32 standalone questions on the map — AI, ML, neural networks, deep learning, GenAI, LLM, GPT and ChatGPT, and where the usual diagram goes wrong.
Audience: absolute beginners. No maths, no code, no prior exposure assumed.
Source: the nested-circles diagram, the "AI terminology" ten-level diagram, and the onion — reconciled, including the two errors that appear in most versions of these pictures.
How to use this: every question stands alone. Pick an option, then read the answer. The sections work outward-in through the rings, then deal with what happens after deep learning.
A · Nested, not competing
Q1Someone asks whether ChatGPT is AI, or machine learning, or generative AI. What is the correct answer?
Answer: (2)It is the same kind of answer as "do you live in India, or Maharashtra, or Mumbai, or Bandra?" Nobody finds that confusing: all four are true simultaneously, and you would use different ones in different conversations. They are not four places you might live; they are four levels of zoom on one place.
Q2A vendor tells you their company is "doing AI." How much have they told you?
Answer: (2)Bigger ring means older, broader and vaguer; smaller ring means newer, narrower and more specific. "AI" covers an enormous amount of ground and tells you very little about any particular system. The useful question is always which ring.
Q3Look at the dates on the four settled rings: AI (1950s), machine learning (1980s), neural networks (1958), deep learning (2012). What is the pattern?
Answer: (2)The rings get smaller as they get more specific, and each one is a narrower answer to a broader ambition. (Note the neural-network date sits out of sequence — the idea is older than the ring that contains it, which is Q11's story.)
B · Ring 1 — Artificial Intelligence
Q4Why is "how does AI work?" not a well-formed question?
Answer: (2)Transport is not a technology; it is a goal — moving things from A to B. Bullock carts, ships, lifts and aircraft are all transport and share almost no engineering. AI is exactly like that, which is precisely why the outer ring is so wide and so vague, and why the term has meant completely different things across seventy years.
Q5A chess machine beat the world champion in 1997 using hand-written evaluation rules and enormous brute-force search, with nothing improving on its own. Where does it sit?
Answer: (3)Unquestionably AI, and a landmark in the field, containing no machine learning whatsoever. Most people assume AI requires learning. It does not, and the counter-example is famous.
C · Ring 2 — Machine Learning, and the correction that matters
Q6What makes machine learning a subset of AI rather than a synonym for it?
Answer: (2)If AI is "transport," machine learning is "engines" — one particular approach that turned out to work so well it now dominates. Plenty of transport has no engine, and there was a lot of transport before engines existed.
Q7Your company has 60,000 rows of customer data in a database, with columns like tenure, last order value and support tickets. Where should you start?
Answer: (3)On rows and columns, a well-tuned random forest routinely matches or beats a deep neural network — while training in minutes instead of days, on a laptop instead of a GPU cluster, on thousands of rows instead of millions, and with an explanation you can show a regulator. If your problem is a spreadsheet, you probably do not need deep learning and you certainly should not start there.
Q8Techniques like logistic regression, decision trees and random forests were mostly invented between the 1950s and 1990s. What is their status today?
Answer: (2)Credit scoring, churn prediction, demand forecasting, insurance pricing, fraud and medical risk scores are overwhelmingly this ring, not the inner ones. Being in the outer, older ring is not a defeat — a model that ships this quarter and can be explained beats one that lands next year and cannot.
Q9Someone proposes a neural network for a customer-churn model. What is the fair question to ask?
Answer: (2)Politely, and without hostility. Deep learning earns its keep on images, audio, video and language — data with no natural columns. A churn model is a table, and the burden of proof sits with the more expensive, less explainable option.
D · Rings 3 and 4 — neural networks and deep learning
Q10What is the honest picture of a single unit inside a neural network?
Answer: (2)Picture an assembly line of extremely dim workers: no individual worker understands the product, each does one trivial thing and passes it on. Complexity comes entirely from how many of them there are and how they are wired, not from any one being clever. Millions of them, in layers, produce ChatGPT.
Q11The neural-network idea dates from 1958 and then failed, repeatedly, for about fifty years. What changed?
Answer: (2)Through the 1980s and 1990s it was a fringe technique that mostly lost to classical methods, and two "AI winters" happened with neural networks sitting right there, not working. The same idea can be worthless and then transformative without changing at all — which is a genuinely useful lesson for anyone judging a technology today.
Q12What is the single most important thing deep learning does that the outer rings cannot?
Answer: (2)Classical machine learning needs a human to decide what to measure — ear length, snout-to-skull ratio, fur texture — and to write code extracting each one. That step had a name, feature engineering, and for thirty years it was the actual job. Deep learning largely automated it away: it gets the raw pixels and works out that ears matter.
Q13Which pairing correctly maps two familiar teaching examples onto classical machine learning and deep learning?
Answer: (2)A smart human found the feature in one case; nobody specified any features in the other. You already taught deep learning in session one — you just did not name it.
E · Where the diagram breaks
Q14Up to deep learning the words nest cleanly. What happens after that?
Answer: (2)Consider: Sport → Ball sports → Cricket → T20 → IPL → Mumbai Indians → Saturday's match. The first three genuinely nest. Then it changes: T20 is a format, IPL a tournament, Mumbai Indians a team, Saturday's match an event. Drawing those as one more nested circle implies something false — that every T20 involves Mumbai Indians.
Q15In the vehicle analogy used to sort the six confusing words, what is "Generative AI" the equivalent of?
Answer: (3)Transformer = front-wheel drive (an architecture — how it is built). Generative AI = a delivery vehicle (a capability — what it is for). LLM = a lorry (a class of model). A delivery vehicle might be front-wheel drive or not — which is exactly why the two cannot be nested.
Q16A widely circulated diagram places LLM inside GPT. What is wrong with it?
Answer: (2)Claude, Gemini, Llama and Mistral are all LLMs and none of them is a GPT. It is the equivalent of a diagram showing "cars" inside "Toyota." A sharp participant will catch this, and once they do it undermines everything else on the slide.
Q17The same diagram places Generative AI inside Transformers. Why is that wrong in both directions?
Answer: (2)Two things that overlap, where neither contains the other, cannot be drawn as one circle inside another. Those two counter-examples are the pair to use if you only have time for two — together they prove the arrows point in both directions.
Q18Autocomplete on a 2005 mobile phone produced text you did not type, with no deep learning anywhere. Which claim does that break?
Answer: (2)It generates, and there is no neural network in sight. It is the counter-example that stops "GenAI ⊂ deep learning" from being stated as a strict rule — though as a rough simplification for a slide, it remains broadly fine today.
F · ChatGPT is not GPT
Q19What is the difference between a model like GPT-4 and a product like ChatGPT?
Answer: (2)When you use the product you get a great deal of surrounding machinery. When you buy access through the API you get closer to the bare engine and build the rest yourself. The same underlying model can answer differently in the two places, because the product wraps it in instructions you never see.
Q20Why is "ChatGPT versus Claude" a category error, strictly speaking?
Answer: (2)It is a perfectly reasonable thing to do in conversation, but be clear what you are comparing. And note where your own organisation actually operates: what your company builds sits at the model layer — the interface, guardrails and integrations are yours to design, which is where most of the value and most of the risk lives.
G · Which ring is a budget decision
Q21Which cost shape belongs to foundation models and generative AI?
Answer: (2)That is a genuinely different economic shape from anything most organisations have bought before, and budgets built for one-off software projects handle it badly. Classical machine learning is the opposite: a one-off build that is cheap to run.
Q22Why do banking and insurance still do most of their machine learning in ring 2?
Answer: (2)It is not timidity. It is the explainability column of the procurement table arriving as a compliance constraint — the "nobody can read the rule" problem showing up as a legal requirement.
Q23Match each ring to its characteristic failure mode.
Answer: (2)Worth knowing before you choose, because the failure mode determines what governance you need. A quietly mediocre model needs monitoring; a confidently wrong one at scale needs a human in the loop.
H · Placing real systems
Q24A bank deciding whether to approve a home loan. How far into the rings does it go?
Answer: (1)Learned from past loans, so it is genuinely machine learning — but kept in the second ring on purpose. The stopping point is a compliance decision, not a technical limitation.
Q25A system that reads incoming support tickets and files each one as billing, technical or complaint — built on the same underlying design as ChatGPT. Is it generative AI?
Answer: (2)Same machinery, different job. This is one half of the pair that shows why generative AI cannot be drawn as a circle inside anything: an image generator generates without being a language model, and this classifier uses ChatGPT's design and generates nothing.
Q26An assistant that reads your inbox, drafts replies and books meetings. Which ring?
Answer: (3)And that is the point. It is exactly the closing line of the session-one whiteboard: many inferences, integrated, aimed at an objective — typically calling several models from several rings. Agents sit outside the diagram, wrapping parts of it.
Q27A thermostat that switches on the air conditioning whenever the room goes above 24°C. Which ring?
Answer: (3)Someone will put it inside AI, and that moment is useful rather than a mistake to correct quickly. It observes, analyses and infers — but it never predicts anything it later checks, so nothing learns and nothing changes.
Q28Excel filling in the rest of a series when you drag a cell corner. Where does it sit?
Answer: (2)The honest teaching point is that there is no sharp boundary here. Saying so plainly buys more trust than inventing a rule that will not survive the first hard case.
I · Confusions worth heading off
Q29"Is generative AI better than machine learning?"
Answer: (2)Like asking whether a lorry is better than an engine. The real question is which is right for your problem — and for anything involving a table of numbers the answer is usually classical machine learning.
Q30"We should move from ML to AI." What is wrong with this sentence?
Answer: (2)And that is a specific, expensive choice that should be argued on its merits rather than smuggled in as an upgrade. Watch for this phrasing — it usually signals that a decision is being made on vocabulary rather than on fit.
Q31"The innermost circle keeps shrinking — are we close to artificial general intelligence?"
Answer: (3)New inner circles have appeared every few years and will keep appearing. Any slide claiming the diagram predicts the future is selling something.
Q32Which four statements close this topic correctly?
Answer: (2)Like India, Maharashtra, Mumbai, Bandra. AI does not require learning and machine learning does not require deep learning. Most working machine learning in business sits in the second ring, and that is a feature rather than a shortfall. And every confusing diagram you will ever see is caused by drawing blueprints, jobs, brands and products as circles.