40 standalone questions on the six-step board — what learning actually is, and where ChatGPT sits on it.
Audience: absolute beginners. No maths, no code, no prior exposure assumed.
Source: the six-step whiteboard — Observe, Analyze, Infer / Predict, Recalibrate, Feedback loop — and the closing line about integrating inferences toward an objective.
How to use this: every question stands alone. Pick an option, then read the answer. The sections follow the board from the top line to the closing sentence.
A · The line across the middle
Q1A system takes something in, works something out, and produces an answer. It does this the same way on its first day and on its ten-thousandth. What is it?
Answer: (2)Taking something in, processing it and producing an answer is one complete pass — and one pass is all automation ever does. What makes something machine learning is the second half: making a prediction, finding out whether it was right, and changing itself so the next one is closer. A calculator observes, processes and outputs. So does a doorbell. Doing is not learning.
Q2Here is a test you can apply to any system for the rest of your career: if you unplugged it for a year and plugged it back in, would it behave any differently? What does "no" tell you?
Answer: (2)If nothing about the system changed while it was running, nothing was being learned. The difference between automation and machine learning is not intelligence, sophistication or how modern it looks — it is whether the system has a feedback loop. No loop, no learning, no matter how impressive the interface.
Q3Which of these best describes the relationship between the three words?
Answer: (2)The top half of the board is automation. Add the bottom half — predict, reconcile, recalibrate — and you have machine learning. Take many such learned inferences, integrate them and point them at an objective, and you have artificial intelligence. They are three different amounts of the same board, not three competing technologies.
Q4A traffic light changes its timing based on live sensor data, following a rule an engineer wrote. Is it machine learning?
Answer: (3)This is the item that catches almost everyone. It genuinely does all three top-half steps, which feels like intelligence. But it never predicts anything it later checks, so it never discovers it was wrong, so nothing ever changes. Reacting to the world is not the same as learning from it. A system that adjusted its timings over months to reduce measured average waiting time would be machine learning.
B · The six steps
Q5A fraud system sees the amount, the merchant category, the city, the time and whether the card was present. The customer's actual situation — that they are on holiday — is nowhere in that list. What does the board say about this?
Answer: (2)Observation is always narrower than reality, and the system can only ever reason about what it was given. Roughly half of all AI failures in business trace back to this one line: the thing that actually mattered was never in the input. No amount of modelling recovers information that was never collected.
Q6The board deliberately uses the word infer rather than know, look up or calculate. What does inferring mean?
Answer: (3)A doctor seeing a fever infers an infection; she does not know it. Everything a model produces is of this kind — a best available conclusion, not a retrieved fact. Which explains something people find baffling: being confidently wrong is not a bug in inference, it is the nature of inference. Humans do it constantly too.
Q7Why does the system have to commit to a prediction before reality reveals the answer?
Answer: (2)The commitment is what makes learning possible. If you never state what you expect, you can never discover you were wrong — and this is as true of organisations as of algorithms. Companies that forecast constantly and reconcile almost never are running the top half of the board and skipping the bottom half, which is why they do not get smarter year over year.
Q8What is recalibration, precisely?
Answer: (2)Not a rewrite — a nudge. Thousands of small nudges is what training actually is. Every technique you will ever read about — gradient descent, loss functions, epochs — is machinery for doing exactly this, faster and across more variables than a person can hold in their head.
C · What the loop actually buys you
Q9A tea-stall owner predicts he will sell 120 cups; about 145 people want one. Next time he predicts 145 and sells 138. His error went from 25 cups to 7. What changed?
Answer: (3)Same stall, same tea, same skill. The only thing that changed was the internal rule, and it changed because a prediction was checked against an outcome. That shrinking error number is the only thing a machine learning system cares about. Training is running this loop millions of times so that number gets smaller.
Q10The same tea-stall owner never counts how many customers he turns away. What follows?
Answer: (2)Without a measured gap between prediction and outcome there is nothing to learn from. No feedback, no learning — no matter how much data you have. A very large number of failed corporate AI projects fail here and nowhere else.
Q11A sales rep marks 30 leads as hot and expects 15 to close. Six close. Digging into the data, she finds enthusiasm on the call predicted nothing, but 5 of the 6 that closed had booked a demo within three days. Next quarter her error falls from 9 to 1. What did she just do?
Answer: (2)Observe, infer, predict, reconcile against reality, find the feature that actually carries signal, recalibrate. A machine learning system does the identical thing across 200 variables instead of 4, and finds combinations no human would think to check. It is not doing something mysterious. It is doing this, at a scale you cannot do by hand. Note too that the good rule was discovered from outcomes, not designed in a meeting.
Q12Google Maps tells you 34 minutes and you arrive in 41. Which part of that interaction is most valuable to Google?
Answer: (2)The ETA is not the product. The gap is the product. You get a slightly useful number; Google gets a perfectly labelled training example, captured automatically, with nobody filling in a form. Multiply by a billion journeys a day and every trip makes the system better for everyone.
Q13A 1998 spam filter blocks any message containing "lottery". Within a week spammers write "l0ttery". What general lesson does this teach?
Answer: (2)Rules are written once and then decay; a learned system adapts because its loop never stops. When enough people click "mark as spam" on the new spelling, the filter picks it up within hours — along with a hundred signals no human would have thought to write down. Most real business problems are of this kind: pricing, fraud, demand, churn, competition.
D · What is actually inside a trained model
Q14A five-year-old can identify dogs they have never seen, from behind, in bad light, half hidden by a car — and cannot tell you the rule. What does this demonstrate?
Answer: (2)Ask an adult to define "dog" precisely enough for a machine to follow and they will fail too. That is exactly what a trained machine learning model is — not a rulebook, but a capability learned from corrected examples that nobody can read back as a sentence.
Q15A master tea taster grades a batch in four seconds and is right. Asked to write the rule down for a new hire, he says things like "it's a bit thin on the finish." What is the executive consequence of this?
Answer: (2)The knowledge is real and reliable, and it does not exist in a form that can be written as instructions. The same is true inside a trained model — the people who built it cannot read it either. Auditing means testing, not reading, and "why did it reject my loan?" therefore has no easy answer.
Q16A hiring tool was found to downgrade CVs containing the word "women's" and to penalise graduates of two all-women's colleges. What had gone wrong?
Answer: (2)The system was working perfectly. The feedback loop has no opinion about what it learns — it learns whatever reality it is shown, including reality you are not proud of. And because the learned rule ends up as an unreadable pile of numbers, nobody notices until someone deliberately goes looking.
Q17In that same hiring example, fewer women get hired, so next year's training data contains even fewer, so the pattern strengthens. What is the general principle?
Answer: (2)The loop runs on the world, not just on the data. This is why the closing line of the board says objective, and why choosing the objective is the most consequential decision in any AI project — a decision that belongs to people, not to the system.
E · A trade-off that is not technical
Q18A fraud model is retuned. Flagged transactions fall from 1,000 a day to 300; real fraud caught falls from 60 to 55; false alarms fall from 940 to 245; fraud missed rises from 0 to 5. Which model should be deployed?
Answer: (3)The second model is better on customer experience and worse on catching fraud. Which one you ship is not a technical question; it belongs to the room, not to the data-science team. Anyone who says "we'll just make the model more accurate" has not understood that these two numbers trade against each other.
Q19Your card is declined, you reply "YES, IT WAS ME", and your evening is ruined. What has the bank just received?
Answer: (2)Your reply reconciles the prediction against reality — the bottom half of the board, running in real time. The irritation is genuine and so is the value: the single most useful thing that happened in that interaction was the correction you supplied.
F · Where ChatGPT sits on the board
Q20While you are chatting with a large language model, which parts of the board are running?
Answer: (2)The predicting, reconciling and recalibrating ran once, across an enormous quantity of text, took months, cost an extraordinary amount, and then stopped. That frozen result is the model. While you chat, no prediction is being scored and nothing is being recalibrated. This single fact explains almost every beginner misconception about these tools.
Q21You correct a chatbot, it apologises — and then makes the identical mistake in a new conversation. Why?
Answer: (2)Nothing was learned. The bottom half of the board is switched off while you chat, so a correction is an observation for the rest of that conversation and nothing more. It never reaches the frozen weights.
Q22The model seems to remember what you said ten messages ago. What is actually happening?
Answer: (2)Each turn is a fresh pass through the top half of the board, with the whole conversation supplied again as observation. Nothing is retained inside the model between turns.
Q23Picture someone who studied intensively for two years and is now sitting a closed-book exam. Which statement follows from that analogy?
Answer: (2)That is a large language model answering your prompt — which is also why it states false things with total confidence. Writing the most plausible available answer is exactly what inference means, and exactly what it was built to do.
Q24Two ways of working around frozen weights: sliding the candidate a reference sheet during the exam, or sending them back for a short refresher course. Which is which?
Answer: (2)RAG changes what the model observes and leaves the model untouched. Fine-tuning re-runs a small slice of the bottom half and does change the model. Getting these the wrong way round leads to expensive mistakes — most notably trying to teach a model facts by fine-tuning when the facts keep changing.
Q25If the model is not learning from you while you chat, what is the only thing you actually control?
Answer: (3)Which is the entire reason prompting is a skill worth teaching. Everything you can influence about the output happens on the observation line of the board.
G · Many inferences, one objective
Q26A self-driving car contains a traffic-sign reader that is genuinely superb — better than you, at night, in rain, at 100 km/h. Put it in a car on its own. What happens?
Answer: (2)Do the same with any of the others and you get the same result. The object-recogniser identifies everything and does nothing about it; the route planner knows the perfect way and cannot see a child at the kerb. Every one of the specialists is individually excellent and individually useless. Not one of them is driving.
Q27Five specialists report at a green light: signs says green, limit 60; route says straight on, you're behind schedule; environment says a distracted child is one step from the road; vehicle says 34 metres to stop on this wet surface. Every report is correct and they point in different directions. What resolves it?
Answer: (3)Nothing in the list resolves this. What resolves it is the goal the whole system is pointed at. That arbitration is the intelligence — not the sensing, not the recognising, not the braking calculation. The integration, under an objective.
Q28If each of five chained components is 99% reliable, roughly how reliable is the combination?
Answer: (2)0.99⁵ ≈ 95% — one failure in twenty. Reliability multiplies along a chain, it does not average. In a car at 60 km/h, one in twenty is a catastrophe, and this is a large part of why self-driving took a decade longer than predicted: the individual pieces got good quickly and integrating them to the standard the objective required did not.
Q29What is the cleanest definition of an AI agent that this board gives you?
Answer: (2)That is precisely the closing line of the board. Swap the five driving capabilities for classify the complaint + retrieve the history + judge the escalation risk + draft a reply + decide whether a human must approve and you have a business agent. Same shape, same arbitration problem, same truth that only the objective settles a disagreement between components.
Q30Of the components that make up such a system, which part is least likely to be something you can buy from a vendor?
Answer: (3)The components are increasingly commodities. The objective is the part that is irreducibly yours — and it is the part that decides what the system does when its own components disagree. Notice also that the objective never sets itself.
H · Confusions worth heading off
Q31"Is AI just a lot of IF statements?"
Answer: (2)The difference is that nobody wrote them, they run to millions in number, and they involve combinations no human would think to write down. A goat has four legs and fur too — the toddler's final rule is not something you could express as an IF.
Q32"We have twenty years of data, so we can do AI." What is missing from that sentence?
Answer: (2)Data is not the requirement. Labelled outcomes are. Twenty years of invoices with no record of which customers turned out to be a problem gives you nothing to reconcile against. Most companies with "lots of data" have observations without outcomes — which is the top half of the board without the bottom half.
Q33"Will AI replace my job?" What does the board suggest?
Answer: (2)Point at the board. It replaces inferences, which is a much narrower claim than "jobs." Deciding what to observe, setting the objective and judging the output are the parts sitting outside the six steps — and the objective never sets itself.
Q34A chess computer from the 1990s beat the world champion using hand-written rules and enormous search, with no learning of any kind. What is it?
Answer: (3)Genuinely AI, and a landmark. All machine learning is AI, but not all AI is machine learning. ML happens to be the approach that currently works best, which is why the press uses the words interchangeably. They are not the same word, and this is the counter-example that proves it.
Q35"So it's basically statistics?" What is the honest answer?
Answer: (2)Say this rather than pretending otherwise. Being straight about it buys credibility for the parts that genuinely are new — and defensiveness on this question tends to cost you the room.
I · Sorting real systems
Q36Spotify's Discover Weekly playlist — machine learning or not?
Answer: (2)Skipping a track is a labelled example you supplied without noticing. Prediction, reconciliation and recalibration are all present, so the full loop is running.
Q37A payroll system calculating tax — machine learning or not?
Answer: (2)The rule is written in law. A system that started improvising on tax would be a catastrophe, not an improvement — which is a useful reminder that learning is not automatically desirable.
Q38Face unlock that keeps working as you grow a beard — machine learning or not?
Answer: (2)The system predicts "this is the owner", reality confirms it when the passcode is not needed, and the representation shifts a little. The full loop, running quietly, on your face.
Q39Excel autocomplete extending a series when you drag a cell down — machine learning or not?
Answer: (2)It does make a prediction, which is why people hesitate here. But nothing ever tells it the answer, so nothing is reconciled and nothing recalibrates. Predicting is not enough; the prediction has to be scored.
Q40Three sentences to carry away. Which one is not on the board?
Answer: (4)Data volume is neither necessary nor sufficient. What matters is whether outcomes are recorded, whether predictions are scored against them, and whether something integrates the resulting inferences toward a goal. The objective never sets itself — that part is still yours.