The GenAI Syllabus
A six-module program I put together to build real working knowledge of generative AI in a business context — foundations, commercial value, risk, opportunity-spotting, delivery, and what's next.
Every source below is one I'm actually working through, not a reading list assembled for show. Progress saves in this browser as I go.
Before week 1
Free accounts on at least one frontier chat model, to do the hands-on parts properly — the free tiers cover this whole plan.
- claude.ai — free tier
- chatgpt.com — free tier, needed for Module 5
- gemini.google.com — optional, worth having for comparison
Module 5's hands-on courses run in notebooks against the OpenAI API, so a free OpenAI API key is worth having too — a few pence of usage covers the whole course.
Modules
Introduction to Generative AI
Foundations: what generative AI and LLMs actually are, how they differ from earlier ML, and the interaction-design challenges of working with them.
- courseGoogle Cloud — Generative AI Learning Path (free, non-technical intro, self-paced)
- video3Blue1Brown — "But what is a neural network?" and the follow-on transformer/attention videos in the same series (~1.5h for the set)
- videoAndrej Karpathy — "Intro to Large Language Models" (1h, the best plain-English mental model of how LLMs work)
- courseAndrew Ng — "Generative AI for Everyone" (Coursera, ~6h across 3 weeks, free to audit)
Can you explain, without jargon: the difference between AI, machine learning, and generative AI · why an LLM sometimes states wrong things confidently (hallucination) · what a "token" and a "prompt" actually are?
How Generative AI Adds Value to Business
The path from research lab to commercial deployment, and case studies of where GenAI has actually moved the needle for organisations.
- readingMcKinsey — "The economic potential of generative AI" (skim the sector breakdowns)
- blogEthan Mollick — One Useful Thing (Wharton professor, free newsletter — read 4–5 back-posts on real workplace use cases)
- readingHBR — "Generative AI Will Change Your Business. Here's How to Adapt."
- practicePick 3 companies in financial or professional services and write one paragraph each on where they've deployed GenAI and what it replaced or sped up
I'm in Transaction Services, so the plan is to map this onto due diligence and reporting workflows I already know — the real test is where GenAI would save hours vs. where it's just noise.
Name one process at your own firm where GenAI would save real hours, one where it would introduce more risk than it removes, and explain the difference in one sentence.
Managing Risks and Ethical Considerations
Data security, legal liability, bias, and emerging threats like deepfakes — the risk side of deployment.
- talkHenry Ajder — "Mapping the Synthetic Age" — a sharp, well-sourced talk on deepfakes and the erosion of digital trust
- frameworkNIST AI Risk Management Framework (skim the "Govern" and "Map" functions)
- readingICO — Guidance on AI and Data Protection (UK-specific, directly relevant to the day job)
- readingUK Government — "A pro-innovation approach to AI regulation" white paper (skim, not full read)
List three ways confidential client or deal data could leak through an AI tool at work, and one control for each. If you can't name three, re-read the ICO guidance.
Identifying Opportunities
A systematic way to spot where GenAI fits — analysing a system, defining the problem properly before reaching for a model.
- bookPrediction Machines by Agrawal, Gans & Goldfarb — the free HBR summary covers the core "AI as cheap prediction" idea
- courseAndrew Ng — "AI For Everyone" (audit free — the opportunity-spotting and workflow-mapping sections are the useful part here)
- practiceAudit one of my own projects for manual steps, and flag which are prediction, classification, or generation problems
Take one workflow and classify it: is the bottleneck a prediction problem, a generation problem, or neither (and therefore not a GenAI opportunity at all)? Being able to say "neither" is the actual skill here.
Delivering Solutions
Building and shipping: development approach, team dynamics, and turning a GenAI idea into something that creates organisational value.
- courseDeepLearning.AI — "ChatGPT Prompt Engineering for Developers" (free, ~1h40)
- courseDeepLearning.AI — "Building Systems with the ChatGPT API" (free, ~2h)
- courseAnthropic Academy — Anthropic's own course catalog on building with Claude, prompting and tool use; no account needed to browse
- practiceShip one small GenAI-powered feature into an existing build — even a single API call that improves one workflow counts as "delivered"
This is the module to actually build in. I already ship my own apps, so the plan is to use this week as a sprint to bolt a genuinely useful GenAI feature onto one of them.
Ship something a user could actually click, not just a prompt in a playground — "delivered" means the first one.
Future Directions of AI
Current limits, unresolved problems, and what's coming next — and how to keep this knowledge from going stale in six months.
- reportStanford HAI — AI Index Report (free annual report; read the executive summary)
- newsletterThe Batch (DeepLearning.AI's weekly, free — subscribe, this replaces "staying current")
- bookCo-Intelligence by Ethan Mollick — the most readable synthesis of where this is heading for knowledge work
- practiceWrite a one-page note: which assumptions from Module 1 already look dated, and what I'd tell myself six months ago
Name one current model limitation that a well-designed business process has to route around rather than trust — and how you'd design that guardrail.
Capstone
Reading isn't the same as knowing, so the plan is to close out the six modules by taking one real opportunity through the full pipeline — on paper and in code.
- Frame it — one paragraph on the business problem, using the "is this actually a prediction or generation problem" test from Module 4.
- Risk-check it — the three questions from Module 3's self-check: what data does it touch, what happens if the model is wrong, who's accountable if it is.
- Build a thin slice of it — the smallest version that's actually usable, not a demo. Reuse whatever came out of Module 5.
- Write the one-pager — what it does, what it cost in time, what it would take to put in front of colleagues or clients.
A working artifact says more than a certificate ever could.
Glossary
The terms worth knowing cold before the rest of this makes sense.
- AI
- Any system that performs tasks we associate with human intelligence — a huge umbrella that includes far more than chatbots.
- Machine learning
- Systems that improve at a task from data, rather than being explicitly programmed rule by rule.
- Generative AI
- ML models that produce new content — text, images, audio, code — rather than just classifying or predicting a number.
- Foundation model
- A large model trained on broad data (GPT, Claude, Gemini) that gets adapted to many downstream tasks rather than built for one.
- LLM
- Large language model — a foundation model trained mainly on text, which is what you're actually talking to in ChatGPT or Claude.
- Transformer
- The neural network architecture (2017) that made modern LLMs possible, via a mechanism called "attention."
- Token
- A chunk of text (roughly ¾ of a word) that a model reads and generates one at a time — what you're actually paying for on a metered API.
- Prompt engineering
- Structuring your input to reliably get the output you want — less "magic words," more applied problem specification.
- Fine-tuning
- Further training a foundation model on your own narrower data so it specialises in a task or tone.
- RAG
- Retrieval-augmented generation — giving a model your own documents to reference at answer-time, instead of (or alongside) fine-tuning.
- Hallucination
- A confident, fluent, wrong answer — the model isn't lying, it's doing what it always does (predicting plausible text) without a ground-truth check.
- Agent
- A model given the ability to take actions — call tools, browse, run code — in a loop, rather than just answer once.
- Multimodal
- A model that handles more than one type of input/output — text, image, audio — in the same system.
Suggested cadence
| Week | Focus | Time |
|---|---|---|
| Wk 0 | Set up free accounts on the model providers below. Ten minutes, not a real week. | ~1h |
| Wk 1 | Module 1 — foundations. Don't skip this even if some feels familiar; it's the vocabulary the rest builds on. | ~6h |
| Wk 2 | Module 2 — business value, with the applied write-up on Transaction Services workflows. | ~6h |
| Wk 3 | Module 3 — risk and ethics. | ~5h |
| Wk 4 | Module 4 — opportunity mapping, applied to one of my own projects. | ~5h |
| Wk 5 | Module 5 — the build sprint. Heaviest week, budget extra time. | ~7h |
| Wk 6 | Module 6 — future directions, plus subscribing to 1–2 ongoing sources so this doesn't go stale. | ~4h |
| Wk 7 | Optional — the capstone below. Where the plan actually pays off. | ~3–4h |
This page documents the sources and structure I'm using to build working knowledge of generative AI in a business context — not a syllabus borrowed from anywhere else. The checkboxes above save in this browser only; they won't follow you to another device.