Learn With Nathan
Sign in
Free Learn With Nathan Academy book
AI From the Ground Up
A plain-language journey from the birth of artificial intelligence to the models transforming business today.
A practical foundation for business readers, from the history of artificial intelligence through transformers, agents, governance, and responsible action.
The complete book is free. No fee or credit card required.
- Author
- Nathan Thiyagarajah
- Chapters
- 37
- Reading time
- About 12 hours
Book contents
Register to read all chapters for free.
Full sample chapters are available below without an account. Register free to read every chapter in this book and access the complete Academy. No fee. No credit card required.
-
01
The Question That Started It AllBefore we ask what AI will do to business, we need a useful answer to a more basic question: what exactly are we calling intelligent?Read sample
-
02
Old Dreams, New MachinesLong before artificial intelligence became a research field, people imagined artificial beings, built lifelike mechanisms, formalized logic, and discovered that one machine could follow many programs.Read sample
-
03
Can a Machine Think?Alan Turing replaced an argument over an undefined word with an observable game—a move that made progress measurable without making philosophy disappear.Read sample
-
04
The Map of AIAI is not one technology arranged on a ladder from simple to smart; it is a landscape of goals, methods, model families, and systems that frequently overlap.Read sample
-
05
1956: AI Gets Its NameA short proposal and a loosely organized summer gathering did not invent machine intelligence—but they gave scattered ambitions a name, a research agenda, and the beginnings of a shared field.Read free in Academy
-
06
The 1960s: When Anything Seemed PossibleEarly programs could prove, converse, learn a boundary, manipulate a miniature world, and plan a robot's actions—achievements whose controlled settings were both the source of their power and the seed of later disappointment.Read free in Academy
-
07
The 1970s: The First AI WinterThe early systems had not been illusions, but their success in small worlds did not scale on schedule; funding and confidence cooled when technical limits met promises made at a much larger scale.Read free in Academy
-
08
The 1980s: Experts in a BoxExpert systems made AI commercially useful by encoding specialist judgment as rules—but every new exception increased the cost of keeping that expertise correct.Read free in Academy
-
09
The 1990s: Learning from DataAs hand-written knowledge became costly and brittle, more AI systems began estimating patterns from examples, probabilities, and feedback—changing both what a model knew and how its performance could be measured.Read free in Academy
-
10
The 2000s: The Web Becomes a Training GroundThe internet turned human activity into machine-readable traces while clusters, programmable GPUs, and shared datasets made it possible to learn from them at a scale no laboratory could manufacture alone.Read free in Academy
-
11
The 2010s: Deep Learning Wakes UpNeural-network ideas matured into decisive systems when large datasets, parallel hardware, improved training methods, and shared benchmarks finally worked together.Read free in Academy
-
12
The 2020s: From Prediction to GenerationLarge pretrained models changed AI from a collection of mostly task-specific predictors into adaptable engines that can generate language, code, images, audio, and actions through a common interface.Read free in Academy
-
13
Data: Experience in Machine-Readable FormA dataset is not reality poured into a computer; it is a designed record of what an organization chose, managed, or happened to capture—and every omission can shape the model that follows.Read free in Academy
-
14
Training: Turning Examples into a ModelTraining is an iterative search for adjustable settings that reduce defined errors on examples while preserving performance on relevant cases the model has never seen.Read free in Academy
-
15
Three Ways to LearnLearning methods differ primarily in the feedback available: correct answers, hidden structure, missing pieces created from the data itself, or rewards that arrive through action.Read free in Academy
-
16
Neural Networks Without the Neuroscience MythA neural network is a layered mathematical system that learns adjustable transformations—not a digital brain, a mind, or proof that a machine understands.Read free in Academy
-
17
Good Scores, Bad DecisionsA model score becomes useful only when it is translated into real outcomes, unequal costs, operating conditions, affected groups, and a decision policy.Read free in Academy
-
18
Language Becomes TokensA language model does not receive words directly: a tokenizer divides text into reusable pieces, assigns each piece an ID, and fits only a finite sequence into the model’s working context.Read free in Academy
-
19
Meaning Becomes GeometryEmbeddings turn items into learned coordinate lists so that useful relationships can be expressed through distance and direction—without reducing meaning to a literal two-dimensional map.Read free in Academy
-
20
Attention: Finding What Matters NowAttention lets each token construct a context-sensitive update by weighing information from other permitted positions—but the weights are routing instructions, not a transparent transcript of thought.Read free in Academy
-
21
The Transformer Assembly LineA transformer repeatedly routes information among token positions, transforms each position, preserves residual paths, and normalizes the evolving representations before a task-specific head turns them into an output.Read free in Academy
-
22
Learning to Predict the Next TokenA decoder language model learns a probability distribution for the next token from preceding context; repeated prediction can produce rich behavior, but fluency, factual support, and a decoding choice remain different things.Read free in Academy
-
23
From Base Model to Helpful AssistantA useful assistant is a layered system: a pretrained model shaped by adaptation, runtime instructions, trusted context, tools, state, safety controls, and an interface with accountable owners.Read free in Academy
-
24
Why Models Make Things UpA fluent answer can be unsupported. Treat factual reliability as a property of the whole evidence and verification system, not as a personality trait of the model.Read free in Academy
-
25
Images from NoiseDiffusion models learn a route from corrupted examples back toward the patterns of images, then follow that route from random noise under guidance from text or other conditions.Read free in Academy
-
26
AI That Sees, Hears, and SpeaksMultimodal systems translate different signals into representations that can be compared, combined, and transformed—but every translation can discard context or introduce error.Read free in Academy
-
27
Retrieval, Tools, and AgentsAn agent is not merely a model with a grand title. It is a bounded control loop that gathers evidence, proposes actions, observes results, manages state, and stops under rules owned by people.Read free in Academy
-
28
The Boundaries of Today’s AICapability has a coastline: dependable ground for some tasks, tidal flats where conditions matter, and uncharted water where confident claims outrun the evidence.Read free in Academy
-
29
Find the Work, Not the HypeThe useful unit of AI strategy is not a job title or a product demo. It is a measurable task inside a real workflow, with known inputs, consequences, owners, and evidence.Read free in Academy
-
30
Build, Buy, or Configure?Choose the least complicated solution that can meet the task, evidence, control, and operating requirements—and preserve the ability to change course.Read free in Academy
-
31
From Pilot to ProductionA pilot is not a small launch. It is a bounded instrument for learning whether a system, its workflow, and its controls deserve wider exposure.Read free in Academy
-
32
Risk Has Many FacesAI risk is not one score or one department’s checklist. It is a set of possible consequences, distributed among people and institutions, that must be described, owned, treated, and revisited.Read free in Academy
-
33
Keep a Human in the Right PlaceHuman oversight works only when a capable person has the information, time, authority, independence, and alternative needed to change what happens.Read free in Academy
-
34
Lead the Human ChangeAI adoption is the continuing redesign of work, roles, learning, incentives, and responsibility—not the installation of a tool.Read free in Academy
-
35
The Futures We Can See from HereThe future is not one line waiting to be revealed. Scenarios help you prepare for several plausible paths without confusing confidence with foresight.Read free in Academy
-
36
The Questions Society Must AnswerAI does not distribute its benefits, costs, voice, power, and risks by itself. Institutions and public choices do.Read free in Academy
-
37
Your Working Philosophy of AIWrite down how you will choose, test, use, challenge, and stop AI—and make those commitments visible in your decisions.Read free in Academy