What is AI?

LaylaOne message reaches your inbox. Another lands in spam. How does the computer decide?

LUNAEasy. Someone told it which words to look for.

LaylaSometimes. Rules can help. But a system can also learn patterns from examples, and still make mistakes.

LaylaI'm Layla. Today we'll make sense of AI: what it does, how we got here, and what to check.

LaylaArtificial intelligence is a broad field. Some systems use rules and knowledge written by people. Others learn patterns from data. Some combine both.

LaylaIn this simplified example, a filter learns from messages labeled spam or not spam. Later, it uses those learned patterns to sort a new message.

LUNASo it can put an invitation in the wrong place, even after all that learning?

LaylaExactly. We need to know the task, the evidence, and what happens when it gets something wrong.

LaylaKeep those three questions close. They will help us look past the impressive surface and see what a system can do. First, let's open the filter and work through one decision together.

LaylaLet's slow down that first decision. Imagine we write one rule: if the subject contains the word free, send the message to spam.

LaylaIt catches a suspicious free prize. It also catches a perfectly good invitation to a free library event. Our rule is doing exactly what we told it to do.

LUNAThe library gets punished for being generous. I object.

LaylaObjection accepted. We could add exceptions. But every exception is another choice someone has to maintain. For this example, let's try learning from labeled messages instead.

LaylaThis approach is called machine learning: learning patterns from data rather than writing every decision as a rule. Here, learning means adjusting a model using examples. Our labels show which messages we wanted in each tray.

LaylaNow we hold back a message the model hasn't used for learning, and try it. Getting old examples right is useful. Handling a new one is the test we care about.

ZEDSo the lesson is bigger than memorizing the answer sheet.

LaylaYes. And one good result isn't enough. We'd try many different messages and inspect the mistakes. This tiny drawing explains the idea; it isn't a measurement of a commercial filter.

La BailarinaWho decides which messages count as spam in those examples?

LaylaIn our demonstration, we do. And we need to be clear about what we mean. Consider the same newsletter reaching two people. One signed up for it. The other didn't. The words alone won't explain that difference.

LaylaBefore judging our filter, we define the job: which messages should this particular person receive? Then we choose examples and labels that fit that job. A pile of examples doesn't choose our purpose for us.

La BailarinaSo the human choices are part of the story, even when the prediction is automatic.

LaylaYour turn. A subject says, Free ticket enclosed. Under our original rule, which tray gets it? Take a moment before I answer.

LaylaSpam, because the rule only checks that word. Now a harder question: is the message unwanted? We can't settle that from the subject alone.

NULLW3AV3I'd check who sent it, and whether I expected a ticket.

LaylaGood. You separated the system's output from the truth of the situation. That's a small move with a long reach. We will use it again when an answer arrives as a beautiful paragraph.

LaylaYou will hear the phrase narrow AI for systems focused on one task or a small group of tasks. Other tools work across many kinds of tasks.

LaylaResearchers still debate what should count as artificial general intelligence, or AGI. You don't need to settle that debate to ask useful questions about a tool.

LaylaCan it do the job you need? Under what conditions? How often does it fail in examples like yours? A spectacular demonstration answers fewer questions than it first appears to.

LUNAIf it says it loves my hat, should I take the compliment?

LaylaYou may enjoy the compliment. But a sentence about a feeling isn't evidence that the system experiences that feeling. What an AI can do is different from whether it has feelings or awareness.

LaylaIn our classroom, characters have stories and personalities. In the lesson, we'll focus on what we can test and observe. Your hat remains beyond the scope of this investigation.

LaylaThese ideas did not begin with a chat window. In nineteen fifty-five, a research proposal used the name artificial intelligence and proposed a summer meeting at Dartmouth in nineteen fifty-six.

LaylaNotice the two dates: the proposal, then the planned project. A timeline should help us understand what happened, not turn different events into one tidy birthday.

ZEDI like seeing the work before the famous moment.

LaylaMe too. A proposal is a beginning, not a finished machine. Let’s follow a few concrete milestones, and ask what each one demonstrated.

LaylaIn nineteen ninety-seven, IBM's Deep Blue defeated reigning world chess champion Garry Kasparov in a six-game match. It evaluated possible chess positions to choose moves.

LaylaThat was an impressive achievement. It was also a specific test, with a board, legal moves, and a clear result. Winning it was evidence about chess performance.

NULLW3AV3So the headline doesn't tell me whether it can organize my desk.

LaylaExactly. Whenever a system succeeds, ask which task the evidence covers. Keep the achievement. Keep its boundaries too.

LaylaIn twenty seventeen, the paper Attention Is All You Need introduced the Transformer architecture. Attention helps connect relevant parts of an input.

LaylaLook at this little example: The ticket is blue. It is on the desk. Understanding what it refers to involves a relationship across the words. Our picture illustrates that idea; it isn't a recording of a model's thoughts.

LaylaThe Transformer became an important design for language models. AI can also use other designs. One diagram cannot explain every system with AI on its label.

LaylaChatGPT launched in November twenty twenty-two. A conversational interface gave people a way to explore language-model capabilities by asking questions.

LaylaOur timeline is a selection, not the whole history. The useful pattern is that different designs demonstrated different capabilities. So when a new tool appears, we look at what changed, and how anyone checked it.

LaylaNow put a confident paragraph beside our misrouted invitation. Generative AI can produce false or unsupported information that sounds plausible. This is often called hallucination.

NULLW3AV3Then how am I supposed to build anything on top of it?

LaylaBy choosing the job and the checks together. Imagine asking for three cheerful invitation drafts. You can judge the tone yourself. Asking for the library's actual opening time creates a different obligation: check the time.

LaylaThe same fluent voice can appear in both answers. Style doesn't tell you which facts were verified.

LaylaLet's do the check. Our invented draft says the event starts at one. The matching entry in our invented library schedule says two. We check the event name and date as well, so we don't compare two different events.

La BailarinaThen I correct the invitation to two, and keep the schedule with it.

LaylaYes. We have evidence for the correction. Asking the same chatbot, Are you sure, would give us another response. It wouldn't replace opening the source.

LaylaFor something consequential, choose checks strong enough for the consequence. A source must exist, be relevant, and support the particular claim. A link that merely looks official is only a starting point.

ZEDAnd if I tell it the correct time, will it remember tomorrow?

LaylaThat depends on the app and its settings. A new chat may not include your earlier conversation. Some apps can save details or use past chats.

LaylaSupplying saved context is different from teaching the underlying model a new fact through training. Check what the app says it saves, what you can review, and what you can remove.

ZEDSo I should bring the schedule again if the next task needs it.

LaylaA good habit. Give the task the evidence it needs, instead of assuming yesterday's correction followed you into the room.

LaylaIt also helps to distinguish a model from the app around it. Think of our classroom project. We could use a model to draft wording, while the app supplies the schedule and keeps the draft in a document.

LaylaIf that schedule is missing, the task has changed. If the app adds a search tool, the available actions have changed. Comparing two answers without noticing those differences can be misleading.

ZEDSo when I ask what an AI can do, I should say which setup I'm using.

LaylaYes. The model, the information it receives, and the tools around it all belong in the description. Now we can make sense of the word agent.

LaylaThere's another distinction. An answer can describe what to do. An agentic system can use tools, inspect results, and decide what step to take next while working toward a goal.

LaylaIn our example, the goal is to prepare an invitation. A permitted tool reads the schedule. The system drafts the message, compares the details, and stops for your review. Sending isn't part of this task.

NULLW3AV3I want that boundary written down before it touches my inbox.

LaylaYes. The tools, permissions, stopping point, and checks belong in the design. More steps can create more opportunities to help, and more places to make a mistake.

LaylaLet's try one decision. The draft is ready, but its time disagrees with the schedule. What should the system do next: send it, inspect the schedule, or make the wording more impressive?

La BailarinaInspect the schedule. Beautiful words won't repair the wrong time.

LaylaRight. The next useful action is the one that resolves the uncertainty. After the repair, we still review the message and its recipient before sending. Finishing a draft and approving an action are different checkpoints.

La BailarinaWhat makes you excited about where this could go?

LaylaThe possibility of more people making the things they imagine. Someone getting help with a first draft, testing an idea, or finding another way into a difficult lesson. That's the future I'm rooting for.

LaylaIt isn't a guarantee. What people build, who can use it, and how carefully it is checked all matter. You can be curious about the possibilities and demanding about the evidence at the same time.

LaylaBefore we leave, close the notes for three questions. First: does every AI system learn from data?

LaylaNo. Some use rules and knowledge written by people, some learn patterns, and some combine approaches.

LaylaSecond: does a confident answer prove it is correct?

LaylaNo. Check the claim against relevant evidence. Our invitation sounded fine while its time was wrong.

LaylaThird: what extra question matters when an AI can use tools to act?

LaylaAsk what it is allowed to do, where it must stop, and how you'll check the result. If you remembered part of that, you have something solid to build on.

LaylaNext time, we'll look more closely at how computers learn: examples, adjustments, and the test of something new. For now, choose one AI tool and name its task, its evidence, and one mistake you'd check.

LaylaRain doesn't fall the same way twice. Bring your questions. I'll see you next time.