Welcome. This deck is a 20-slide tour of the whole book: the central idea,
how the 20 chapters are organized, and what a reader can do after finishing.
Each slide maps to one or two chapters. Speaker notes carry the detail.
Chapter 1. Graduate research in EE and computer design is a chain of repeatable
tasks with checkable outputs. The pipeline is only as trustworthy as its weakest
stage. AI absorbs much of the procedural layer; what it cannot supply is the
question, the domain judgment, and the accountability.
The 20:10 ratio reflects that plans are cheap to discard while executions are
expensive to check. The number of agents matters least; dependencies, review
cost, and handoff quality set what the team can deliver (Chapter 10).
The course deliberately does not cover training LLMs, ML theory, AI chip design,
vendor EDA flows, or legal advice. It assumes students bring their own EE domain
knowledge and teaches how to direct AI at it.
Every chapter ships with annotated references and a quiz. The book also has a
glossary of 379 terms, a learning graph with 379 concepts, 48 interactive
MicroSims, 20 infographic posters, and two appendices (Personal Agents and
Personal Knowledge Graphs).
Chapter 2 explains tokens, context windows, training data, knowledge cutoffs,
and hallucination. Chapter 3 turns that into choices: API access (keys are
secrets), cost and rate limits, when a local open-weight model is the right
answer, and the first prompting techniques.
Reliability comes less from clever wording than from complete specification,
selected and grounded context, checks between steps, and tested, versioned
prompts. Defining domain vocabulary (resolution, gain, noise) prevents silent
misreadings.
An agent is only as reliable as the project it works in. Plain text, a short
checkable instruction file, a predictable layout, and version history give
memory, boundaries, and recovery. A knowledge graph with typed links can be
queried to ground an agent's answers; Appendix B covers building one.
A skill is worth the effort when a task recurs and the result can be checked.
Chapter 7 covers trigger tests including near misses, behavior tests with known
answers, evaluation of pass rate across repeated runs, and a research skill
template.
An agent's power and its risk are the same quantity: what its tools can reach.
Early errors compound across loop iterations, so version control is the safety
net. Setting limits is what lets you use agents with confidence.
Supervision is the design of where a human looks. Each failure mode has a
mechanical defense. Roles make a group of agents legible.
Orchestration is the design of how work, checks, and decisions flow so that your
attention prevents the most damage. Span of control and review load limit the
useful team size; the MicroSims on critical path and review load let students
experiment with both.
AI changes the speed of every stage of the literature workflow and none of the
standards. A fluent fake citation looks identical to a real one, so verification
is a required step, not an optional one.
Example from Chapter 13: a 10 versus 20 log10 mistake turned a 10-bit converter
into an 18-bit one. Code from an agent is a hypothesis about how to compute a
result; the chapter is how you test that hypothesis.
Chapter 14 separates divergent and convergent idea phases and ranks ideas in a
table with assumptions and failure conditions. Chapter 15 follows a paper from
outline through LaTeX, figures, references, venue rules, and arXiv submission.
Argument-bearing sections stay human-written.
This book contains 48 MicroSims that students use to sort, predict, calculate,
and judge. Chapter 16 shows how to generate one from a precise specification,
and the same approach extends to intelligent textbooks and slide decks like
this one.
Agreement between models is weak evidence; disagreement is a flag. Verification
is not a final stage but a habit that runs through the whole workflow. AI raises
the volume of output and adds failure modes that sound plausible, so every check
matters more.
Unpublished results pasted into a hosted tool are disclosed to a third party.
Keep a log as you work so the disclosure statement is accurate. The chapter ends
with drafting a responsible-use policy for a research group.
Products change monthly, so Chapter 19 teaches durable criteria: blast radius,
usage metering, data exposure, vendor-claim verification, and overnight activity
audits. Chapter 20 applies the whole toolkit; each case records cause,
detection, control, and change for what failed.
Close by pointing to the quizzes, MicroSims, and posters for self-study, and to
the capstone chapter for an adoption plan. Invite questions.