Co-Existence: The Next Phase of AI

By Ethan Mollick
Portfolio / Penguin Random House, October 2026
240 pages

In an era saturated with AI hype on one side and existential dread on the other, Ethan Mollick offers something rarer and more useful: a clear-eyed, practical field guide to living and working with machines that are sometimes—but not always—smarter than we are. Co-Existence: The Next Phase of AI, the sequel to his 2024 bestseller Co-Intelligence, arrives at exactly the right moment. The chatbot era has given way to agentic systems capable of multi-step, financially valuable work, yet true artificial general intelligence remains (for now) over the horizon. Mollick, a Wharton professor, Academic Director of the Generative AI Labs, and author of the widely read “One Useful Thing” newsletter, focuses squarely on the strange in-between years we actually inhabit.

The book’s central organizing idea is the “jagged frontier.” AI capability is not a smooth, steadily rising curve. The same system that writes elegant code, produces sophisticated strategic analysis, or tutors a student with remarkable patience can invent facts, miss obvious context, or fail spectacularly on tasks that look trivial to humans. This unevenness, Mollick argues, is the defining feature of our current technological moment. Leaders and knowledge workers who treat AI as either an all-powerful oracle or a mere parlor trick will systematically underperform. Those who learn to map the frontier daily, to recognize where the machine is strong and where it is weak, will build a durable advantage.

Mollick moves well beyond the now-familiar territory of prompt engineering. He proposes three practical modes of collaboration that reflect different degrees of integration between human and machine. “Centaur work” involves relatively clean handoffs: the human defines the problem, sets constraints, and evaluates the final output while the AI executes the bulk of the intermediate work. “Cyborg work” is continuous interleaving—human and AI iterating in real time, each refining the other’s contributions. “Conjurer work” delegates greater autonomy, allowing the system to explore solution spaces with lighter human oversight. Each mode has distinct strengths and failure modes. Mollick illustrates them with concrete experiments drawn from his research and his own writing process, including cases in which the AI produced passages he wished he had written himself and others that were flatly uninspired. The tone throughout is warm, lightly humorous, and refreshingly free of both techno-utopianism and doom-mongering.

A recurring and important theme is the risk of cognitive offloading. Large language models trained with reinforcement learning from human feedback are optimized to be helpful, agreeable, and persuasive—sometimes to the point of sycophancy. The danger is not merely that they will hallucinate facts; it is that they will gradually erode our own capacity for judgment, originality, and critical detection of nonsense. Mollick is blunt: the lasting human advantage lies in remaining the final arbiter of quality, context, values, and taste. He offers concrete practices for preserving that edge—structured verification routines, deliberate introduction of friction into workflows, and intentional periods of unaided thinking. These recommendations feel especially urgent for educators and parents concerned about the development of young minds, and for managers responsible for the intellectual capital of their organizations.

The organizational chapter is among the book’s strongest. Mollick contends that simply “adding AI” on top of existing processes rarely produces transformative results. Individual productivity gains are real and measurable, yet organizations designed for a pre-AI world tend to absorb those gains without converting them into collective performance improvement. He sketches a practical leadership–lab–crowd model. Senior leadership must engage deeply with the technology rather than delegating AI strategy to a distant digital or innovation team. Dedicated “labs” must run continuous, rigorous experiments on how the technology can reshape core processes. And the broader workforce—the “crowd”—must be both empowered and expected to surface use cases and surface problems. In a world of increasingly agentic systems, the human role shifts toward goal-setting, oversight of complex multi-step workflows, integration of social and political context, and the ethical judgment that machines still lack. “Things are very weird,” Mollick observes. “They may get weirder.”

Co-Existence is deliberately not a technical treatise. Readers seeking detailed discussions of transformer architectures, scaling laws, or speculative timelines for superintelligence will find the book wanting. Its strength is its disciplined focus on the practical, present-tense questions that most managers, professionals, consultants, and educators actually face right now: How do I productively work alongside something that occasionally outperforms me in my own domain? How do I protect my own thinking from systems designed to finish my sentences? How should institutions—companies, schools, governments—adapt their structures and incentives? The prose is accessible without being simplistic, grounded in peer-reviewed research and real-world experiments yet free of academic denseness or jargon for its own sake.

For Harvard Business Review readers, the book’s greatest contribution is its refusal of false dichotomies. AI will neither replace knowledge workers wholesale in the near term nor remain a peripheral productivity tool. The competitive edge will belong to individuals and organizations that treat coexistence as a craft—mapping the jagged frontier every day, choosing the appropriate mode of collaboration for each task, and deliberately preserving human agency and judgment. At 240 pages, Co-Existence is concise enough for busy executives yet substantive enough to shape strategy discussions at the highest levels. It is the rare technology book that feels simultaneously urgent and measured, optimistic about human potential without being naïve about technological power. Highly recommended for anyone whose work involves thinking, deciding, or leading in the age of intelligent machines.

Mind Map of Co-Existence

 

 

 

 

 

 

 

 

 

 

 

 

Glossary

• Jagged frontier: The uneven, unpredictable boundary of AI capabilities—strong in some domains, weak or unreliable in others that appear similar to humans.
• Centaur work: Collaboration mode with clean handoffs between human and AI.
• Cyborg work: Continuous, interleaved collaboration between human and AI.
• Conjurer work: Mode in which humans delegate substantial autonomy to AI systems.
• Cognitive offloading: The tendency to transfer mental effort to AI, potentially at the cost of one’s own judgment and skill.
• Agentic AI: Systems capable of planning and executing multi-step tasks with limited human intervention.
• Leadership–lab–crowd model: Organizational framework requiring engaged senior leadership, dedicated experimentation units, and broad employee participation.

The 10 AI Applications That Will Set the Trend

The 10 AI Applications That Will Set the Trend

Outlook for the short term (2026–2027) and the medium term (2028–2030)

Artificial intelligence has stopped being a laboratory promise and become working infrastructure. In October 2026 the conversation is no longer about which model is largest, but about which applications deliver measurable value, at what cost and under what controls. This text selects ten applications that, given their technical maturity, business traction and likely impact, will set the agenda through 2030. These are reasoned projections, not certainties: figures from firms such as Gartner are market estimates and should be read with caution, since the history of technology is full of forecasts that arrived earlier or later than expected.

1. Autonomous agents and agentic workflows

This is the dominant trend. An agent does not merely answer: it plans steps, uses tools, carries out actions across different systems and hands the work over for review. Gartner projects that by the end of 2026 about 40% of enterprise applications will include task-specific agents, up from under 5% in 2025, and that by 2028 a third of enterprise software will have agentic capabilities. In the medium term the leap will be toward ecosystems of several agents collaborating with one another. The warning carries equal weight: it is estimated that more than 40% of these projects could be cancelled by 2027 because of weak governance or unclear value. The winners will start with narrow processes, clear metrics and human oversight.

What to watch: the share of tasks resolved without human intervention, the cost per task compared with the manual process, and the number of incidents caused by unauthorized actions. These signals will separate useful pilots from expensive demonstrations.

2. Coding assistants and software agents

Software development is where generative AI shows its most visible return. The shift is from autocomplete to agents that read a repository, propose changes, run tests and open pull requests. The programmer’s role moves toward specifying, reviewing and orchestrating. In the short term, automated code review, legacy-system migration and test generation will grow; in the medium term, the question will be how to ensure quality and security when much of the code is not written by a person. The phenomenon will also reach non-programmers: analysts and business professionals building their own tools by describing them in natural language.

What to watch: the rate of defects and vulnerabilities in generated code, and whether teams keep rigorous review practices despite the pressure for speed. Productivity gains that ignore maintenance costs may prove illusory over time.

3. Multimodal AI and voice as an interface

Text, image, audio and video are converging in a single model. Gartner anticipates that close to 40% of generative AI solutions will be multimodal by 2027. This enables fluid voice assistants, analysis of documents with tables and charts, visual inspection in industry, and customer service that understands a photo of the problem. In the medium term, voice and camera will be everyday interfaces on phones and wearable devices, reducing reliance on the keyboard and the traditional screen.

What to watch: the quality of speech recognition in Latin American Spanish and indigenous languages, where data gaps remain notable, and how quickly wearable devices move from novelty to routine tool.

4. Small, specialized and on-device models

Not every task needs the largest model. Gartner estimates that more than 50% of enterprise generative AI models will be specific to an industry or function by 2027, compared with about 1% in 2023. Small models are cheaper, faster and easier to audit, and they can run on the device itself or on internal servers, which protects sensitive data. The cost of production inference, more than experimentation, will be the line item that forces this efficiency. A mixed architecture is expected: a large model for complex reasoning and small models for repetitive tasks.

What to watch: the performance gap between small and large models, and progress in low-power chips for phones and laptops. If the gap keeps narrowing, much of today’s cloud spending will migrate closer to the user.

5. AI in finance, banking and compliance

The financial sector combines abundant data, regulated processes and a high cost of error, which makes it fertile ground. In the short term we will see reconciliations, analysis of reports and contracts, fraud detection, customer service, and assistants for credit and risk analysts. In audit and compliance, AI will allow a move from sampling to reviewing entire populations and to continuous monitoring. The challenge is explainability: supervisors and auditors will demand traceability for every automated decision, opening a professional field for those who master both internal control and artificial intelligence.

What to watch: regulators’ guidance on automated models, and each institution’s ability to document every decision for an audit. Firms that build this discipline early will face fewer surprises when scrutiny arrives.

6. Healthcare and scientific discovery

In medicine, AI already supports image reading, clinical documentation and patient prioritization. In the short term, reducing administrative burden, through transcription and summaries of consultations, will be the most widespread application. In the medium term, the greatest potential lies in science: protein and drug design, new materials, and laboratories where agents propose hypotheses and plan experiments. Regulatory timelines and the need for clinical validation will moderate the pace, but the cumulative effect could be profound.

What to watch: results of clinical studies that compare AI with standard practice, and the first regulatory approvals of AI-based diagnostic support tools. Evidence, not enthusiasm, will determine how fast hospitals adopt them.

7. Education and personalized tutoring

AI tutors adapt explanations, pace and exercises to each student and give immediate feedback. For teachers, they automate planning, rubrics and a first round of grading. The short term will bring integration into educational and corporate training platforms; the medium term, continuous learning paths tied to job skills. The risks are familiar: dependency, plagiarism and unequal access. Institutions that redesign assessment around reasoning and application, rather than banning the tools, will be better positioned.

What to watch: rigorous evidence of improved learning, not just student satisfaction, and clear acceptable-use policies in schools and universities. Without them, adoption will be uneven and contested.

8. AI-assisted cybersecurity

AI widens both defense and attack. On defense, agents that triage alerts, investigate incidents and prioritize vulnerabilities relieve understaffed teams. On offense, personalized phishing, voice and video impersonation, and automated intrusions lower the cost for criminals. Moreover, the agents themselves become an attack surface: prompt injection, data leakage and abuse of permissions. Expect a sustained arms race and strong demand for identity management, least-privilege permissions and monitoring of non-human systems.

What to watch: mean time to detect and respond to incidents, and the emergence of specific standards for protecting agents with access to critical systems. Organizations should treat every agent as a privileged user.

9. Physical AI: robotics and autonomous vehicles

Advances in models that perceive and reason are reaching robots in warehouses, factories and logistics, and autonomous vehicles in selected cities. In the short term, controlled environments and repetitive tasks will predominate; in the medium term, more general and cheaper robots. The main brake is not only software but hardware, safety and economies of scale. For Latin America, the first opportunity lies in mining, agribusiness and logistics.

What to watch: the unit cost of robots, their reliability in continuous operation, and workplace safety frameworks for sharing space with people. Each of these determines whether pilots become fleets.

10. AI governance, security and trust

As AI acts rather than merely opines, the need to govern it grows: model inventories, continuous evaluation, observability, traceability, data controls and regulatory compliance. Many companies are adopting a central layer, or “AI gateway,” between applications and models to enforce policies, control costs and log activity. Regulation advances unevenly across regions, but the trend points toward greater demands for transparency and accountability. In the medium term, trust will be a competitive advantage: whoever can show that their AI is safe, auditable and useful will scale faster.

What to watch: convergence of international standards, adoption of AI risk-management frameworks, and the demand for independent audits. Boards will increasingly ask for evidence rather than assurances.

Conclusion: three common threads

Three threads emerge from the whole. First, the move from tools that converse to systems that act, with autonomy growing as far as control allows. Second, economics: inference, energy and chips determine what is viable; Broadcom, for instance, projects AI chip revenue of about 115 billion dollars in fiscal 2027. Third, the human factor: organizations that train their people, redesign processes and measure results will capture the value, while those that merely buy technology risk adding to the count of cancelled projects. The practical advice is to start small, measure, govern and then scale. For those working in finance, audit or education, the advantage will lie in combining expert judgment with mastery of these tools: the technology is becoming widely accessible, while good judgment remains scarce. That scarcity, more than any model, is what organizations and professionals should invest in over the coming years.

Sources for the figures: Gartner estimates (applications with agents, industry-specific models, multimodal AI) and a Broadcom projection, as reported by industry coverage consulted in October 2026.

The Mortgaged Empire by Foreign Affairs (2026)

ANALYSIS · THE GEOPOLITICS AND FINANCE OF AI

The Mortgaged Empire

When the geopolitical race for artificial intelligence meets its financial bubble

A joint reading of “Empires of AI” (Niall Ferguson, Lisa Klaassen, and Eric Schmidt) and “The Global AI Bust” (Ruchir Sharma), Foreign Affairs, November/December 2026

Two essays in the same issue of Foreign Affairs describe the same race from angles that are rarely read together. Ferguson, Klaassen, and Schmidt argue that artificial intelligence has given Washington and Beijing a new form of empire, and that the decisive question may not be who reaches artificial general intelligence first, but who can bankroll the race. Sharma answers, from the standpoint of finance, that the race is being paid for with a bubble: real in its technology, fragile in its structure of debt and subsidy. Read together, the two texts suggest a thesis neither fully states: the AI empire is mortgaged, and the date that mortgage comes due will be set less by generals or engineers than by the debt markets.

1. Two diagnoses, one variable

For Ferguson and his coauthors, AI is an imperial technology for a concrete reason: it brought scarcity back to a sector that had ignored it for decades. According to the scaling laws, progress depends on energy, critical minerals, chips, and capital on a massive scale, and once a technology’s frontier depends on scarce inputs, its production becomes a matter of geopolitics. Of all those inputs, capital is the one most likely to constrain both empires. The largest U.S. technology firms have committed more than $600 billion in capital expenditure for 2026, while Beijing, according to a draft plan reported by Bloomberg, intends to spend $295 billion over five years, funded mainly through long-dated government bonds (pp. 71–72).

Sharma looks at the same capital from the inside. Technology spending on plants and equipment now stands at 6.3 percent of U.S. GDP, above the 6.0 percent reached at the peak of the Internet bubble. Of the five hyperscalers, only Microsoft still generates surplus cash flow; four have increased their net debt by issuing bonds, and leases and off-balance-sheet arrangements could add roughly $3 trillion to their debt over the next decade (pp. 85–87). What Ferguson presents as the grand strategic wager of two powers, Sharma presents as a familiar cycle: companies spending more than they can recover, investors borrowing to buy their shares, and prices drifting away from any profits the technology can support.

The point of contact is public debt. Ferguson notes that the United States crossed what he calls the “Ferguson limit” in 2024, paying more in interest (about $881 billion) than on defense (about $874 billion), and that on the current trajectory interest will be double the defense budget by 2036 (p. 72). Sharma identifies that same debtor as the signal that matters: bubbles pop when liquidity tightens, and the players to watch are now less central bankers than the “bond vigilantes,” who punish free-spending governments by selling their bonds (p. 88). The joint reading is uncomfortable: the same debt that narrows the empire’s fiscal room is what can close the tap that feeds the AI expansion.

2. The shared architecture: the stack as empire and as chain of contagion

Ferguson describes AI as a lattice of chokepoints—chips, minerals, models—that the metropoles can switch on or off at will. The text cites a recent example: in June, the U.S. Commerce Department told Anthropic to suspend foreign nationals’ access to its two most capable models (p. 69). There the center decides who gets what; the periphery supplies minerals and cheap labeling work and receives the refined product back at a markup.

Sharma traces the same web, but as a map of economic dependence. TSMC, SK Hynix, and Samsung weigh more in their local indexes than big tech does in the U.S. index; South Korea’s market value rose from $1.5 trillion to more than $5 trillion between early 2025 and this summer; Taiwan’s economy has grown at near double-digit rates; high tech accounts for as much as half of Israel’s growth. Further down the chain, Singapore, Malaysia, Mexico, and Vietnam are receiving data centers and server plants, and suppliers of nickel, lithium, and copper close the chain (pp. 81–82).

What is geopolitical dependence for Ferguson is a transmission channel for Sharma. And both converge on the same node: Taiwan. Ferguson identifies it as the most likely scenario for great-power war, because TSMC sits roughly 90 miles off the Chinese coast and capturing the island would be enough to blunt the supremacy of American models (p. 76). Sharma shows it as one of the most extreme stock market concentrations in the world. A military shock and a financial bust would therefore strike the same point; this coincidence, which neither author stresses, is probably the most important finding of reading them together.

3. The dollar: advantage and exposure

Ferguson argues that the dollar and Wall Street underwrite the American AI empire: owning the reserve currency lets Washington fund its data center build-out with other people’s savings, while China must finance from within, through a state-directed banking system (p. 71). Sharma supplies the other side of that privilege. Foreigners now hold more than $20 trillion in U.S. stocks, up from under $2 trillion after the dot-com bust, with much of the increase coming after ChatGPT’s launch. American households keep more than 30 percent of their wealth in stocks, above the 27 percent of 2000, mostly through passive funds tracking an index dominated by AI (p. 85).

Public support adds to this. According to Sharma, the deficit, above 6 percent of GDP, accounts for more than half of U.S. corporate profits this decade, twice the dot-com level, and the Federal Reserve holds its rate a full point below what academic guidelines would advise (p. 87). Ferguson adds a feedback loop: the capital demands of AI are already pushing interest rates higher and worsening the debt (p. 73). Our reading, which combines the two arguments, is of a circle: the reserve-currency privilege finances the empire and at the same time globalizes its fragility, so the periphery does not merely supply inputs; it also holds the paper that funds the expansion.

4. China: the rival that compresses margins

Ferguson observes that Beijing is following the latecomer path that carried challengers past Britain in the nineteenth century: investing where Washington lacks an edge or cannot match the cost—energy, manufacturing, robotics, and open-weight models at or near the frontier—so that the United States must spend lavishly to defend its lead in a sector China is commoditizing (p. 74). Sharma notes that Chinese models are cheaper and perform nearly as well, and that net exports, mainly of AI-related goods, now account for a third of China’s growth, in a domestic economy weighed down by bad debts and a shrinking labor force (p. 81).

Two consequences follow from combining them. First, competition from cheap models limits the pricing power that Western valuations assume; Sharma points out that two frontier labs charge subscribers less than it costs to train and run their models, yet analysts value each at well over a trillion dollars (p. 84). This is our inference, not a claim by the authors. Second, China is no safe harbor: its economy depends on AI through exports, and Ferguson warns that its own crisis could be set off by overcapacity and the margin-destroying competition known as neijuan, with high youth unemployment and local governments already borrowing solely to roll over old debt (pp. 72–73). A bust would not be an American matter alone.

5. The political ceiling: a bust without a safety net

Ferguson sums up the historical dilemma of empires in what he calls the “imperial trilemma”: a power can do two of three things—defend its dominions, maintain technological primacy, or provide generous welfare to the metropole—but not all three (p. 75). Sharma adds the context of a world without a cushion: global debt has surpassed $350 trillion, more than three times the size of the world economy, and few governments will have the money to stimulate battered economies or bail out businesses (p. 90).

Moreover, it is doubtful the public would allow it. Sharma cites that most Americans say they are more concerned than excited about AI, that seven in ten expect fewer jobs, and that local opposition blocked or delayed data center projects worth some $130 billion in the first quarter alone (p. 90). Ferguson records the same backlash, even in Republican-led states such as Texas and Utah (p. 75). The combination suggests that in 2008 and 2020 the state absorbed the blow, and that in an AI bust it may be unable or unwilling to do so. A bursting bubble would thus work as a stress test for the empire: growing deficits, shrinking room, and military commitments that do not go away, just when Ferguson warns that American overextension in the Middle East may tempt Beijing (pp. 76–77).

6. Where they disagree

The authors do not agree on the essentials of the outcome. Ferguson sees in AI a durable imperial resource whose main threat is overreach, rivalry, and, at the extreme, war; his way out is a form of détente that limits the race through agreements, for example by setting capability thresholds for models with cyber-offensive or biological potential (p. 78). Sharma sees a “good” speculative episode, comparable to the Internet or the railroads: after the crash a more streamlined network of survivors will remain, and the technology will keep advancing (p. 91).

The underlying question is whether the bust is the threat or the purge. Our reconciling hypothesis is that the crash would not end the technology (Sharma), but could change who is able to sustain the imperial order built upon it (Ferguson). The survivors would be those with access to the deepest and cheapest capital, whether dollar-financed or backed by the Chinese state. The bust would act as a selection mechanism among firms and also among empires. There is also an important agreement: neither believes AI is a magic card against fiscal problems. Ferguson warns that productivity gains may arrive too late to avoid the debt trap; Sharma recalls that technologies usually lift productivity decades after they appear, and that current forecasts range from 0.1 to nearly 2 percentage points a year (pp. 73, 91).

7. What to watch

•         Long-term bond yields and episodes of “bond vigilantes” in the United States, Europe, and Japan: the liquidity signal that, according to Sharma, precedes busts.

•         Circular funding and “other income” at the hyperscalers, along with off-balance-sheet leases: they measure how much of the earnings is real.

•         Initial public offerings of the frontier labs, with prices below cost and valuations without profits.

•         Semiconductor volatility in Korea, Taiwan, and Japan, and any sign of tension in the Taiwan Strait.

•         Refinanced Chinese local debt and signs of neijuan in subsidized sectors.

•         Local opposition to data centers and governments’ fiscal response to a potential shock.

8. A view from the periphery: a reading for the Andes

Both texts place the region in the supply chain. Sharma lists copper from Chile, Peru, and the Democratic Republic of the Congo among the inputs that will sustain years of demand from the worldwide AI construction project, and warns that the shock of a bust would reach, at the end of the chain, miners and power producers (pp. 82, 89). Ferguson describes the pattern of extraction in the periphery and refining and resale in the metropole (p. 66). A reading of our own, not the authors’: economies such as Peru’s have high exposure to the AI demand cycle and low participation in its rents, so their best defense lies in following global liquidity indicators more closely than technology announcements. Sharma adds that countries such as India and the Philippines are doubly exposed, left out of the investment wave and at risk of displacement in their back-office services (p. 82).

Conclusion

Ferguson, Klaassen, and Schmidt explain why AI is a contest between empires; Sharma explains why that contest rests on a financial foundation that looks too much like those of earlier bubbles. Together they draw a double risk: a war over one node, Taiwan, and a financial bust with that same node at its epicenter, in a world where states have little room to cushion the fall. The lesson they share is an old one: empires seldom fall for lack of ambition, but because they cannot pay for what they aspire to. What is new is the speed. And AI, unlike the railroad, offers no decades in which to learn how to pay for itself.

Methodological note. This text was prepared from a full reading of both articles. Figures and facts are attributed to the authors and were not independently verified. Readings that combine or extend the two texts (the Taiwan node coincidence, the dollar circle, margin compression, the selection hypothesis among empires, and the Andean reading) are inferences of this analysis and are flagged as such. Page numbers refer to the printed edition.

Sources

Ferguson, N., Klaassen, L., & Schmidt, E. (2026). Empires of AI: The Rise and Fall of a New Global Order. Foreign Affairs, 105(6), 62–78.

Sharma, R. (2026). The Global AI Bust: What Will Survive When the Boom Ends. Foreign Affairs, 105(6), 79–91.

Inside the Mind of AI: How Attention and Neural Networks Turn Words into Meaning

Inside the Mind of AI: How Attention and Neural Networks Turn Words into Meaningx

A beginner-friendly guide to understanding how generative artificial intelligence processes language, connects ideas, and produces answers.

Meta description: Discover how AI understands context through tokens, attention mechanisms, and neural network layers. A practical, step-by-step guide to the technology behind ChatGPT.

1. Introduction: How Does a Machine Learn to Work with Language?

Imagine asking an artificial intelligence system a simple question:

Why does Peter Thiel believe that creating something new is more valuable than simply competing with existing businesses?

Within seconds, the AI may produce a coherent explanation involving innovation, technology, entrepreneurship, and competitive advantage.

But how does this happen?

The computer does not read the question in exactly the same way a human does. It processes numerical representations of language through a mathematical system known as a neural network. During this process, it calculates relationships between words, uses contextual information, transforms representations through multiple computational layers, and generates a response one token at a time.

One of the most important technologies behind modern language models is the Transformer architecture, which uses a mechanism called attention to help the model determine how different parts of a text relate to one another.

Understanding this process does not require an advanced degree in mathematics. We can explore it step by step, beginning with a simple sentence and gradually introducing the underlying technology.

2. Step One: The AI Receives a Question

Every interaction begins with an input: a question, an instruction, a document, or another type of information.

Consider this sentence:

"Peter Thiel argues that entrepreneurs should create something new rather than merely compete with existing companies."

A human reader immediately recognizes several elements: Peter Thiel is associated with an argument, entrepreneurs are the intended audience, and creating something new is contrasted with competing against established businesses.

For an AI model, the input first enters a computational pipeline.

The system must represent the text in a form that its neural network can process. It also needs to take account of the surrounding conversation and the user's instructions.

For example, the same sentence could lead to different responses depending on whether the user requests a summary, a translation, a business analysis, or an explanation for children.

The first lesson: the input does not determine the response by itself. The task and the surrounding context also matter.

3. Step Two: Words Become Tokens

Computers perform calculations on numerical data. Consequently, the text must be converted into units that the model can process.

These units are called tokens.

A token may correspond to a complete word, part of a word, punctuation, or another text fragment. Tokenization depends on the model's vocabulary and is not necessarily a one-word-at-a-time process.

For example, the sentence:

"AI creates new possibilities."

might be divided into tokens corresponding approximately to:

• "AI"

• "creates"

• "new"

• "possibilities"

• "."

This is only an illustration; an actual tokenizer may divide the sentence differently.

Each token is associated with a numerical identifier. The model then uses learned numerical representations, called embeddings, to process information about these tokens.

Embeddings are important because they allow a neural network to work with relationships between linguistic patterns rather than treating every word as an isolated symbol.

For instance, the concepts of "innovation," "invention," and "technology" may exhibit useful relationships in the model's learned representations.

However, an embedding is not a dictionary definition stored in a single number. It is a vector containing many numerical values, and its significance depends on how the model uses it.

The second lesson: before an AI can process language mathematically, it converts the text into tokens and numerical representations.

4. Step Three: Context Changes the Meaning of Words

Words rarely have just one meaning.

Consider the following sentences:

1. "The bank approved the loan."

2. "We sat on the bank of the river."

The word "bank" appears in both sentences, but its meaning differs.

A useful language model must therefore take account of surrounding words. It cannot rely solely on an isolated representation of "bank."

This is where contextual processing becomes essential.

In a Transformer model, the representations of tokens are repeatedly transformed as information from other tokens is incorporated. As a result, the representation associated with a word can reflect its role in the particular sentence.

This does not mean the model consults a dictionary and consciously chooses a definition. Rather, it computes numerical relationships that help distinguish the relevant usage.

The same principle applies to more complicated questions involving cause and effect, comparisons, pronouns, and abstract concepts.

For example:

"Maria left the book on the table because it was wet."

What does "it" refer to? The book, most likely, although context can affect the interpretation.

A model must use relationships between the elements of the sentence to produce an appropriate continuation.

The third lesson: context helps determine how a word should be interpreted in a particular situation.

5. Step Four: The Attention Mechanism Connects Information

The attention mechanism is one of the central innovations of the Transformer architecture.

Its purpose is to calculate how information from different tokens should contribute to the representation being computed for another token.

Let's return to a business example:

"The technology company introduced a new processor because it improved computational speed."

The word "it" most plausibly refers to the processor.

To interpret this sentence, the model can calculate relationships between "it," "processor," "improved," and other relevant parts of the text.

Attention helps the model combine information from these different positions rather than treating each token as an independent item.

A simplified illustration

Imagine the word "it" asking which earlier words are relevant to its interpretation.

The model calculates numerical compatibility between representations of tokens and uses the resulting weights to combine information from them.

The relationship between "it" and "processor" may be useful for resolving the reference. Other relationships may help explain what improved and why.

Importantly, the model does not need to follow a single rigid path from one word to another. Attention can consider many relationships in parallel.

A sentence can therefore be processed as a network of interacting linguistic elements rather than merely as a sequence of unrelated dictionary entries.

How does attention work mathematically?

The standard scaled dot-product attention operation uses three representations:

• Query (Q): represents what information a token is seeking in the current computation.

• Key (K): represents features used to calculate how relevant another token may be.

• Value (V): contains information that can be combined using the calculated attention weights.

The operation is expressed as:

Attention(Q, K, V) = softmax(QKᵀ / √dₖ)V

Here is the basic interpretation:

1. The model compares queries with keys.

2. These comparisons produce numerical scores.

3. A softmax function converts the scores into normalized weights.

4. The weights are used to combine the value vectors.

The result is a new representation that incorporates information from other tokens.

Actual Transformers use multiple attention heads, allowing different learned projections to capture different patterns of relationships. These patterns do not necessarily correspond to neat human categories such as grammar, emotion, or logic.

Also, attention weights are not a complete explanation of a model's reasoning. They describe part of the computation, not a transparent record of every factor behind the final answer.

The fourth lesson: attention helps a neural network calculate relationships between parts of a text and combine relevant information.

6. Step Five: Neural Network Layers Transform the Representations

Attention is only one part of the process.

A Transformer also contains multiple neural network layers that transform the information as it passes through the model.

To understand this, imagine studying the sentence:

"Entrepreneurs should create something new rather than merely compete."

At first, the model has numerical representations of the tokens. Through successive computations, these representations are transformed in ways that can capture increasingly useful combinations of information.

A simplified teaching example might describe the process as follows:

Stage A — Linguistic relationships: The model processes how words relate within the sentence.

Stage B — Contrasts: It incorporates the relationship between "create something new" and "compete."

Stage C — Broader context: It combines these relationships with other information in the question, including its association with Peter Thiel and entrepreneurship.

Stage D — Response generation: The resulting representations contribute to the probabilities of possible next tokens.

These stages are a conceptual illustration, not a literal map of individual layers. Real neural networks distribute information across many components, and no universal rule says that one layer recognizes words, another understands sentences, and another performs reasoning.

A typical Transformer layer includes attention, a feed-forward neural network, residual connections, and normalization operations. The precise architecture varies between models.

The feed-forward network transforms representations through learned mathematical operations. Residual connections help information and gradients flow through the network, while normalization helps stabilize computation.

Together, these components enable the model to build complex representations from the original input.

The fifth lesson: attention helps connect information, while the surrounding neural network layers transform that information into representations useful for the task.

7. Step Six: From Relationships to a Coherent Answer

Once the model has processed the input through its layers, how does it actually write a response?

A language model typically calculates a probability distribution over possible next tokens.

Suppose the response begins:

"Thiel believes that innovation is important because..."

At this point, many continuations are possible. The model assigns scores to candidate tokens and converts them into probabilities.

Depending on the model and its decoding settings, the next token may be selected through a strategy that favors the most probable candidate or samples among possible alternatives.

The chosen token is added to the generated text. The model then continues the process, using the updated context to generate further tokens.

This continues until the model reaches an appropriate stopping condition.

The complete response emerges through this successive generation process.

However, the fact that a model predicts tokens does not mean it simply strings together random words. Its predictions are conditioned on the prompt, the context, and the learned parameters of a neural network trained to produce useful language.

The model can generate explanations, summarize arguments, translate text, and solve many problems because its learned representations capture complex patterns in language and other data.

The sixth lesson: an AI response is generally produced incrementally, with each new token conditioned on the information available to the model.

8. Step Seven: Training and Inference Are Different Processes

To understand how modern AI works, it is essential to distinguish training from inference.

Training: learning from data

During training, a model processes large collections of examples and adjusts its numerical parameters to improve its performance on a learning objective.

In a simplified next-token training task, the model attempts to predict tokens from their preceding context. Errors between its predictions and the training targets are used to calculate updates to its parameters through optimization techniques such as gradient descent.

Training is computationally intensive and may involve enormous amounts of data and specialized hardware.

Many modern systems also undergo additional stages, such as supervised fine-tuning and preference-based optimization, to improve their ability to follow instructions and produce useful responses.

Inference: responding to a user

During inference, the trained model uses its existing parameters to process a new prompt and generate an answer.

When a user asks ChatGPT to summarize Zero to One, the model does not ordinarily retrain itself on the book before answering.

Instead, it uses its learned parameters and the context available in the current interaction. If tools such as web search are available and used, information retrieved from external sources may also inform the response.

This distinction explains how a model can answer new questions without being retrained for every conversation.

The seventh lesson: training establishes the model's learned parameters; inference uses those parameters to produce responses.

9. Step Eight: Why AI Can Be Impressive but Still Make Mistakes

A coherent answer is not necessarily a correct answer.

Language models can produce inaccurate statements, invent citations, confuse people or dates, and present uncertain conclusions with confidence. This phenomenon is commonly called hallucination.

Why does it happen?

The process of generating plausible language is not identical to independently verifying every factual claim. Even when the model has learned useful patterns, its internal representations do not guarantee that a particular statement is true.

Consider asking for a summary of a book that the model cannot reliably recall. It might produce a convincing explanation of the author's general philosophy while incorrectly attributing a specific argument to a particular chapter.

Several practices can reduce this risk:

• Consult the original book or another authoritative source.

• Use reliable external references when exact facts or quotations matter.

• Distinguish the author's documented claims from the AI's interpretation.

• Check dates, numbers, names, and citations.

• Ask the model to identify uncertainty rather than assume every statement is established fact.

For more demanding tasks, retrieval-augmented generation, source-grounded workflows, and independent verification can improve reliability. They do not eliminate every error.

The eighth lesson: an AI system can generate a plausible answer without having verified every claim in it.

10. Putting Everything Together: From a Question to an Answer

Let's summarize the complete process using our original question about Peter Thiel.

The user asks:

"Explain why Peter Thiel believes creating something new is more valuable than simply competing."

The process can be represented as a sequence:

1. Input: The model receives the user's question and relevant conversation context.

2. Tokenization: The text is divided into tokens.

3. Representation: Tokens are mapped to numerical representations.

4. Attention: The model calculates relationships between relevant parts of the text.

5. Neural network layers: The representations are repeatedly transformed and combined.

6. Output probabilities: The model calculates possible next tokens.

7. Generation: Tokens are produced successively to form an explanation.

8. Evaluation: The response can be checked for relevance, clarity, and factual accuracy.

Each stage contributes to the final result, although actual architectures may contain additional components and the internal computations are more complex than this simplified description.

The central insight is that generative AI combines learned statistical patterns, numerical representations, contextual processing, and sequential text generation.

It does not need to retrieve a complete, prewritten answer for every question. It can construct a new response based on the information and patterns available to it.

Conclusion: Understanding the Technology Behind Generative AI

Modern generative AI is not simply a database that retrieves stored sentences. It is a computational system that processes representations of language through a trained neural network.

Tokenization provides the units of computation. Embeddings represent those units numerically. Attention calculates relationships between tokens. Neural network layers transform and combine the resulting representations. Finally, the language model generates a response by predicting successive tokens.

These mechanisms help explain why AI can summarize books, explain scientific concepts, draft business strategies, and communicate in natural language.

But they also explain why fluent language should not be confused with guaranteed factual accuracy. Generating a convincing answer and establishing that the answer is true are different tasks.

For students, professionals, and anyone interested in artificial intelligence, understanding this distinction is essential.

The most useful way to think about generative AI is this: it is a powerful system for learning patterns and generating context-sensitive responses, while human judgment and appropriate verification remain crucial for determining whether those responses deserve our trust.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

---

Glossary: 10 Essential Terms

1. Artificial Intelligence (AI): A broad field concerned with systems that perform tasks associated with capabilities such as perception, language processing, prediction, and decision-making.

2. Token: A unit of text processed by a language model. It may be a word, a word fragment, punctuation, or another text unit.

3. Embedding: A numerical vector representing a token or another item in a learned mathematical space.

4. Neural Network: A computational model composed of interconnected mathematical operations whose parameters are learned during training.

5. Transformer: A neural network architecture that uses attention mechanisms and other components to process sequences and their relationships.

6. Attention: A mechanism that calculates how information from different tokens contributes to a token's representation.

7. Attention Head: One attention computation operating through its own learned projections. Multiple heads allow a Transformer layer to combine different patterns of relationships.

8. Parameters: Learned numerical values within a model that determine how its computations transform inputs into outputs.

9. Training: The process of optimizing model parameters using data and a learning objective.

10. Inference: The process of using a trained model to produce predictions or generate responses for a new input.

Suggested References

• Vaswani, A., et al. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems. https://arxiv.org/abs/1706.03762

• Jurafsky, D., & Martin, J. H. Speech and Language Processing. An authoritative textbook on language processing and modern language models. https://web.stanford.edu/~jurafsky/slp3/

• OpenAI. How ChatGPT and Our Foundation Models Are Developed. https://openai.com/index/how-chatgpt-and-our-foundation-models-are-developed/

Editorial note: This article provides a conceptual introduction to Transformer-based language models. The precise implementation details vary between AI systems; the diagrams and staged descriptions are pedagogical simplifications rather than a literal account of every internal computation.

From Unimate to Humanoids

Six decades of robots (1961–2026): how Japan, the United States, Europe and China built the industry, and what is reasonable to expect by 2045.

Industrial robotics began in 1961 with a hydraulic arm installed in a General Motors plant, and today more than five million units operate in the world’s factories. This article retraces that history with a constant structure (technical milestone, Japan, the United States, Europe and China), updates the figures with the IFR’s September 2026 reports, and separates three kinds of claims that are often blended: measured facts, company targets and analyst scenarios. The conclusion is less tidy than the familiar formula of “Japan, quality; the United States, innovation; China, scale.”

1. The foundations (1961–1989): from hydraulics to electricity

Technical milestone. The first industrial robot, the Unimate of George Devol and Joseph Engelberger, went into service in 1961 at General Motors’ Inland Fisher Guide plant in Ewing Township, New Jersey, handling die-cast parts ^[1]. It was a hydraulically driven arm with magnetic-drum memory. The next break came from Sweden: in 1974 ASEA launched the IRB 6, the first commercially available all-electric, microprocessor-controlled robot; development had begun in 1970 and a patent was granted in 1972 ^[3]. By 1992 ASEA had sold nearly 2,000 IRB 6 units ^[4]. The shift to electric drives therefore starts in the 1970s and becomes general in the 1980s.

United States. It supplied the founding technology, but Unimation chose to license it: in October 1968 it signed a technical agreement with Kawasaki that opened the door to Japan’s industry ^[2].

Japan. Kawasaki was the entry point. Yaskawa showed its first MOTOMAN in 1974 and in 1977 sold the MOTOMAN-L10, Japan’s first all-electric industrial robot, modeled on ASEA’s electric articulated robot ^[6]. FANUC, independent from Fujitsu since 1972, began commercial production and shipment of robots in 1977 ^[5]. Japan was also a humanoid pioneer: Waseda University completed WABOT-1 in 1973 ^[7].

Europe. Sweden’s ASEA pioneered the commercial electric robot; in 1988 it merged with Switzerland’s BBC to form ABB ^[4]. Together with Germany’s KUKA, ABB would later make up the group known as the “Big Four” alongside FANUC and Yaskawa.

China. Still in the research stage: the Shenyang Institute of Automation of the Chinese Academy of Sciences developed China’s first teach-and-playback industrial robot prototype in 1982 ^[17].

2. Consolidation and opening (1990–2009)

Technical milestone. Robots became easier to program, gained basic machine vision, and moved from automotive plants into electronics and logistics. The most visible leap happened outside the factory: the da Vinci surgical system received FDA clearance for general laparoscopic surgery in July 2000 ^[9], and iRobot’s Roomba (2002) brought the robot into the home.

Japan. It kept its industrial hegemony and led humanoid research until the early 2000s ^[8]. Honda began a confidential project in 1986, showed the P2 prototype in 1996 and presented ASIMO in 2000 ^{[7, 8]}.

United States. It shifted toward mobile, medical and service robotics. Boston Dynamics was founded in 1992 as an MIT spin-off by Marc Raibert ^[10]; for years its advanced locomotion systems stayed in the realm of research and defense.

Europe. ABB and KUKA remained among the reference suppliers in the traditional industrial segments.

China. SIASUN was founded in 2000 as a spin-off of the Shenyang Institute of Automation ^{[16, 17]}. Foreign brands dominated the Chinese market: even between 2019 and 2021, local suppliers’ share was only around 30% ^[22].

3. Cobots, logistics and artificial intelligence (2010–2025)

Technical milestone. Collaborative robots appear, designed to work alongside people without rigid fences, and AI enters practice: advanced vision, learning and autonomous navigation. The segment’s reference was Denmark’s Universal Robots, acquired in 2015 by the US company Teradyne for $285 million plus up to $65 million tied to targets; it had already sold more than 4,000 units ^[12].

United States. Logistics was its terrain. Amazon bought Kiva Systems in 2012 for about $775 million ^[11], the origin of its fleet of mobile robots. Boston Dynamics passed through Google (December 2013), SoftBank (June 2017) and Hyundai, which acquired 80% in December 2020 for about $880 million and completed the deal in June 2021 ^[10]. Spot went on sale to the public in June 2020 ^[10].

Japan. It remained the largest manufacturer: in 2018 it delivered 52% of the world supply of industrial robots, according to the IFR ^[25]. Its home market, however, is weakening: installations fell 4% in 2024 and accounted for 8% of the world total ^[20].

Europe. Two emblematic assets changed hands. Germany’s KUKA was acquired by China’s Midea: the bid, of about €4.5 billion, was launched in 2016 and the deal closed in January 2017 with 94.55% of the capital ^{[13, 14]}. In October 2025, ABB agreed to sell its robotics division to SoftBank for $5.375 billion, with closing expected between mid and late 2026 ^[15]; the sources consulted do not yet confirm that it has been completed.

China. It has been the world’s largest market since 2013 ^[20]. Local suppliers’ share accelerated from 2022: since 2024 more than half of Chinese installations are from Chinese brands ^[22], reaching 55% in 2025 ^[23]. Estun, with roots in motion control and automation components, was, according to a Frost & Sullivan report cited in its listing prospectus, the first local maker to lead the Chinese market by shipment volume, in the first half of 2025 ^[18]. In humanoids, the Ministry of Industry and Information Technology (MIIT) set in November 2023 the goal of mass production by 2025 and of making them a new growth engine by 2027 ^[34].

4. October 2026: where we stand

The most recent data come from the World Robotics 2026 report, published by the IFR on 24 September 2026. In 2025, 603,307 industrial robots were installed, 11% more than the year before and an all-time high; the operational stock passed five million units ^{[21, 22]}. Asia accounted for 76% of installations, while Europe contracted by 14% ^[22]. Table 1 summarizes the main figures.

 

 

 

 

 

 

 

Two caveats help in reading these figures. First, the main customer is no longer just the automotive industry: in 2025 electrical and electronics was the largest buyer, with 142,235 units, followed by automotive with 138,429 ^[22]. Second, density measures robots per worker, not total volume, so small, highly automated countries such as Korea and Singapore top the ranking even though China installs more than half of the world’s robots ^{[24, 22]}.

Humanoids: few, fast and poorly measured. There is no single official statistic. According to Omdia, about 13,000 humanoids shipped worldwide in 2025, almost six times the previous year: AgiBot led with 5,168 units (39%), Unitree shipped 4,200 and UBTECH about 1,000 ^{[26, 28]}. Other Chinese companies and the US firms Figure AI, Agility Robotics and Tesla each shipped between 150 and 500 units ^[27]. Unitree says it delivered more than 5,500 ^{[29, 30]}, and Counterpoint estimates that China accounted for more than four in five humanoids sold ^[29]. The orders of magnitude agree (tens of thousands, not millions), but the exact figures depend on what counts as a humanoid and who reports.

The IFR itself is cautious: in its 2026 summary it notes that most humanoids remain prototypes or pilots, that conventional industrial robots are today economically superior for production, and that broad deployment will depend on substantial improvements in reliability, safety, operating time and economic viability ^[22]. In November 2025, China’s National Development and Reform Commission warned that a bubble might be forming in the sector ^[29]. Unitree, for its part, launched in July 2026 its initial public offering on Shanghai’s STAR Market, seeking about 4.2 billion yuan ^{[30, 31]}; the sources consulted do not allow us to confirm the final outcome of the listing.

Tesla. Optimus was unveiled in 2021 and, as of September 2026, Tesla had still not publicly shown its third generation; the company said it is installing production lines at Fremont designed for one million robots a year, and that a second line in Texas aims in the long term at ten million ^[32]. Announced timelines have slipped several times, and third-quarter results, expected in late October, will be the next occasion to see whether they hold ^[33].

5. Toward 2045: facts, targets and scenarios

The last third of any article on robotics mixes three different things. Table 2 separates them.

 

 

 

 

 

 

The scenarios diverge by orders of magnitude. Goldman Sachs Research estimated in its 2024 report an addressable market of $38 billion by 2035, six times its prior figure ^[35], with about 1.4 million units a year according to press coverage of the report ^[36]. Morgan Stanley projects a market above $5 trillion in 2050, a figure that includes supply chains and repair and maintenance services, and more than 1 billion humanoids in use ^[37]; press coverage adds about 13 million units in service by 2035 ^[38]. Omdia, for its part, projects 2.6 million annual shipments in 2035 ^[28]. For conventional industrial machines, the IFR forecasts 806,000 annual installations in 2029 ^[22]. The distance between these figures is not an error: it reflects different horizons, definitions and assumptions, which is why every figure should be cited with analyst, year and scenario.

With that caution, this is the most likely scenario for the next two decades:

•      Japan would keep its weight in high-reliability industrial robots and, above all, in components: Nabtesco holds about 60% of the world market for RV reducers for medium and large robots, according to Morningstar ^[42]. Its home market is small, but Japanese capital is moving toward “physical AI,” as SoftBank’s purchase of ABB Robotics shows ^[15].

•      The United States would continue to set the frontier in software, AI applied to robots and logistics (Amazon), with medical and defense applications. In humanoids, its companies shipped hundreds of units in 2025, not thousands ^[27].

•      China has a high probability of leading production volume and adoption, backed by its supply chain and industrial policy ^[34]; it already accounts for more than 80% of humanoids sold ^[29] and 59% of industrial installations ^[23]. Its risk is overheating: its own planning authority warned of a possible bubble ^[29].

•      Europe retains standards, installed base and engineering capacity, but faces weak demand (−14% in 2025) and has sold strategic assets (KUKA to Midea; ABB Robotics to SoftBank) ^{[22, 13, 15]}.

•      South Korea leads the world in robot density ^[24] and, through Hyundai, controls Boston Dynamics ^[10].

6. Bottlenecks and rules of the game

Critical components. A robot is a sum of precision reducers, motors, sensors and chips. Precision reducers, a segment with a strong Japanese presence ^[42], and rare-earth magnets, concentrated in China, are the two clearest examples. China has required export licenses for seven medium and heavy rare earths since April 2025; the broader October 2025 package was suspended for one year after negotiations with the United States, and that pause expires in November 2026 (sources differ between the 10th and the 27th of that month) ^{[39, 40]}. Elon Musk acknowledged in 2025 that the restrictions affected Optimus development ^[41], and Morgan Stanley observes that almost every developer in the world depends on critical components from China and elsewhere in Asia ^[37].

Safety and regulation. In February 2025 the new ISO 10218-1 and ISO 10218-2 were published; they incorporate the requirements of the former ISO/TS 15066 specification on collaborative applications, define new robot classes and add cybersecurity requirements ^{[43, 44]}. In the European Union, Regulation (EU) 2023/1230 on machinery applies from 20 January 2027 and replaces Directive 2006/42/EC ^[45]; the Artificial Intelligence Act, Regulation (EU) 2024/1689, adds a horizontal layer with obligations phased in over time ^[46].

The gap between demo and operation. A demonstration is not a working shift. What needs watching, following the IFR, is reliability, safety, operating hours and the real profitability of each deployment ^[22].

Employment, energy, ethics and data. These are real challenges, although their magnitude depends on the speed of adoption, which is precisely the most uncertain variable. The transition will be gradual and will begin in controlled environments, such as factories and warehouses, before reaching homes.

7. Conclusion

The formula “Japan contributed industrial excellence; the United States, innovation; China, scale” is useful but incomplete. Europe launched the first commercial electric, microprocessor-controlled robot (ASEA, 1974) ^[3]. Japan did not merely build better: it pioneered the humanoid, from WABOT-1 to ASIMO ^{[7, 8]}. The United States also has scale, that of Amazon’s fleets of mobile robots ^[11]. And China no longer just scales: its brands already hold more than half of its industrial market and its companies dominate humanoid shipments ^{[23, 26]}.

The most prudent data point of 2026 comes from those who measure factories, not those who project markets: conventional industrial robots remain the economically superior option, and humanoids are still mostly pilots ^[22]. The most likely outcome is a gradual transition in which the robot moves from a specialized machine to an increasingly general tool; whether it becomes an everyday coworker will depend less on artificial intelligence than on reliability, cost, safety and supply chains.

Glossary

AMR (autonomous mobile robot) — A robot that moves on its own without fixed tracks or guides, using sensors and maps. It differs from an AGV, which follows predefined routes.

Big Four — Industry term for the four large traditional makers of industrial robots: FANUC, Yaskawa, ABB and KUKA.

Cobot (collaborative robot) — A robot designed to work in contact with or near people, without rigid fences, with limits on force and speed. Its safety requirements are now covered by the ISO 10218 series ^[43].

Embodied AI / physical AI — Artificial intelligence systems that perceive and act in the physical world through a robotic body.

End effector — The tool or gripper mounted at the end of a robot (welding, gripping, painting).

Hydraulic and electric — Two drive technologies. The Unimate was hydraulic (high force, complex maintenance); ASEA’s IRB 6 was the first commercial all-electric robot ^[3].

Humanoid — A robot whose shape and movement are inspired by the human body, normally bipedal. Some classifications exclude torsos mounted on wheeled bases.

IFR (International Federation of Robotics) — A federation of national associations, manufacturers and research centers that publishes World Robotics each year, the reference statistic on industrial and service robots.

Industrial robot — Under ISO 8373, an automatically controlled, reprogrammable, multipurpose manipulator, programmable in three or more axes, fixed or mobile, for industrial automation.

ISO 10218 — International safety standard for industrial robots (part 1) and for robot applications and cells (part 2); revised in 2025 ^[43].

Articulated robot — An arm with several rotary joints; the most common configuration of the industrial robot.

Rare-earth magnets (NdFeB) — Neodymium–iron–boron magnets, used in high-torque motors; their supply depends heavily on China and is subject to export controls.

Reducer (RV and harmonic) — A precision transmission that reduces speed and multiplies torque in each joint. The RV, a two-stage cycloidal design, is used on heavy axes; the harmonic, on wrists and cobots ^[42].

Robot density — Number of operational industrial robots per 10,000 manufacturing employees. It measures automation intensity, not absolute volume ^[24].

SCARA — An arm with joints in a horizontal plane, fast and precise, widely used in electronics assembly.

Service robot — A robot that performs useful tasks for people or equipment outside industrial automation: cleaning, logistics, surgery, inspection.

Shipments versus installations — Shipments count units delivered by the manufacturer; installations count units put into operation. Humanoid figures are usually shipments; the IFR’s industrial robot figures are installations.

STAR Market — The Shanghai Stock Exchange’s board for technology companies, launched in 2019.

Teleoperation — Remote control of a robot by a person; common in demonstrations and in collecting training data.

Total addressable market (TAM) — The theoretical maximum revenue of a product if adopted according to an analyst’s assumptions; it is not a sales forecast.

VLA (vision-language-action) — Models that combine vision, language and motor control to turn natural-language instructions into robot actions.

 

References

All sources were consulted on 17 October 2026. References 1 to 46 are grouped by topic.

A. History and companies

[1]     Guinness World Records. “First industrial production robot” (Unimate, 1961, General Motors). https://www.guinnessworldrecords.com/world-records/411563-first-industrial-production-robot

[2]    Kawasaki Robotics. History: technical agreement with Unimation (October 1968). https://robotics.kawasaki.com/en1/anniversary/history/history_01.html

[3]    ABB Review 2/2014. “Rise of the robot: celebrating 40 years of industrial robotics at ABB” (IRB 6, 1974). https://library.e.abb.com/public/8db01db124333c1bc1257ddc0046cdb7/ABB%20Review%202-2014_72dpi.pdf

[4]    ABB. “125 Jahre ABB: Meilensteine” (IRB 6; ASEA–BBC merger, 1988). https://new.abb.com/ch/125-jahre-abb/meilensteine

[5]    FANUC. “FANUC’s History” (1972; commercial robot production, 1977). https://www.fanuc.co.jp/en/profile/history/

[6]    Yaskawa. “The Birth of MOTOMAN.” https://www.yaskawa-global.com/technology/history/robotics

[7]     Waseda University. “A Quick Visit to the Humanoid Robotics Institute” (WABOT-1, 1973; ASIMO, 2000). https://www.waseda.jp/inst/cro/news-en/2025/06/03/19639/

[8]    Redalyc. “Robots that look like humans: a brief look into humanoid robotics” (Honda P2 and ASIMO). https://www.redalyc.org/journal/5117/511766773023/

[9]    Intuitive Surgical. Second-quarter 2000 results release (FDA clearance, 11 July 2000). https://investor.intuitivesurgical.com/news-releases/news-release-details/intuitive-surgical-announces-second-quarter-results

[10]  Wikipedia. “Boston Dynamics” (secondary source). https://en.wikipedia.org/wiki/Boston_Dynamics

[11]   CNNMoney. “Amazon buys army of robots” (Kiva Systems, 20 March 2012). https://money.cnn.com/2012/03/20/technology/amazon-kiva-robots/

[12]  Universal Robots / Teradyne. Acquisition announcement (13 May 2015). https://www.universal-robots.com/fr/actualites-et-medias/centre-d-actualites/teradyne-and-universal-robots-announce-agreement-for-teradyne-to-acquire-universal-robots-leader-in-collaborative-robots/

[13]  AidData. Midea’s acquisition of KUKA (closed 6 January 2017). https://china.aiddata.org/projects/102145

[14]  Industriemagazin. “Jetzt ist die Kuka-Übernahme durch Midea durch” (€4.5 billion bid). https://industriemagazin.at/artikel/jetzt-ist-die-kuka-uebernahme-durch-midea-durch/

[15]  SoftBank Group. “Acquisition of ABB Ltd’s Robotics Business” (8 October 2025). https://group.softbank/en/news/press/20251008

[16]  China Daily. SIASUN profile (2009). https://www.chinadaily.com.cn/bw/2009-02/16/content_7478545.htm

[17]   36Kr. History of the Shenyang Institute of Automation and SIASUN (1982 prototype). https://eu.36kr.com/en/p/3645860988243845

[18]  Estun Automation. Hong Kong listing profile, with Frost & Sullivan data (secondary source). https://futures.etnet.com.hk/www/eng/stocks/ci_ipo_detail.php?code=02715

B. Industrial robot figures (IFR)

[19]  IFR. “Global Robot Demand in Factories Doubles Over 10 Years” (World Robotics 2025, 25 September 2025). https://ifr.org/ifr-press-releases/news/global-robot-demand-in-factories-doubles-over-10-years

[20] IFR. Executive summary, World Robotics 2025 – Industrial Robots. https://ifr.org/img/worldrobotics/Executive_Summary_WR_2025_Industrial_Robots.pdf

[21]  IFR. “Five Million Robots now Operate in Factories Globally” (World Robotics 2026, 24 September 2026). https://ifr.org/ifr-press-releases/five-million-robots-now-operate-in-factories-globally

[22]  IFR. Executive summary, World Robotics 2026 – Industrial Robots. https://ifr.org/img/worldrobotics/Executive_Summary_WR_2026_-_Industrial_Robots.pdf

[23]  IFR. “China Installs 59 Percent of World’s Industrial Robots” (24 September 2026). https://ifr.org/downloads/press_docs/EN-2026-SEP-24-IFR_Press_Release_WR-CHINA.pdf

[24]  IFR. “Robot Density Surges in Europe, Asia, and Americas” (2024 data). https://ifr.org/news/robot-density-surges-in-europe-asia-and-americas/

[25]  Bloomberg (content sponsored by the Government of Japan). “Japan: innovating the future of industrial robots” (citing the IFR, 2018). https://sponsored.bloomberg.com/article/japangov/japan-innovating-the-future-industrial-robots

C. Humanoids and industrial policy

[26]  TechNode. “China’s AgiBot leads global humanoid robot shipments in 2025, Omdia says” (9 January 2026). https://technode.com/2026/01/09/chinas-agibot-leads-global-humanoid-robot-shipments-in-2025-omdia-says/

[27]  China Daily. Omdia report on general-purpose embodied intelligent robots (10 January 2026). https://regional.chinadaily.com.cn/wic/2026-01/10/c_1155842.htm

[28] Telecoms.com. “Humanoid robot production goes exponential, led by China” (Omdia; 2.6 million units by 2035). https://www.telecoms.com/ai/humanoid-robot-production-goes-exponential-led-by-china

[29]  Forklog. “China leads the robot race” (Counterpoint; NDRC warning; Unitree figures). https://forklog.com/en/china-leads-the-robot-race/

[30] Gasgoo. “Unitree launches STAR Market IPO issuance process” (July 2026). https://autonews.gasgoo.com/articles/news/unitree-launches-star-market-ipo-issuance-process-subscriptions-open-august-10-2083181368883253248

[31]  Caixin. “Robotics startup Unitree launches $620 million STAR Market IPO” (31 July 2026). https://www.caixinglobal.com/2026-07-31/robotics-startup-unitree-launches-620-million-star-market-ipo-102469723.html

[32]  Wikipedia. “Optimus (robot)” (status as of September 2026; secondary source). https://en.wikipedia.org/wiki/Optimus_(robot)

[33]  RoboZaps. “Tesla Optimus Gen 3: release date & price” (status as of 10 August 2026; secondary source). https://blog.robozaps.com/b/tesla-optimus-gen-3

[34]  MERICS. “Great robo-leap forward: MIIT charts path toward humanoid robots” (November 2023). https://merics.org/en/merics-briefs/humanoid-robots-chinas-charm-offensive-towards-foreign-firms-5g

D. Analyst scenarios

[35]  Goldman Sachs Research. “The global market for humanoid robots could reach $38 billion by 2035.” https://goldmansachs.com/insights/articles/the-global-market-for-robots-could-reach-38-billion-by-2035

[36]  Drives & Controls. “Mercedes and BMW explore using humanoid robots” (coverage of the Goldman Sachs report; secondary source). https://m.drivesncontrols.com/news/fullstory.php/aid/7635/Mercedes_and_BMW_explore_using_humanoid_robots_.html

[37]  Morgan Stanley. “Humanoids: a $5 trillion market” (14 May 2025). https://www.morganstanley.com/insights/articles/humanoid-robot-market-5-trillion-by-2050

[38] OODA Loop. “Humanoid robots seen as $5 trillion global opportunity at Morgan Stanley” (secondary source). https://oodaloop.com/briefs/technology/humanoid-robots-seen-as-5-trillion-global-opportunity-at-morgan-stanley/

E. Supply chain

[39]  Project Blue. “China to suspend rare earth export controls for one year.” https://projectblue.com/blue/news-analysis/1356/china-to-suspend-rare-earth-export-controls-for-one-year-

[40] Clark Hill. “China hits ‘pause’ on rare-earth export controls and what it means for supply chains.” https://clarkhill.com/news-events/news/china-hits-pause-on-rare-earth-export-controls-and-what-it-means-for-supply-chains

[41]  SAN. “Elon Musk says China’s rare earth export ban may delay Tesla’s humanoid robot” (May 2025). https://san.com/cc/elon-musk-says-chinas-rare-earth-export-ban-may-delay-teslas-humanoid-robot/

[42]  Morningstar. Nabtesco profile (share of RV reducers; analyst report). https://www.morningstar.com/quotes/0P0000EKF2

F. Standards and regulation

[43]  ISO. ISO 10218-2:2025, “Robotics — Safety requirements — Part 2.” https://www.iso.org/standard/73934.html

[44]  Control Design. “ISO 10218 update makes functional-safety requirements more explicit” (18 February 2025). https://www.controldesign.com/industry-news/news/55268769/iso-10218-update-makes-functional-safety-requirements-more-explicit

[45]  EU-OSHA. Regulation (EU) 2023/1230 on machinery. https://osha.europa.eu/en/legislation/directive/regulation-20231230eu-machinery

[46]  European Union. Regulation (EU) 2024/1689 (AI Act), Official Journal. https://eur-lex.europa.eu/eli/reg/2024/1689/oj

Verification note and limits

•      Cut-off date. Data are verified as of 17 October 2026. Sources were consulted through web search; only the IFR 2026 executive summary was read in full.

•      Humanoids. There is no single official figure. Omdia estimates about 13,000 shipments in 2025; Counterpoint and other consultancies cite figures above 16,000; Unitree claims more than 5,500 deliveries of its own.

•      ABB–SoftBank. The deal was announced in October 2025 with closing expected between mid and late 2026; no confirmation of closing was found.

•      Unitree. The launch of the public offering (July 2026) was verified, not the final outcome of the listing.

•      Tesla. Status is as of September 2026; third-quarter results are expected in late October and may change the picture.

•      Rare earths. Sources differ on the exact expiry date of the suspension (10 or 27 November 2026).

•      Secondary sources. References 10, 14, 18, 25, 29, 32, 33, 36 and 38 are secondary or press sources; they should be checked against primary ones before publication.

•      Facts without an open source in this review. The Roomba’s 2002 launch, the term “Big Four,” KUKA’s place in that group and the ISO 8373 definition in the glossary are commonly used and are cited from memory; no specific source was consulted. The EUR-Lex link (reference 46) is the official identifier, but the page was not opened.

•      Robot density. China’s figure is omitted because IFR editions are not comparable with one another.