› Free · Mini textbook · 14 chapters
The people behind the ideas
A short book of fourteen figures, in the order they help a parent teach.
AI did not arrive as one invention. It is a chain of people who asked what a machine could do with symbols, feedback, rules, and examples. This book walks that chain. Each chapter keeps three kinds of notes: what they contributed, what they noticed, and what to take away.
How to read this
- Read one chapter, then send that person's stop page (`/figures/<slug>`) to a child or student.
- The order is the walk — foundations, then information and feedback, then symbols, then networks and data.
- Side doors into free Hi, Bot activities appear on some stops. The book does not replace them.
Prefer one screen at a time? Start the figure walk →
Chapter 1 · Foundations of Computing
Ada Lovelace
1815–1852 · British
Realized Babbage's Analytical Engine could manipulate symbols, not only numbers.
Ideas: First computer program, general-purpose computing. Affiliations: Independent / associate of Charles Babbage.
Contribution
- Wrote what is often called the first computer program — notes for how Babbage's Analytical Engine could compute Bernoulli numbers.
- Saw that a calculating machine could also follow rules about symbols, music, and other non-numeric patterns.
- Published her analysis as notes on Menabrea's account of the Engine, expanding it into a vision of general-purpose computing.
Key observations
- A machine that only adds numbers is useful; a machine that manipulates symbols by rule can represent many kinds of work.
- Imagination about what a machine might do can outrun the hardware that exists in your lifetime.
- Careful notes — steps, conditions, expected results — are how you hand a plan to a future machine.
Conclusions
- Programming is older than electronic computers; it begins as precise instructions for a general engine.
- When someone says AI "understands," ask what symbols and rules (or examples) it was given.
- General-purpose machines matter because the same substrate can serve many jobs.
Chapter 2 · Foundations of Computing
Alan Turing
1912–1954 · British
Formalized computation and algorithms, and proposed a practical test for machine intelligence.
Ideas: Turing Machine, Turing Test. Affiliations: University of Cambridge, Bletchley Park, University of Manchester.
Contribution
- Defined the Turing machine — a simple model that can run any algorithm a digital computer can run.
- Helped break wartime ciphers at Bletchley Park, showing that systematic computation could change real outcomes.
- Proposed the Imitation Game (the Turing Test) as a practical way to talk about machine intelligence.
Key observations
- Computation is not a particular brand of box; it is a class of step-by-step procedures.
- Some questions about "thinking" get clearer if you ask what a machine would have to do to fool a careful observer.
- Limits matter too: some problems cannot be decided by any algorithm.
Conclusions
- When people debate whether AI is "really intelligent," they are often still arguing in Turing's frame — behavior under a test.
- Teach algorithms as procedures anyone (or anything) can follow, not as magic inside a gadget.
- Knowing what computation can and cannot decide is part of AI literacy.
Chapter 3 · Foundations of Computing
John von Neumann
1903–1957 · Hungarian-American
Architected the modern computer structure and pioneered game theory used in AI decision-making.
Ideas: von Neumann architecture, game theory. Affiliations: Institute for Advanced Study, Princeton.
Contribution
- Helped shape the stored-program computer design still called the von Neumann architecture.
- Co-founded game theory, giving a mathematical language for strategies, payoffs, and rational play.
- Worked across math, physics, and early computing at the Institute for Advanced Study.
Key observations
- Code and data can live in the same memory — a machine that rewrites its own instructions becomes far more flexible.
- Decision-making under conflict can be modeled as games with rules and scores.
- Architecture choices (how memory, control, and arithmetic connect) shape what software can do later.
Conclusions
- Today's computers still echo his layout: a processor, memory that holds both programs and data, and input/output.
- AI "decision-making" often borrows game-theory language — agents, rewards, opponents.
- Hardware design and AI ideas grow together; neither is only software.
Chapter 4 · Information Theory
Claude Shannon
1916–2001 · American
Laid the mathematical foundation for digitizing information, and pioneered computer chess and search.
Ideas: Information theory, bits. Affiliations: Bell Labs, MIT.
Contribution
- Founded information theory — a math of how much surprise a message carries, measured in bits.
- Showed how to send signals reliably over noisy channels with encoding and redundancy.
- Built early experiments in computer chess and search, treating play as a tree of choices.
Key observations
- Information can be measured separately from meaning — bits are about possibility, not understanding.
- Noise is normal; good systems plan for errors instead of pretending channels are perfect.
- Search through move trees is a concrete way machines can "look ahead."
Conclusions
- Digital life — files, networks, compression, encryption — sits on Shannon's bit.
- When AI models "compress" language into patterns, they are still in the business of representing information efficiently.
- Chess and games remain a classroom for search, evaluation, and limits of lookahead.
Chapter 5 · Cybernetics
Norbert Wiener
1894–1964 · American
Founded the study of control and communication in animals and machines.
Ideas: Cybernetics, feedback loops. Affiliations: MIT.
Contribution
- Named and founded cybernetics — the study of control and communication in animals and machines.
- Put feedback loops at the center: sense → compare → correct → sense again.
- Warned early that powerful automation would reshape work, war, and society.
Key observations
- A thermostat and a nervous system share a pattern: they steer by measuring error.
- Communication without control, or control without sensing, fails in living and engineered systems alike.
- Ethics belongs beside engineering when machines amplify human action.
Conclusions
- Modern AI training (adjust weights when you are wrong) is a feedback story Wiener would recognize.
- Ask of any AI system: what does it sense, what does it compare to, and who sets the target?
- Technical power without social foresight was already a cybernetics warning.
Chapter 6 · Symbolic AI
John McCarthy
1927–2011 · American
Organized the 1956 Dartmouth conference and invented the dominant AI programming language of the following decades.
Ideas: The term Artificial Intelligence, Lisp. Affiliations: Stanford University, MIT.
Contribution
- Coined the name Artificial Intelligence and co-organized the 1956 Dartmouth conference that launched the field.
- Invented Lisp, the dominant language for symbolic AI for decades.
- Pushed for machines that could reason with formal knowledge, not only calculate numbers.
Key observations
- Naming a field focuses attention, funding, and shared problems.
- A language shaped for lists and recursion makes symbolic reasoning easier to write.
- Early AI optimism assumed intelligence was mostly searchable rules and representations.
Conclusions
- "AI" is a research program with many methods — not one product.
- Symbolic approaches (rules, logic, knowledge graphs) still matter beside learning from data.
- When you hear Dartmouth or Lisp in history, you are hearing the birth certificate of the field's name.
Chapter 7 · Symbolic AI
Marvin Minsky
1927–2016 · American
Co-founded the MIT AI Lab and championed symbolic AI, robotics, and cognitive science.
Ideas: Society of Mind, SNARC. Affiliations: MIT.
Contribution
- Co-founded the MIT AI Lab and built early hardware such as SNARC, a neural-net learning machine.
- Championed symbolic AI, robotics, and cognitive science as one intertwined project.
- Wrote The Society of Mind — intelligence as many simple processes cooperating.
Key observations
- Mind may not be one algorithm; it may be a society of specialized parts.
- Labs that mix hardware, software, and psychology invent problems nobody's syllabus listed.
- Strong opinions about the "right" path for AI can both drive progress and miss alternatives.
Conclusions
- Today's multi-agent and modular systems rhyme with Society of Mind.
- Robotics keeps AI honest — ideas must survive contact with the physical world.
- A healthy field holds more than one research tradition at once.
Chapter 8 · Symbolic AI
Allen Newell
1927–1992 · American
Co-created early programs that treated intelligence as symbol manipulation.
Ideas: Logic Theorist, General Problem Solver. Affiliations: Carnegie Mellon University, RAND Corporation.
Contribution
- With Herbert Simon, built Logic Theorist and the General Problem Solver — early programs that searched for proofs and plans.
- Treated intelligence as the manipulation of symbols according to rules.
- Helped establish Carnegie Mellon as a center for AI and cognitive science.
Key observations
- If you can write the rules of a puzzle, a machine can search the space of moves.
- Human problem-solving often looks like heuristics — shortcuts that usually work — not endless brute force.
- Psychology and computer science can share one model of mind-as-process.
Conclusions
- Search + representation is still a core AI pattern (pathfinding, planning, theorem proving).
- "Thinking as symbol crunching" is one powerful lens — not the only one.
- Early AI proved machines could do intellectual labor once thought uniquely human.
Chapter 9 · Symbolic AI
Herbert A. Simon
1916–2001 · American
Collaborated with Newell on the first AI programs, and won the Nobel Memorial Prize in Economics for decision-making.
Ideas: Logic Theorist, General Problem Solver, bounded rationality. Affiliations: Carnegie Mellon University.
Contribution
- Collaborated with Newell on Logic Theorist and GPS, among the first AI programs.
- Won the Nobel Memorial Prize in Economics for work on decision-making and bounded rationality.
- Bridged organization science, psychology, and artificial intelligence.
Key observations
- People (and machines) rarely optimize perfectly; they satisfice — choose something good enough under limits.
- Attention and time are scarce, so smart systems need heuristics.
- Institutions and algorithms both shape what decisions get made.
Conclusions
- AI literacy includes limits: compute, data, and human attention are all bounded.
- When a model "optimizes," ask what it was scored on — and what it ignored.
- Economics, psychology, and AI share one question: how do agents decide under constraint?
Chapter 10 · Connectionism
Frank Rosenblatt
1928–1971 · American
Invented the first artificial neural network that could learn from trial and error, initially in hardware.
Ideas: The Perceptron. Affiliations: Cornell Aeronautical Laboratory, Cornell University.
Contribution
- Invented the Perceptron — an early artificial neural network that learned from trial and error.
- Built perceptron hardware, not only equations on paper.
- Showed that a simple network could classify patterns it had not been hard-coded to recognize.
Key observations
- Learning can mean adjusting connection strengths when the answer is wrong.
- Hardware demos persuade when slides do not.
- Simple models hit walls — some patterns need deeper or differently structured networks.
Conclusions
- Modern deep learning is a descendant of this idea: learn weights from examples.
- Hi, Bot's Weight Pond is a hands-on echo of the perceptron lesson.
- Connectionism and symbolic AI competed for decades; both left tools we still use.
Chapter 11 · Deep Learning
Geoffrey Hinton
1947–present · British-Canadian
Championed neural networks through AI winters, leading breakthroughs in deep learning. Godfather of AI.
Ideas: Backpropagation, Boltzmann machines. Affiliations: University of Toronto, Google.
Contribution
- Kept neural-network research alive through "AI winters," refining learning methods such as backpropagation and Boltzmann machines.
- Helped lead the deep-learning breakthroughs that remade vision and speech in the 2010s.
- Shared the 2018 Turing Award with Bengio and LeCun — often called a "Godfather of AI."
Key observations
- Ideas can look dead in a field and still be right — they need compute, data, and patience.
- Stacking layers lets networks learn features of features, not only flat patterns.
- Researchers who train powerful systems also carry a duty to talk about risk.
Conclusions
- Deep learning's success was scientific persistence plus scale, not a sudden miracle.
- Backpropagation — sending error backward through layers — is the workhorse of modern nets.
- Celebrate capability and ask hard safety questions in the same breath.
Chapter 12 · Deep Learning
Yoshua Bengio
1964–present · Canadian
Made fundamental contributions to deep learning architectures and natural language processing. Godfather of AI.
Ideas: Word embeddings, language modeling. Affiliations: University of Montreal (Mila).
Contribution
- Made foundational contributions to deep architectures and representation learning.
- Advanced neural language modeling and word embeddings — how networks map words into usable geometry.
- Built Mila in Montreal into a major deep-learning research center. Godfather of AI.
Key observations
- Language can be treated as vectors in space where "nearby" means "related in use."
- Deep models learn intermediate representations that transfer across tasks.
- Scientific communities (labs, open papers, students) multiply individual insight.
Conclusions
- Today's language models sit on decades of representation-learning work.
- Embeddings are a classroom idea: meaning as position and distance.
- Localization of talent (cities, institutes) shapes who gets to build the next wave.
Chapter 13 · Deep Learning
Yann LeCun
1960–present · French-American
Developed convolutional neural networks for vision, document recognition, and pattern recognition. Godfather of AI.
Ideas: Convolutional neural networks. Affiliations: New York University, Meta.
Contribution
- Developed convolutional neural networks (CNNs) that exploit local patterns in images and signals.
- Applied CNNs to document recognition and other pattern tasks long before the 2012 boom.
- Continues to argue for architectures that learn world models, not only next-token prediction. Godfather of AI.
Key observations
- Images have structure — nearby pixels belong together — and architecture should respect that.
- Weight sharing (the same filter reused across the image) cuts parameters and improves generalization.
- Supervised classification is powerful; unsupervised and predictive learning of the world may be the next leap.
Conclusions
- CNNs made computer vision practical for handwriting, photos, and medical images.
- Inductive bias — baking in assumptions about structure — still matters beside "just add data."
- Debates about the right future architecture are part of a living field, not settled dogma.
Chapter 14 · Deep Learning
Fei-Fei Li
1976–present · Chinese-American
Created the labeled dataset that showed deep neural networks working at scale in 2012.
Ideas: ImageNet. Affiliations: Stanford University.
Contribution
- Led ImageNet — a massive labeled image dataset that let researchers train and compare vision models at scale.
- Enabled the 2012 deep-learning surge when networks trained on ImageNet shocked the field with accuracy jumps.
- Advocates for human-centered AI: capability paired with care for people and society.
Key observations
- Data quality and labeling can unlock methods that looked weak on tiny sets.
- Benchmarks focus a community — shared tests make progress visible and competitive.
- Who appears in the photos, and who labels them, shapes what the model "sees" as normal.
Conclusions
- Scale of carefully labeled examples was as important as clever algorithms.
- When you evaluate AI, ask what dataset it was graded on — and what that dataset left out.
- Building AI responsibly includes the people in the data, not only the accuracy score.
Keep walking
These fourteen people do not cover every contributor. They cover a parent-sized path: from the idea that machines can follow rules about symbols, to the idea that machines can learn from labeled examples at scale. The next step is still the same — open one stop, send the link, and keep walking.
Canonical book URL: hibot.space/figures/book