Learn With Nathan
Sign in
Full sample chapter · AI From the Ground Up
The Question That Started It All
Before we ask what AI will do to business, we need a useful answer to a more basic question: what exactly are we calling intelligent?
The Question That Started It All
Before we ask what AI will do to business, we need a useful answer to a more basic question: what exactly are we calling intelligent?
At nine on a Monday morning, the leadership team of a fictional insurance company gathers around a conference table. The head of customer service has brought a proposal. A new system can read incoming claims, extract the important facts, draft a response, and send straightforward cases to the right queue. In a demonstration, it completes in seconds what once took an employee several minutes.
The room fills with questions. Is it accurate? What does it cost? Can customer information leak? Who is responsible when it is wrong? Then someone asks a question that sounds deeper than all the others: “Does it actually understand the claim?”
That question changes the mood because it mixes several different concerns. The leaders want to know whether the system can perform a task, whether its process is dependable, whether its output can be trusted, and perhaps whether something inside it has a human-like experience. Those are not the same question. Yet in ordinary conversation, the single word intelligence often stands in for all of them.
This confusion is understandable. We recognize intelligence in people through behavior. A colleague explains a difficult idea, adapts when circumstances change, remembers a customer, notices a contradiction, or invents a new approach. We cannot directly inspect the colleague's inner life; we infer it from action, language, and our shared humanity. When a machine produces similar behavior, the familiar signals arrive without the familiar person behind them. Our instincts have little practice with that combination.
This chapter builds a working vocabulary for the rest of the book. It will not settle the philosophical question of whether a machine could ever think or feel. It will do something more immediately useful: separate what a system does, how it does it, and what—if anything—it experiences. That separation is the beginning of sound business judgment.
A decision hiding inside a word
“Is it AI?” sounds like a classification question, but in business it is usually a proxy for a decision. A buyer may be asking whether a product is genuinely new. A risk officer may be asking whether a special control process applies. An employee may be asking whether a role will change. A customer may be asking whether a human reviewed an outcome. An investor may be asking whether a company has a durable advantage or a fashionable label.
A product can contain AI without being intelligent as a whole. Consider a customer-service platform. One component may classify the topic of a message. Another may retrieve a policy. A language may draft a reply. Conventional software may check required fields, enforce permissions, calculate a deadline, and record the result. A human may approve exceptional cases. The useful unit of analysis is therefore not just “the AI.” It is the complete system: model, data, rules, software, people, policies, and operating environment.
Audiobook description
The figure is a set of nested rectangles. A small center block labeled AI model sits inside a larger product and workflow layer. Around that are data and software, human judgment, governance, and the real-world customer context. The visual point is that a model is only one component of the system a business actually deploys.
Model, application, and system
A model turns inputs into outputs based on patterns or rules. An application packages one or more models into software people can use. An AI system includes the application plus the data, people, processes, controls, and environment that shape what happens in practice.
This system view prevents two common errors. The first is to give the model credit for everything the product does. The second is to blame the model for every failure that occurs around it. A response might be wrong because the model generated a false claim. It might also be wrong because the retrieval system supplied an outdated policy, a database held the wrong customer record, a rule routed the case incorrectly, or the organization removed human review too soon.
When a leader asks, “Is this intelligent?” the better opening question is often, “What decision or action does this system influence, and what evidence would make that use acceptable?” The change in wording is small. The improvement in clarity is large.
Intelligence is a bundle, not a single dial
People speak of intelligence as though every mind had one dial marked from low to high. Real abilities are more uneven. A person can be brilliant at spatial reasoning and ordinary with words. A talented negotiator may forget names. A new employee can understand sarcasm but know little about the company. Human intelligence involves memory, perception, language, learning, planning, social judgment, motor skill, creativity, and more. These abilities interact, but they are not interchangeable.
Machine capabilities are even more jagged. A system may identify patterns in millions of transactions yet fail when a column heading changes. A language model may produce an elegant market analysis and then make an elementary counting error. A vision system may detect tiny manufacturing defects under controlled lighting but misclassify an object after a modest change in angle. Extraordinary performance in one region of the map does not imply general competence everywhere else.
Audiobook description
At the center is the word intelligence. Eight separate nodes orbit it: language, memory, perception, reasoning, planning, creativity, social judgment, and action. Lines connect some abilities but no single outer node contains all the others. The intended message is that intelligence is a constellation of capabilities, not one universal score.
This is why comparisons such as “AI is smarter than humans” are usually incomplete. Smarter at what task? Under what conditions? With access to which tools and data? Measured by speed, accuracy, cost, adaptability, explanation, or consequences? Compared with which people: a novice, a typical worker, or a specialist with twenty years of experience?
A narrow system is built or trained for a bounded purpose. Fraud detection, route optimization, image classification, product recommendation, and speech transcription are familiar examples. A general-purpose model can perform many tasks and can be adapted to uses its builders did not specify one by one. But general-purpose is not the same as generally intelligent. Breadth of useful behavior does not prove human-level flexibility across every domain, and neither breadth nor fluency proves consciousness.
One benchmark is one window
A high score can demonstrate a valuable capability. It cannot, by itself, establish broad understanding. Examine whether the evaluation resembles real work, whether test material appeared in training data, how performance varies across groups and conditions, and what a mistake costs.
Performance, process, and experience
Three questions are often compressed into the phrase “Does it understand?” Pulling them apart gives us a durable framework.
Performance: What can the system do, and how well does it do it? This is an empirical question. We can test whether a system extracts the correct claim number, summarizes a contract faithfully, recommends a useful product, or detects a defect. Performance includes not only average accuracy but also speed, cost, consistency, robustness, and the consequences of error.
Process: What mechanism produces the result? A rule-based program follows explicit instructions. A statistical model estimates patterns from examples. A modern language model repeatedly estimates likely continuations from representations learned during training, while an application may add retrieved documents and external tools. Understanding the process helps us predict failure and choose controls.
Experience: Is there anything it feels like to be the system? Does it possess awareness, intentions, or subjective experience? This is the question of consciousness. A system's use of “I,” emotional language, or self-description does not settle it. Those outputs are observable behavior; consciousness concerns an inner condition that the output alone cannot establish.
Audiobook description
The figure has three columns. Performance asks what the system can do and points to tests and outcomes. Process asks how the result is produced and points to architecture, data, and workflow. Experience asks whether there is subjective awareness and is marked as an open philosophical and scientific question. Arrows between columns are crossed out to show that success in one does not automatically prove the next.
For business purposes, performance and process usually deserve priority because they connect directly to evidence and action. If a system drafts clinical notes, leaders must measure fidelity, privacy, time saved, correction burden, and the effect on care. They must understand how the tool receives data, where information is stored, how the draft is produced, and who approves it. The consciousness question may be fascinating, but it does not remove the organization's responsibility for the system's use.
This is not an argument that consciousness is unimportant. If credible evidence ever showed that an artificial system could suffer or possess morally relevant experience, the consequences would be profound. It is an argument against smuggling that conclusion into a product decision through fluent language or a humanlike interface.
The chatbot that sounds certain
Suppose a support assistant says, “I understand how frustrating this is,” and offers a refund. The phrase may improve the interaction, but it does not prove empathy. The operational questions are whether the policy permits the refund, whether the customer was treated fairly, whether the model used private information appropriately, and whether a human can intervene. The organization remains accountable for those choices.
The moving finish line
AI has a peculiar public-relations problem: when a capability becomes reliable and familiar, people often stop calling it AI. Optical character recognition turns a scanned page into editable text. Spam filters sort unwanted mail. Navigation software finds a route through traffic. Recommendation systems arrange products and entertainment. Each once looked like a striking example of machine intelligence. Success turned many of them into ordinary software.
This habit is sometimes called the AI effect. The label does not mean every conventional program was once AI. It points to a moving cultural boundary. We reserve the word intelligence for abilities that remain surprising, then redefine yesterday's achievement as mere computation. The field advances, and the finish line moves.
Audiobook description
A rising staircase represents public expectations. On lower, completed steps are game playing, speech recognition, and image recognition. Language generation sits near today's higher frontier, while more dependable autonomous action is shown farther ahead. Beneath the staircase, completed capabilities flow into a band labeled everyday infrastructure.
The moving boundary explains why a permanent definition based on a list of tasks will age badly. If we say AI is “software that plays chess,” the definition becomes absurdly narrow. If we say it is “anything computers cannot yet do,” the definition chases novelty rather than describing a technology.
It also creates two opposite business errors. One is to dismiss mature AI because it no longer feels futuristic. A proven forecasting or anomaly-detection system may deliver more value than a dazzling conversational demo. The other is to reward novelty simply because it sits at the frontier. A surprising capability is not automatically a reliable product, and a reliable product is not automatically suitable for a high-stakes decision.
New is not the same as useful
Ask whether a proposed capability improves an outcome that matters after integration, supervision, correction, and change-management costs are included. Novelty can earn attention. Evidence earns deployment.
A working definition for this book
No single definition of artificial intelligence satisfies every scientist, philosopher, regulator, and engineer. That is not a reason to give up. Business readers need a definition that is broad enough to survive changing techniques and precise enough to guide questions.
The National Institute of Standards and Technology describes an as an engineered or machine-based system that, for a given set of objectives, can generate outputs such as predictions, recommendations, or decisions that influence real or virtual environments. The OECD's updated definition uses a similar structure and explicitly includes content among the possible outputs. Both definitions focus on what a system infers from inputs and what its outputs can influence, rather than requiring the system to resemble a person.12
For this book, we will use the following plain-language version:
An AI system
An AI system is an engineered system that uses rules or patterns learned from data to infer an output—such as a prediction, recommendation, decision, or piece of content—that can influence a digital or physical environment.
Audiobook description
The diagram moves left to right. Objectives sit above inputs and both feed an inference process. The process produces four possible output types: prediction, recommendation, decision, and content. Those outputs influence a real or virtual environment. A feedback arrow returns from the environment to monitoring, emphasizing that deployment consequences must be observed.
The definition includes both major traditions we will meet later. In symbolic AI, people encode rules and representations that a machine can manipulate. In , a system learns patterns or parameters from examples. Many useful products combine the two, wrapping a learned model in explicit rules and conventional software.
The definition also shifts attention from imitation to influence. A pricing model does not need to talk like a person to affect livelihoods. A recommendation system does not need a face to shape what millions of people see. A warehouse robot does not need emotions to create a safety hazard. The more an output can affect a person or environment, the more carefully we should examine objectives, evidence, boundaries, and accountability.
UNESCO's ethics recommendation reinforces the human side of this view: responsibility for decisions and actions involving AI must remain attributable to people or existing legal entities, and human oversight should protect human agency rather than becoming a ceremonial checkbox.3 An AI system can participate in a decision. It does not absorb the organization's duty to make that decision responsibly.
Five questions before the label
- Purpose: What objective is the system optimizing, and is that objective a good stand-in for what we truly value?
- Evidence: Under which real conditions does it perform acceptably, and where does it fail?
- Influence: Who or what can be affected by its output?
- Boundaries: What may the system do on its own, and what requires review or must remain prohibited?
- Responsibility: Which person or role owns the outcome, monitoring, escalation, and remedy?
These questions will recur throughout the book. History will show why different technical approaches rose and fell. The conceptual chapters will explain how data, learning, tokens, embeddings, attention, and generation work. The business chapters will turn those ideas into choices about opportunities, risks, controls, and change. We will repeatedly return to the distinction that began in the conference room: capability is not consciousness, fluency is not truth, and a model is not the whole system.
Remember this
Intelligence is not one dial. It is a bundle of capabilities that can be strong, weak, or brittle in different combinations.
Separate performance, process, and experience. Evidence that a system performs a task does not by itself prove humanlike understanding or consciousness.
Evaluate the system, not only the model. Data, rules, interfaces, people, and governance shape the outcome.
Ask about influence and accountability. The practical importance of an AI output depends on what it can affect and who remains responsible.
Audiobook description
A five-point compass surrounds a center labeled accountable action. The five directions are purpose, evidence, influence, boundaries, and responsibility. The diagram summarizes the chapter's practical lesson: leaders do not need to settle the nature of intelligence before they can ask disciplined questions about an AI system.
Five key terms
- AI system
- Model
- Capability
- Consciousness
- AI effect
Reflection questions
- Think of an AI product your organization uses. What parts of the outcome come from the model, and what parts come from data, rules, people, or workflow design?
- When someone says a system “understands,” are they making a claim about performance, process, subjective experience, or some mixture of the three?
- Which familiar technology in your work may rely on mature AI even though nobody calls it AI anymore?