Exam practice

• Artificial intelligence (AI): a computer doing a job that needs a mind when a person does it, such as recognizing pictures, understanding speech or translating.
• Machine learning: a way of making a program in which the computer learns from examples with the right answers instead of following rules that a person wrote.
• Training: the computer finding the pattern in the examples.
• Model: what the computer ends up with after training, which it uses to answer for new things.

An ordinary program follows rules that a person wrote, step by step, and does the same thing every time; it learned nothing. Example: the calculator, the alarm clock.
A program that learns was given many examples with the right answers and found the pattern in them by itself; its answers depend on its examples. Example: the photo gallery finding photos of dogs, voice typing, a translation app.

Any four, such as:

• Not alive or feeling: it has no feelings or wishes, even when it writes “I’m happy to help”.
• Not understanding as we do: it knows only the numbers it measures, not what a thing is.
• Not always right: it learned from examples, and examples never cover everything, so it makes mistakes.
• Not aware of what it doesn’t know: it gives one of its answers even for something it was never taught.
• Not every clever machine: a rice cooker or a factory robot following rules that people wrote is not AI.
• Not magic: it is a computer doing arithmetic on numbers, very fast.

1. It measures the new fruit: its size and its color.
2. It puts it on the map of its examples, at the place those two measurements give.
3. It finds the example closest to the new fruit on the map.
4. It gives the same answer as that example: kagati if the closest example is a kagati, suntala if it is a suntala.

No rules were written into it; its answers come only from its examples.

Reason: none of its examples was a green suntala. On its map, green is much closer to yellow than to orange, so the green fruit is closer to the yellow kagati examples than to the orange suntala, and it copies a kagati. A model only knows what its examples showed it.

Fix: add examples of unripe, green suntala (and green kagati) with their right answers, so the examples cover all the fruit it will meet. Then test it again on many fruit, not just this one, because adding examples changes other answers too.

1. Too few examples.
2. Examples that don’t look like the new thing (they didn’t cover it).
3. It measures only some things about each example, and can’t see the rest.
4. It can only give the answers it was taught, so it answers even for something it never learned.
5. Some examples had the wrong answer, and it learned the mistake.

Any four, such as:

• Photo search on a phone: from millions of photos, each named with what is in it.
• Voice typing in Nepali: from many hours of people speaking, with what they said written out.
• Translation apps: from a great many sentences already translated by people, in both languages.
• Video and shopping recommendations: from what millions of people watched, clicked and bought.
• A bank’s fraud check: from past payments, honest ones and frauds.
• Crop disease apps: from photos of sick leaves whose disease an expert had named.
• Face unlock, spam filters, traffic on maps.

The three questions: What happens if it’s wrong? Has it seen things like this? Can I check it?

• i.A wrong label does no harm and you can see it is wrong: fine to use, or just correct the label.
• ii.A wrong spray wastes money and harms the soil and the water, and the app may not know diseases common in Nepal; you can check: use it, then check with an agriculture technician before spraying.
• iii.A wrong dose could seriously harm him, and you can’t check it yourself: don’t rely on it; ask a doctor, a health worker or a pharmacist.

1. A calculator, or your own working, does it exactly right every time; an AI’s answer is a guess from examples and can be wrong.
2. AI takes much more computing, and so much more electricity, than a calculator for the same answer: data centers use a lot of electricity and water.

(Also: it’s homework, so working it out yourself is how you learn it, and handing in an answer you didn’t work out isn’t your own work.)

Many AI apps send what you type, say or show them over the internet to the company’s computers, where it may be stored, read by people who check the AI, or used to train the next model. Once it is sent, you can’t take it back. Private information could then be seen by strangers and used to get into your accounts or find you.