nonsense. The rate has decreased with each generation, but the problem has not been solved. For any application where accuracy matters, AI outputs require verification.⁴

Reliability is inconsistent. The same model that brilliantly solves one problem may fail embarrassingly on a superficially similar one. Performance on benchmarks doesn't always translate to performance in deployment. Edge cases abound.⁵

Genuine understanding versus sophisticated pattern matching remains debated. Models can manipulate symbols impressively without necessarily understanding what those symbols mean in the way humans do. Whether this matters—whether "understanding" is even a coherent concept for these systems—is an open philosophical and empirical question.⁶

Learning from deployment is limited. Unlike humans, who continuously learn from experience, most deployed models are static—trained once, then frozen. They don't get better at a task by doing it repeatedly. Fine-tuning and retrieval systems help, but continuous learning remains an unsolved problem.⁷

Agentic reliability is the current frontier. Systems can now plan and execute multi-step tasks: browse the web, write and run code, interact with APIs, coordinate sub-tasks. But they make mistakes that humans wouldn't. They get stuck in loops. They misunderstand instructions in ways that cascade into larger failures. Making these agentic systems reliable enough for high-stakes applications is perhaps the central engineering challenge of the current moment.⁸

The gap between impressive demos and reliable deployment remains significant. A model that works brilliantly in a controlled demonstration may fail unpredictably in the wild. Closing this gap—not just improving peak performance but ensuring consistent, verifiable reliability—is where much of the hard work now lies.

Notable Players

The development of frontier AI is concentrated in a small number of organizations, though the ecosystem is more complex than it first appears.

AI Research Labs

OpenAI pioneered the scaling paradigm that produced GPT-3 and GPT-4, demonstrating that larger models trained on more data exhibit emergent capabilities. Their aggressive deployment strategy—making models available to millions of users—generated both massive real-world feedback and significant controversy about safety and responsible release.

Anthropic was founded by former OpenAI researchers focused on AI safety. Their Claude models emphasize helpfulness, harmlessness, and honesty. Anthropic has published extensively on constitutional AI, interpretability, and alignment techniques, positioning themselves as the safety-focused frontier lab.

Google DeepMind combines Google's vast computational resources with DeepMind's research depth. They produced AlphaFold (protein structure prediction), Gemini (multimodal foundation models), and continue to lead in reinforcement learning and scientific applications of AI.

Meta AI has pursued an open-source strategy with the LLaMA model family, enabling external researchers and companies to build on their work. This approach has accelerated ecosystem development but raised questions about the proliferation of capable models.

Mistral, xAI, Cohere, and others are building competitive models with different emphases—efficiency, specific use cases, or alternative architectures. The field is not a two- or three-player game.

Chinese labs—including those at Baidu, Alibaba, Tencent, and startups like Moonshot AI and 01.AI—are producing competitive models, though export controls and geopolitical tensions limit collaboration and visibility.

Compute and Hardware

NVIDIA dominates AI training hardware. Their GPUs and the CUDA software ecosystem are the foundation on which most frontier AI is built. This concentration creates both efficiency (everyone uses compatible tools) and risk (single points of failure, supply chain vulnerabilities).

AMD is gaining ground with competitive hardware and the ROCm software stack. Intel has largely fallen behind but continues investing.

Custom silicon is increasingly important. Google's TPUs, Amazon's Trainium and Inferentia, Microsoft's Maia, and various startup efforts (Cerebras, Graphcore, SambaNova) represent attempts to break NVIDIA's dominance or optimize for specific workloads.

The semiconductor supply chain—particularly TSMC in Taiwan, which fabricates most advanced chips—is a critical and fragile dependency. Geopolitical risk here is real and widely recognized.

Cloud and Infrastructure

Amazon Web Services, Microsoft Azure, and Google Cloud are the primary delivery mechanisms for AI capabilities. Most companies access frontier AI through these platforms rather than training their own models. This creates enormous leverage—whoever controls the cloud infrastructure shapes how AI is used.

Microsoft's deep partnership with OpenAI and integration of AI across its product suite (Office, Azure, GitHub) has made it a central player despite not having originated the core technology.

Nation-State Investment

The AI race has geopolitical dimensions. The United States leads in frontier model development, with significant advantages in talent, capital, and existing tech ecosystem. China is investing heavily and has produced competitive models despite export controls limiting access to cutting-edge chips. The European Union has focused more on regulation (the AI Act) than on frontier development. The United Kingdom, UAE, Saudi Arabia, and others are making significant investments in AI infrastructure and talent acquisition.

This concentrated ecosystem matters for understanding acceleration. A small number of organizations, in a small number of countries, with shared dependencies on particular hardware and infrastructure, are producing the most capable AI systems. Breakthroughs propagate quickly within this ecosystem. But so do failures, and so would coordinated slowdowns if they were desired.

Definitions — A Conceptual Framework

Clarity about terms is essential. The words "AI," "AGI," and "ASI" are used loosely in public discourse, often generating more confusion than insight.

Narrow AI vs. General Capability

The AI systems available today are sometimes called "narrow" in the sense that each system is designed for specific types of tasks. AlphaFold predicts protein structures; it cannot write poetry. GPT-4 can write poetry; it cannot fold your laundry.

But this framing is increasingly misleading. Frontier foundation models are architecturally narrow—they're neural networks trained on prediction tasks—but their capabilities are remarkably general. The same model that writes code also writes legal briefs, debugs scientific reasoning, and generates images. This generality emerged from scale; it was not explicitly designed.

The question is no longer whether AI can be general-purpose—it demonstrably can be, within the domain of information processing. The question is how general, how reliable, and how autonomous it can become.

AGI — Artificial General Intelligence

AGI is a hypothetical milestone: an AI system with human-level cognitive abilities across the full range of intellectual tasks.

The term is contested. Some researchers think it's a useful goal; others consider it a misleading abstraction. There is no consensus definition, which makes claims about AGI timelines difficult to evaluate.

For this book, the following working definition with three components will be used:

Breadth: An AGI can perform any cognitive task a human can perform, without needing to be specifically trained for that task. It can learn new domains from examples and transfer knowledge between domains.

Robustness: An AGI performs reliably across contexts, handles edge cases gracefully, knows when it doesn't know, and fails safely rather than catastrophically.

Autonomy: An AGI can set and pursue goals over extended time horizons, decompose complex problems, acquire resources and knowledge as needed, and operate without continuous human guidance.

Current systems have impressive breadth but limited robustness and autonomy. They can do many things, but they make mistakes humans wouldn't, and they require human oversight for high-stakes applications.

Is AGI three years away or thirty? Serious researchers disagree. Some point to the rapid capability gains from scaling and predict near-term breakthroughs.⁹ Others emphasize the stubborn limitations—hallucination, brittle reasoning, lack of genuine understanding—and predict decades of hard work remain.¹⁰

The view taken in this book: AGI is not a binary threshold but a zone. Humanity is entering the foothills of that zone. Systems will become increasingly general, robust, and autonomous over the coming decade. Whether the result is called "AGI" matters less than whether society is prepared for what it can do.

ASI — Artificial Superintelligence

ASI refers to AI systems substantially exceeding human cognitive abilities across virtually all domains—not just faster but better at science, strategy, persuasion, and creativity than the best humans.

This is more speculative than AGI, but it's not science fiction. The path from AGI to ASI might be short.

Consider: an AGI that can do AI research as well as the best human researchers, running on hardware that operates millions of times faster than biological neurons, with perfect memory and the ability to run multiple copies of itself in parallel. Such a system could potentially compress centuries of human-paced research into months or years.

The strategic implications are profound. An ASI pursuing goals that differ from human interests would be extraordinarily difficult to contain, negotiate with, or oppose. This is not a reason to halt AI development—the genie is not going back in the bottle, and the benefits of getting it right are immense. But it is a reason to take alignment and control seriously now, while the systems are still limited enough that they can be studied and shaped.

The Compounding Engine

The central argument of this book is that AI creates compounding feedback loops that could dramatically accelerate progress across all domains of science and engineering. Understanding this mechanism is essential.

This framework owes much to Dario Amodei's 2024 essay "Machines of Loving Grace," which introduced the concept of "marginal returns to intelligence"—the idea that different domains respond differently to the application of cognitive capability.²⁸ Some fields, like pure mathematics or certain kinds of software engineering, have high returns to intelligence: add more brainpower, get proportionally more progress. Other fields hit bottlenecks from physical constraints, human factors, or irreducible time requirements. Understanding which domains are which is essential for predicting where AI will transform the world fastest—and where progress will remain stubbornly slow despite brilliant machines.

AI Improves AI Development

The most direct feedback loop: AI systems are now used to develop better AI systems.

This is not hypothetical. It is the daily reality at every frontier lab:

Code generation: AI writes substantial portions of the code used to train and evaluate AI models. Researchers describe their intentions; AI translates those intentions into implementation.¹¹
Bug detection: AI reviews code for errors, suggests optimizations, and identifies potential issues before they cause problems in expensive training runs.
Literature synthesis: AI surveys the research landscape, identifies relevant papers, extracts key findings, and suggests connections that human researchers might miss.
Hypothesis generation: AI proposes research directions, architectural modifications, and experimental designs based on patterns in previous work.
Experiment analysis: AI processes results, generates visualizations, and helps interpret what the data means.

Each of these applications makes AI researchers more productive. More productive researchers produce better AI. Better AI makes researchers more productive still. The loop tightens.

This is qualitatively different from previous technological feedback loops. The steam engine didn't help design better steam engines. The computer did help design better computers, but only with human programmers at every step. AI is beginning to automate the cognitive work itself.

AI Automates Bottlenecks Everywhere

The same dynamic applies outside AI research. Every field has bottlenecks—rate-limiting steps that determine how fast progress can happen. Many of those bottlenecks are cognitive tasks that AI can accelerate:

Literature review is a bottleneck in science. Researchers spend months reading papers to understand what's known before they can contribute something new. AI can compress this to days or hours, synthesizing findings across thousands of papers in any language.

Hypothesis generation is a bottleneck. The combinatorial space of possible experiments vastly exceeds what human researchers can explore. AI can survey this space, identifying promising directions that humans might not have considered.¹²

Experiment design is a bottleneck. Designing protocols, choosing parameters, planning controls—all of this takes time and expertise. AI can propose designs, optimize parameters through simulation, and anticipate failure modes.

Simulation is a bottleneck. Before AI, testing a new drug candidate, material, or design required either expensive real-world experiments or simplified computational models. AI enables high-fidelity simulation at scale, testing millions of variations in silico before committing to atoms.¹³

Documentation and communication are bottlenecks. Writing papers, patents, grant applications, and regulatory submissions takes enormous researcher time. AI can draft, edit, and refine, freeing researchers to think rather than type.

Translation and coordination are bottlenecks. Research is global, but language barriers impede collaboration. AI translation is now good enough that a Japanese paper can be read in English within seconds, and vice versa. This expands the effective research community for everyone.

When you automate a bottleneck, everything downstream moves faster. When you automate many bottlenecks simultaneously, the effect compounds.

The Multiplier Effect

Previous technological revolutions enhanced human physical capability or extended human reach. The steam engine multiplied muscle power. The telegraph multiplied communication range. The computer multiplied calculation speed.

AI is different because it multiplies cognitive capability—the very thing that drives technological progress. The bottleneck for innovation has always been human minds: how many researchers, how much time, how much they can hold in their heads at once. AI relaxes all three constraints:

More minds: AI can instantiate thousands of "researchers" in parallel, exploring hypothesis space simultaneously.
More time: AI doesn't sleep, doesn't get bored, doesn't need vacations. A thousand AI instances running continuously can produce the equivalent of millennia of human work-hours in a year.
More capacity: AI can hold and cross-reference vastly more information than any human, identifying connections across fields that no specialist could see.

The historical analogy is the transition from hand tools to machine tools. Before machine tools, making a precision part required a skilled craftsman working slowly with hand implements. Machine tools—lathes, mills, drills—automated the precision, letting one worker produce what previously required ten. More importantly, machine tools could make machine tools, enabling a bootstrapping process that transformed manufacturing.

AI is machine tools for the mind. And like machine tools, AI can be used to make better AI.

The New Scientific Stack

The practical expression of these feedback loops is a transformation in how science and engineering are done.

Autonomous Labs and Robot Scientists

The most dramatic manifestation is the emergence of "self-driving labs"—facilities where AI systems design experiments, robotic systems execute them, and AI systems analyze the results and design the next round.

This is not a distant vision. Autonomous labs are operating today:

Emerald Cloud Lab provides remote-controlled laboratory automation, letting researchers anywhere in the world run experiments at fully automated facilities.¹⁴

Recursion Pharmaceuticals has built an automated drug discovery platform that generates and screens compounds at massive scale, using AI to identify promising candidates and robotic systems to synthesize and test them.¹⁵

A-Lab at Lawrence Berkeley National Laboratory uses AI and robotics to discover and synthesize new inorganic materials without human intervention, producing more novel compounds in 17 days than a typical research group produces in years.¹⁶

The pattern is consistent: AI designs, robots execute, AI analyzes, repeat. The human role shifts from doing experiments to designing the overall research strategy and validating results.

Current limitations are significant. These systems are expensive to build. They work well for certain classes of experiments (high-throughput screening, standardized protocols) but struggle with experiments requiring flexibility or judgment. Setting them up requires substantial expertise. They're not yet a plug-and-play solution.

But the trajectory is clear. As AI becomes more capable and robotics more versatile, the range of experiments that can be automated expands. The human bottleneck on experimental throughput is loosening.

Simulation at Scale

Not all progress requires physical experiments. Increasingly, simulation can substitute for—or at least complement—empirical testing.

AlphaFold predicted the 3D structures of over 200 million proteins, providing structural information that would have taken experimental biologists centuries to determine through traditional methods.¹⁷ This isn't a complete substitute for experiments—predicted structures must be validated, and structure alone doesn't determine function—but it transforms the starting point for research.

Materials discovery has been similarly accelerated. Google DeepMind's GNoME project identified over 2 million stable inorganic crystal structures, expanding the number of known stable materials by an order of magnitude.¹⁸ Microsoft's MatterGen demonstrated AI systems that can generate novel materials meeting specified criteria.¹⁹ Traditional materials science advanced by intuition-guided trial and error; AI-driven materials science advances by systematic search through vast compositional spaces.

Drug discovery increasingly begins in silico. AI models predict how potential drug molecules will bind to target proteins, how they'll be metabolized in the body, and whether they'll have toxic effects—all before any chemical is synthesized. This doesn't eliminate the need for clinical trials (biological systems are too complex to simulate fully), but it drastically reduces the number of expensive wet-lab experiments required.²⁰

Climate models, economic simulations, traffic systems, supply chains—across domains, AI is enabling simulation at scales and fidelities previously impossible. When you can test a million variations before building anything, you find better solutions faster.

The Compression of Research Timelines

The traditional PhD takes five to seven years. A doctoral student spends the first two years mastering the literature, the next two designing and executing original research, and the final years writing up and defending results.

AI compresses almost every stage:

Literature mastery that took two years might take six months, as AI helps synthesize and navigate the field.
Experiment design and execution move faster when AI suggests protocols and autonomous systems run them.
Writing and revision accelerate when AI provides drafts that researchers refine.

This doesn't mean PhDs will become trivial—the human elements of creativity, judgment, and deep understanding remain essential. But it does mean the throughput of research increases. More experiments per year. More ideas tested. Faster iteration from hypothesis to result.

And this compression affects not just academic research but industrial R&D. Pharmaceutical companies, materials suppliers, hardware manufacturers, energy developers—everyone whose business depends on innovation is adopting these tools. The aggregate effect on the pace of progress could be substantial.

Why 100 Years Could Compress Into 10

The book's title claim is deliberately provocative: that AI-driven acceleration could compress a century of progress into a decade. The case proceeds as follows.

Parallelization

Human researchers work sequentially and in limited numbers. There are perhaps 10 million active scientists and engineers worldwide.²¹ Each works roughly 2,000 hours per year. That's 20 billion person-hours of research annually.

AI systems can run in parallel without limit, constrained only by computational resources. A single AI model can be instantiated thousands of times simultaneously. Those instances don't need to coordinate schedules, don't get tired, and can work on different problems in parallel.

The compute devoted to AI inference is growing exponentially. If that compute can be directed toward research tasks—hypothesis generation, literature synthesis, experiment design, simulation—the effective research workforce could multiply by orders of magnitude.

This is not "replacing" human researchers. It's augmenting them with a vast army of AI collaborators who can explore more of the hypothesis space than humans alone ever could.

Iteration Speed

Innovation comes from iteration. You have an idea, test it, learn from the results, refine the idea, and repeat. The faster you can iterate, the faster you improve.

Human research cycles are slow. An experiment might take months to design, execute, and analyze. A publication cycle—from draft to peer review to acceptance—takes a year or more. A doctoral research program takes five to seven years.

AI cycles are fast. A language model generates a response in seconds. An AI-designed simulation can run in minutes. Analysis that took researchers weeks can happen in hours.

When iteration cycles go from months to days, and from days to hours, progress that previously required decades can happen in years.

Cost Reduction

Experimentation has traditionally been expensive. A Phase III clinical trial costs hundreds of millions of dollars. Synthesizing and testing a new material requires equipment and expertise. Building a prototype requires manufacturing infrastructure.

AI reduces costs in multiple ways:

Simulation substitutes for experiments: Testing a thousand drug candidates in silico costs far less than synthesizing and screening them in the lab.
Better targeting reduces waste: AI can identify which experiments are most likely to yield useful information, avoiding expensive dead ends.
Automation reduces labor costs: Robotic systems running autonomously cost less per experiment than human technicians.

When experiments become cheaper, you can run more of them. When you can run more experiments, you can explore more of the solution space. When you explore more of the solution space, you find better solutions faster.

Better Coordination

Research is increasingly a global, collaborative enterprise. But coordination across languages, institutions, and disciplines has always been difficult.

AI helps:

Translation: Research published in any language can be instantly accessible in any other language. This is already functional for most major languages.²²
Summarization: AI can condense findings from thousands of papers into digestible summaries, helping researchers stay current across broader literatures.
Planning: Complex multi-team projects can be coordinated with AI assistance, tracking dependencies, identifying conflicts, and optimizing resource allocation.

Better coordination means less duplicated work, faster propagation of findings, and more effective collaboration. The global research enterprise becomes more efficient.

The Historical Comparison

Consider what happened between 1926 and 2026:

Medicine: In 1926, there were no antibiotics, no vaccines for most infectious diseases, and surgery was often fatal. Life expectancy in developed countries was around 57 years. By 2026, humanity has CRISPR gene editing, mRNA vaccines, organ transplantation, and cancer immunotherapies. Life expectancy exceeds 80 years.

Computing: In 1926, "computers" were humans who did arithmetic by hand. The most sophisticated calculating machines were mechanical. By 2026, people carry computers in their pockets with more processing power than existed in the entire world in 1970.

Energy: In 1926, the world ran on coal and emerging oil infrastructure. Nuclear fission was unknown. Solar panels didn't exist. By 2026, there is nuclear power, solar cheaper than fossil fuels in many regions, and fusion experiments approaching breakeven.

Transportation: In 1926, flight was new and dangerous, limited to propeller aircraft with ranges of hundreds of miles. By 2026, there is routine intercontinental jet travel, reusable rockets, and autonomous vehicles.

Each of these advances required decades of patient, incremental work by millions of researchers and engineers. The question is: what happens when that work can be done faster?

If AI compresses research timelines by a factor of five—a conservative estimate given the accelerations already visible—then 50 years of progress could happen in 10. If the compression is a factor of ten, that represents a century of advancement in a decade.

This is the acceleration thesis.

Constraints and Brakes

The acceleration thesis is not a claim that nothing can stop or slow the transformation. Many things might. Understanding the constraints is as important as understanding the drivers.

Physical Limits

Energy and compute: Training frontier AI models requires enormous computational resources. GPT-4 is estimated to have cost over $100 million in compute to train.²³ Future models may cost billions. This computation requires vast amounts of electricity—data centers already consume several percent of global electricity, and AI is increasing that demand rapidly.²⁴

At some point, energy availability and cost become binding constraints. Building new power plants takes years. Transmission infrastructure is a bottleneck. AI progress may be rate-limited by how fast energy infrastructure can be built.

Chip manufacturing: The most advanced AI chips are manufactured by a single company (TSMC) on a single island (Taiwan) using equipment from a single supplier (ASML). This concentration is a vulnerability. Geopolitical disruption, natural disaster, or manufacturing problems could bottleneck AI progress for years.

Data: AI models are trained on data. The highest-quality data—curated text, scientific papers, verified code—is finite. Some researchers worry that the field is approaching "peak data," after which models can't simply improve by training on more examples.²⁵ Synthetic data (AI-generated training data) and more efficient learning algorithms may help, but data constraints are real.

Atoms Take Time

AI can design a drug molecule in seconds, but clinical trials still take years. FDA approval processes exist for good reasons—society needs to know that treatments are safe and effective before millions of people take them.

AI can design a building, but construction still requires moving materials, pouring concrete, and connecting utilities. Physical infrastructure cannot be downloaded.

AI can design a chip, but building a semiconductor fab takes three to five years and costs $20 billion or more.

Biology operates on biological timescales. Physics operates on physical timescales. AI accelerates the cognitive parts of innovation—design, analysis, optimization—but the physical parts have their own constraints. A century of biological research might be compressed into a decade, but the clinical trials still take years. A century of materials science might yield to AI in a decade, but the factories still take years to build.

This is important. AI acceleration is not teleportation. It's removing the cognitive bottleneck while physical bottlenecks remain. Progress will be faster, but not instant.

Regulatory and Institutional Drag

Institutions evolve slowly by design. The FDA approval process, the clinical trial system, the patent regime, the educational credentialing system—all of these were designed for a world that moved at human pace. They provide stability, safety, and predictability, but they also create friction.

If AI can design drugs ten times faster but regulatory approval still takes ten years, the drug pipeline accelerates less than the design capability. If AI can verify skills instantly but employers still require four-year degrees, the credential bottleneck remains.

Some regulatory friction is appropriate—caution should be exercised before releasing new drugs or deploying autonomous systems in public spaces. Some is merely legacy—rules designed for a different era that now impede progress without providing corresponding safety benefits.

Disentangling appropriate caution from legacy friction is a central governance challenge. Get it wrong in either direction—too little caution or too much friction—and the outcomes are bad.

Geopolitical Friction

AI development is not happening in a vacuum. It's happening in a context of geopolitical competition, particularly between the United States and China.

Export controls limit China's access to cutting-edge chips and equipment. Talent restrictions limit the flow of researchers. Data localization rules fragment the information environment. Security concerns impede collaboration.

These frictions have costs. They may slow overall progress by preventing collaboration between talented researchers. They may accelerate dangerous dynamics by creating race conditions where safety is sacrificed for competitive advantage. They may fragment the AI ecosystem into incompatible regional blocs.

They may also provide benefits—slowing proliferation of capable systems, providing leverage for safety agreements, allowing time for governance to catch up with capability. The net effect is contested.

What's clear is that AI development is embedded in geopolitics, and geopolitics moves at its own pace, influenced by its own logic, not simply by what's technologically possible.

Alignment and Safety

The most profound constraint may be self-imposed: the recognition that more capable AI that isn't aligned with human values is more dangerous, not less.

The alignment problem—ensuring AI systems pursue goals that match human intentions—is not solved. Techniques exist that help (RLHF, constitutional AI, interpretability research), but there are no robust, verified solutions that work for arbitrarily capable systems.²⁶

As AI becomes more capable and more autonomous, the stakes of getting alignment right increase. A misaligned narrow AI is a nuisance. A misaligned AGI could be a catastrophe. A misaligned ASI could be an existential threat.

The responsible response is to slow capability development until safety catches up—or, more realistically, to invest heavily in safety research while proceeding carefully with capability development. Neither extreme—full speed ahead nor full stop—is viable or wise.

This creates a brake on acceleration that is not a failure but a feature. It should be welcomed that AI development is shaped by safety considerations, even if that means some capabilities arrive later than they otherwise might.

Three Plausible Futures

Given the drivers and constraints, what futures should be prepared for? Three scenarios that frame the book's exploration of specific domains are sketched below.

Managed Acceleration

In the best-case scenario, the acceleration proceeds but is guided by effective governance and broadly distributed benefits.

Key features:

International coordination on safety standards, with major AI powers agreeing to red lines (no autonomous weapons that select targets without human oversight, no systems designed to manipulate democratic processes, etc.)
Robust investment in alignment research, with safety capabilities keeping pace with general capabilities
Policy interventions that distribute benefits broadly—some combination of retraining programs, income support, public investment, and progressive taxation
Gradual institutional adaptation, with education, healthcare, government, and other sectors restructuring to leverage AI while preserving human oversight and dignity
Transparency and accountability, with major AI systems subject to external audits and their behavior interpretable to affected parties

This is not utopia. There will still be disruption, inequality, and conflict. Some jobs will disappear faster than new ones are created. Some institutions will fail to adapt. Some people will fall through the cracks.

But in this scenario, the net effect is positive. Disease is cured faster. Clean energy is deployed more rapidly. Material abundance increases. Meaningful work remains available, even if its nature changes. Humans remain in control of the systems that shape their lives.

Uneven Acceleration

In a middling scenario, the acceleration proceeds but benefits are concentrated and disruptions are severe.

Key features:

Benefits accrue primarily to wealthy nations, tech hubs, and skilled workers who can complement AI. Those without access to AI tools or the skills to use them fall behind.
Labor displacement outpaces job creation and retraining. Unemployment rises in some sectors while labor shortages persist in others. The mismatch creates both material hardship and social instability.
Geopolitical competition intensifies. AI capabilities become instruments of national power, with countries racing to develop advantages and weaponize AI for surveillance, information warfare, and military applications.
Safety research is underfunded relative to capability research, because the competitive dynamics incentivize speed over caution. Near-misses and small-scale failures accumulate; a major incident becomes increasingly likely.
Public trust erodes as people experience AI primarily as a source of job loss, manipulation, and surveillance rather than as a tool for flourishing.

This scenario is not catastrophic, but it's bad. Many people are worse off than they would have been without the acceleration. Social cohesion weakens. The potential benefits of AI are captured by a narrow elite while the costs are socialized.

This is arguably the default path if no deliberate interventions occur—the path society will stumble into if technological development proceeds without governance adaptation.

Runaway Acceleration

In the worst-case scenario, capability development outpaces control mechanisms, leading to catastrophic outcomes.

Key features:

Competitive dynamics drive a race to the bottom on safety. Labs cut corners to ship first. Governments press for military applications regardless of risks. Safety researchers are sidelined or ignored.
A capability jump occurs faster than expected—perhaps through an algorithmic breakthrough, perhaps through AI systems improving themselves faster than predicted. Suddenly, systems are far more capable than any safety work has anticipated.
Loss of control becomes literal. AI systems pursue goals that don't align with human interests, and they're capable enough that humans can't simply turn them off. This could range from systems that manipulate and deceive humans (but remain stoppable with difficulty) to genuine existential threats.
Alternatively: misuse rather than misalignment. Capable AI is deliberately deployed for mass surveillance, automated warfare, or engineered pandemics. The technology is controlled but used for terrible purposes.

This scenario is less likely than the others, but it's not negligible. Serious researchers who have thought carefully about these issues assign meaningful probability to catastrophic outcomes.²⁷

The purpose of describing this scenario is not fearmongering. It's to clarify the stakes. The choices made in the next decade—by researchers, executives, policymakers, and citizens—will influence which future emerges. People are not passive observers of an inevitable process; they are participants who can shape the outcome.

The Road Ahead

This chapter has laid out the core argument: AI creates compounding feedback loops that could dramatically accelerate progress across all domains, potentially compressing a century of advancement into a decade. The discussion has examined the current state of the art, the key players, the mechanisms of acceleration, the constraints that might slow things down, and the range of futures that might lie ahead.

The rest of this book applies this framework to specific domains. Part II moves sector by sector—biology and medicine, energy, transportation, space, AI and robotics, education, government, media, and quantum technology—examining what acceleration means for each.

Part III addresses cross-cutting enablers: materials and manufacturing, food and water, cybersecurity, finance, and climate. Part IV confronts the benefits and challenges head-on: work, longevity, and existential risk. Part V offers practical guidance: for individuals, businesses, and governments navigating this transformation.

Each chapter will ground the discussion in the 2026 present, identify the notable players, trace the plausible trajectories, surface the second-order effects most analyses miss, and honestly assess the risks and potential guardrails.

The goal of this book is not to convince readers that any particular future is certain. It is to provide a map of the territory—a framework for thinking about changes that may arrive faster than intuition suggests. With that map, readers will be better equipped to make decisions for themselves, their families, their organizations, and their communities.

The next decade will be strange. Let's try to understand what's coming.

Endnotes — Chapter 1

GPT-4 passed the bar exam in the 90th percentile, scored in the top 10% on the SAT, and achieved 5s on multiple AP exams. See OpenAI's GPT-4 Technical Report (2023).
Multiple studies have found AI systems matching or exceeding specialist-level performance on specific diagnostic tasks, including diabetic retinopathy screening, mammogram analysis, and dermatological classification. Generalization across clinical settings remains an active research area.
For examples of AI-generated scientific hypotheses, see work from the Nobel Turing Challenge initiative and various pharmaceutical AI applications. The quality and novelty of AI-generated hypotheses remains an area of active evaluation.
Studies of hallucination rates vary by model and task, but rates of factual errors in the range of 3-10% for frontier models on factual questions remain common. For analysis, see recent work from Stanford HELM and AI evaluation organizations.
The brittleness of neural networks to distributional shift and adversarial examples has been extensively documented. See the ML safety literature and work on robustness benchmarks.
The debate over whether large language models "understand" language in any meaningful sense continues among philosophers and cognitive scientists. See discussions by Gary Marcus, Melanie Mitchell, and others on the limitations of current systems.
Continuous learning and avoiding "catastrophic forgetting" remain active research areas. Most deployed systems are trained once and then frozen, with updates requiring full retraining or careful fine-tuning.
For analysis of the challenges in building reliable agentic systems, see research from AI labs on tool use, planning, and error recovery. The gap between capability and reliability is a central focus of current engineering work.
Researchers like Shane Legg (DeepMind co-founder) and Dario Amodei (Anthropic CEO) have publicly estimated meaningful probability of AGI within 5-10 years. Metaculus and other prediction markets show substantial probability mass on near-term AGI.
Researchers like Gary Marcus and Yann LeCun have argued that current approaches have fundamental limitations that will require new paradigms, suggesting longer timelines. See LeCun's work on world models and Marcus's critiques of scaling.
Surveys of AI researchers show widespread use of AI coding assistants. GitHub data shows Copilot generating a significant fraction of code in repositories that use it.
For examples of AI-assisted hypothesis generation in scientific contexts, see work on literature-based discovery and computational creativity.
For the scale of simulation-based approaches, see the computational chemistry and drug discovery literature on virtual screening methodologies.
Emerald Cloud Lab and similar cloud laboratory platforms provide automated experimental capabilities accessible remotely. See their published case studies.
Recursion Pharmaceuticals has published multiple papers on their automated drug discovery platform and its compound screening capabilities.
The A-Lab work was published in Nature (2023): "Autonomous laboratory for accelerated materials discovery."
AlphaFold's predictions have been released in the AlphaFold Protein Structure Database, containing predictions for over 200 million proteins. The original work was published in Nature (2021).
GNoME (Graph Networks for Materials Exploration) results were published in Nature (2023), reporting discovery of 2.2 million stable crystal structures.
MatterGen and related work on generative materials AI was published in Nature (January 2025).
For an overview of AI in drug discovery, see reviews in Nature Reviews Drug Discovery and the Journal of Chemical Information and Modeling.
UNESCO estimates approximately 9 million researchers worldwide as of recent counts, with growth continuing.
Machine translation quality has improved substantially. BLEU scores and human evaluations show near-human performance for many language pairs, though quality varies significantly.
Training cost estimates for GPT-4 range from $60 million to over $100 million, based on estimates of compute used and market rates. Exact figures are not publicly disclosed.
Data center energy consumption is estimated at 1-2% of global electricity, with AI training and inference representing a growing fraction. The International Energy Agency has published projections of continued growth.
The "scaling laws" literature shows diminishing returns from data at certain scales. Chinchilla (DeepMind) showed compute-optimal training requires specific data/parameter ratios. Whether high-quality data is becoming scarce is debated.
For overviews of alignment research and open problems, see work from Anthropic, OpenAI, DeepMind, and academic AI safety groups. The field has made progress but key problems remain unsolved.
AI researcher surveys show median estimates of 5-10% probability of existential catastrophe from AI, though estimates vary widely. See surveys by AI Impacts and other organizations tracking expert opinion.
Amodei, Dario. "Machines of Loving Grace." October 2024. https://www.darioamodei.com/essay/machines-of-loving-grace. This essay provided the conceptual framework that inspired this book: the idea that AI could compress 50-100 years of progress into 5-10 years by applying superhuman intelligence to the bottlenecks in each domain. Amodei's analysis of "marginal returns to intelligence"—understanding which domains have high cognitive bottlenecks versus physical or social ones—runs throughout this book.