A6, Lesson 1: How AI actually learns
Tags: Sourced stories / Our view / Illustrative examples. Self-paced reading and practice. No math or code. Every case or news claim carries its source and the date we read it; invented figures are labeled invented.
1. The story
On September 30, 2012, a neural network built by three researchers at the University of Toronto won the ImageNet contest, the world's best-known competition in computer vision. Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton trained their network, AlexNet, on 1.2 million labeled photographs, using two gaming graphics cards in a bedroom. On the contest's separate test images, its top-5 error rate - the share of photos where the right answer was not among its five best guesses - was 15.3%, more than ten percentage points clear of the runner-up. Earlier winners had relied on visual features hand-designed by researchers; AlexNet learned its features from the data itself, and the margin was so wide that the whole field changed direction.
2. The one idea
Modern AI learns from examples, not from hand-written rules. Nobody described a cat to AlexNet. It saw an enormous number of labeled photos and adjusted itself until its guesses improved. What an AI learns depends heavily on the examples it is shown - the quantity, the quality, and how well they match the real situations it will face.
3. The kitchen-table version
Teaching a child to spot a ripe tomato. You can hand over a rulebook - "deep red, slight give, sweet smell" - or you can spend a summer at a market stall handling ten thousand tomatoes. The stall teaches more, including things nobody thought to write down. But if half the stall's "ripe" crate was actually overripe, the child learns that too. The examples are the teacher, for better and for worse.
4. Worked example
Wick and Ember is an invented two-person candle shop we will follow through this course. Its examples live in four places. Three years of order rows: what sold, when, at what price. A photo library: every product shot ever taken. The support inbox: every customer question and complaint. And the supplier ledger: what was bought, when it arrived, what it cost. A tool that learns from the order rows can find the shop's real rhythm - which scents peak in which weeks. A tool that learns from the support inbox can draft replies in the shop's voice. But each tool is only reading its own pile, and each pile contains the shop's mistakes along with its wisdom.
WORKED EXAMPLE, invented and illustrative, arithmetic script-checked: Wick and Ember's best candle, Hearth, sold these November unit counts: 2023: 410, 2024: 455, 2025: 505. Year over year: 455/410 is about +11%, and 505/455 is about +11% again. A simple learner with a simple rule - next year grows by this year's growth rate - reads those rows and projects November 2026 at roughly 505 x 1.11 = about 560 units. Now the catch: the owner remembers that the 2024 row includes a 120-unit one-off wedding order that was never labeled as special. Strip it out and the same spreadsheet reads 410, 335, 505 - down 18%, then up 51% - and the same rule now projects 505 x 1.51 = about 761 units. Same spreadsheet, same rule, one unlabeled row: the forecast swings by about 200 units, because removing one exceptional order exposed that the year-to-year pattern was never stable to begin with. That does not mean no trustworthy forecast can exist; it means these three points and this simple rule are not a sound basis for a purchasing decision. Nothing miscomputed. The input was mixed.
Instructor: "So the fix is more data?" Learner: "More data can help through better coverage, but adding mislabeled examples can also carry old mistakes forward. AlexNet learned from about 1.2 million labeled training images. Our shop example is a simple illustration of input-quality problems, not a reconstruction of neural-network training. Training can pick up patterns in errors as well as useful patterns - and ten years of mislabeled orders teach mess with confidence."
Excluded costs, stated: this is not a purchasing plan. Supplier minimums, storage, cash tied up in stock, and the asymmetric cost of ordering too many versus too few are all real and all excluded. And this simple extrapolation illustrates the teaching role of examples - it is not a model of how neural-network training itself works. The example teaches one thing only: the rows are the teacher.
5. What it costs and what it breaks
Keeping records clean takes weekly discipline, and labeling exceptions takes judgment nobody can fully automate. Tools that learn from your data cost more than free generic ones, in money or in setup time. And examples are not everything - a tool can learn your past perfectly and still miss a changing market, because examples describe yesterday. The 2020 spring taught every forecaster that lesson at once.
6. Monday checklist
Pick the one record set a tool would most likely learn from (usually orders). Sample twenty rows and check whether what they say matches what actually happened. Find your shop's version of the mislabeled wedding order - bulk, wholesale, staff purchases - and mark them. Note where your photos, support emails and supplier records actually live and who keeps them current. Write down which single record set you would clean first.
7. Try it yourself (five minutes)
Take twenty orders from last year and write down the pattern you see. Then paste the same twenty into any AI chatbot and ask for the pattern. Compare the answers and mark what neither of you can infer from twenty rows. This is an exercise in inspecting data and questioning a chatbot, not training a model. Use invented or de-identified rows; do not upload private customer details.
8. Self-check
- What does "top-5 error rate of 15.3%" mean, in plain words? Answer: On the contest's separate test photos, the right answer was missing from the model's five best guesses about one time in seven.
- Were the 1.2 million images the training set or the test set? Answer: The training set; the 15.3% was measured on different, held-out images.
- In the worked example, why did the forecast swing from about 560 to about 761? Answer: One row mixed a 120-unit one-off order with regular demand; the learner had no way to know.
- If an AI tool gives odd answers about your products, what is the first question to ask? Answer: What examples can it draw on - and are they clean, current, and about your shop?
- Which of your own records would be worth cleaning first, and why that one?
Sources
- Krizhevsky, Sutskever and Hinton, "ImageNet Classification with Deep Convolutional Neural Networks" (2012), and the ImageNet ILSVRC-2012 results: trained on 1.2 million images, 15.3% top-5 error on the test set, more than 10.8 points ahead of the runner-up, two GTX 580 GPUs, bedroom.
- Read 3 Nov 2026 via the paper and Wikipedia; contest videos not watched.
- Wick and Ember, its order rows and all figures are invented, illustrative and script-checked, not a reported story.
- Review flags: "depends heavily on examples" is a simplification; the changing-market caution is our framing, not a sourced finding.
- Go deeper, free: Harvard's session 2, Deep Neural Networks, builds this idea from n-grams on Shakespeare to a tiny two-by-two-pixel network in its slide deck. Watch it here.
Review flags
- "depends heavily on examples" is a simplification; the changing-market caution is our framing, not a sourced finding.
Claim ledger
| Claim or example | Basis | Checked |
|---|---|---|
| 1.2 million training images; 15.3% top-5 test error; two GPUs. | AlexNet paper | 3 Nov 2026 |
| Invented order-series example: about 560 vs 761 units under the same rule. | Invented, illustrative; arithmetic script-checked | 3 Nov 2026 |
Independence
Independently developed by Commerce Stories, not affiliated with Harvard. We link to the archived 2024 course; we do not reproduce its videos, slides or exercises. The linked videos were not watched for this written-material review.