Название: Seven Layers of Intelligence: AI, Machine Learning, Deep Learning, Generative AI, LLMs, RAG and Agentic AI
Автор: Sanjeev Vaishya
Издательство: Independently published
Год: 2026
Язык: английский
Формат: pdf, epub
Размер: 13.2 MB

Seven words. Seven different things.

AI. Machine learning. Deep Learning. Generative AI. Large language models. RAG. Agentic AI.

They get thrown around as if they were synonyms for the same new thing. They are not. Each one sits inside the one before it, and each buys a capability at a price.

Treat them as interchangeable and you pay for it. You pay in money, when you buy the most expensive layer of the stack to solve a problem that a cheaper layer solves better and faster. You pay in credibility, when you promise the board an autonomous agent and ship a chatbot that invents the refund policy. And you pay in time — months spent building the wrong layer, discovering the mismatch in user testing, and starting again.

Seven Layers of Intelligence is a field guide to choosing correctly. It is written for the person who has to make the decision and be accountable for it: the architect, the engineering lead, the product manager, the CTO of a company small enough that the CTO also writes code.

What's inside:

A four-question decision framework that tells you, in about two minutes, which layer a requirement actually needs — and which layer someone is proposing.
A reference architecture for systems that use several layers at once, designed so that models, prompts and knowledge bases stay independently replaceable.
The real economics. Cost per decision differs across the seven layers by a factor in the tens of thousands. Where the money goes, what routing and semantic caching actually save, and when self-hosting starts to make sense.
The failure catalogue. Data leakage, distribution mismatch, chunking through a table, hallucinated dosages, prompt injection into a tool-using agent — what each one looks like, what it costs, and how it was fixed.
Governance you can take to a board, including the six questions executives are really asking about AI risk, and how to answer them.
Twenty-two original diagrams, a glossary, a one-page decision checklist, and ten questions to ask any AI vendor.

Part I covers the foundation: AI, Machine Learning and Deep Learning. If you already know the difference between Supervised Learning and a neural network, skim it — but read the sections on where each still wins, because the most common architectural error of the current era is reaching for a language model when a gradient-boosted tree would have been faster, cheaper and more accurate.
Part II covers the generative turn: Generative AI, large language models, RAG and agents. This is where most of the current confusion lives and where most of the money is being spent.
Part III is the part I would read first if I were you. It contains the decision framework, the reference architecture, the cost model, the governance chapter, and the full FieldLedger case study. It is where the seven layers stop being a taxonomy and start being a set of choices.
The appendices are meant to be photocopied and taken into meetings.

Built on one real system, end to end:

Abstract explanations of AI concepts are easy to nod along to and impossible to apply. So this book carries a single worked example from first page to last: a working agricultural platform serving five thousand smallholder farmers, which genuinely needs all seven layers — in different places, for different reasons — and whose unit economics leave no room to get any of them wrong.

It is deliberately unglamorous. It has a supply chain, a compliance regime, intermittent connectivity, and users who cannot read English. If a way of thinking about AI survives contact with it, it will survive contact with your problem.

The argument

The industry's incentives point toward the top of the stack. Vendors sell it. Conference talks feature it. Boards ask about it. Against that, the discipline of writing a rule when a rule will do, of shipping a gradient-boosted tree when it beats the neural network, of stopping at classification when no new content is needed — that discipline is what separates systems that work from systems that demo.

Ask one question of every requirement: what is the smallest layer that solves this correctly? This book shows you how to answer it.

Contents: