WEBVTT 00:00:00.200 --> 00:00:02.485 One message reaches your inbox. 00:00:02.485 --> 00:00:04.390 Another lands in spam. 00:00:04.390 --> 00:00:06.240 How does the computer decide? 00:00:06.625 --> 00:00:10.305 Easy. Someone told it which words to look for. 00:00:10.708 --> 00:00:12.668 Sometimes. Rules can help. 00:00:12.668 --> 00:00:17.428 But a system can also learn patterns from examples, and still make mistakes. 00:00:17.905 --> 00:00:22.973 I'm Layla. Today we'll make sense of AI: what it does, how we got here, 00:00:22.973 --> 00:00:23.833 and what to check. 00:00:24.666 --> 00:00:26.656 Artificial intelligence is a broad field. 00:00:26.656 --> 00:00:30.082 Some systems use rules and knowledge written by people. 00:00:30.082 --> 00:00:32.608 Others learn patterns from data. 00:00:32.608 --> 00:00:33.888 Some combine both. 00:00:34.361 --> 00:00:39.012 In this simplified example, a filter learns from messages labeled spam or not 00:00:39.012 --> 00:00:43.892 spam. Later, it uses those learned patterns to sort a new message. 00:00:44.292 --> 00:00:48.772 So it can put an invitation in the wrong place, even after all that learning? 00:00:49.747 --> 00:00:53.439 Exactly. We need to know the task, the evidence, and what happens when it 00:00:53.439 --> 00:00:54.287 gets something wrong. 00:00:54.801 --> 00:00:56.667 Keep those three questions close. 00:00:56.667 --> 00:01:01.874 They will help us look past the impressive surface and see what a system can do. 00:01:01.874 --> 00:01:06.027 First, let's open the filter and work through one decision together. 00:01:06.553 --> 00:01:09.272 Let's slow down that first decision. 00:01:09.272 --> 00:01:14.467 Imagine we write one rule: if the subject contains the word free, send the 00:01:14.467 --> 00:01:15.697 message to spam. 00:01:16.083 --> 00:01:18.803 It catches a suspicious free prize. 00:01:18.883 --> 00:01:23.683 It also catches a perfectly good invitation to a free library event. 00:01:23.783 --> 00:01:27.283 Our rule is doing exactly what we told it to do. 00:01:27.667 --> 00:01:30.005 The library gets punished for being 00:01:30.005 --> 00:01:32.707 generous. I object. 00:01:33.250 --> 00:01:36.210 Objection accepted. We could add exceptions. 00:01:36.210 --> 00:01:39.650 But every exception is another choice someone has to maintain. 00:01:39.690 --> 00:01:43.810 For this example, let's try learning from labeled messages instead. 00:01:44.416 --> 00:01:48.658 This approach is called machine learning: learning patterns from data rather than 00:01:48.658 --> 00:01:51.104 writing every decision as a rule. 00:01:51.104 --> 00:01:54.988 Here, learning means adjusting a model using examples. 00:01:54.988 --> 00:01:58.448 Our labels show which messages we wanted in each tray. 00:01:58.833 --> 00:02:02.673 Now we hold back a message the model hasn't used for learning, and try it. 00:02:02.703 --> 00:02:06.104 Getting old examples right is useful. 00:02:06.104 --> 00:02:08.833 Handling a new one is the test we care about. 00:02:09.208 --> 00:02:13.528 So the lesson is bigger than memorizing the answer sheet. 00:02:13.917 --> 00:02:16.610 Yes. And one good result isn't enough. 00:02:16.610 --> 00:02:20.077 We'd try many different messages and inspect the mistakes. 00:02:20.173 --> 00:02:24.857 This tiny drawing explains the idea; it isn't a measurement of 00:02:24.857 --> 00:02:25.997 a commercial filter. 00:02:26.375 --> 00:02:30.455 Who decides which messages count as spam in those examples? 00:02:30.972 --> 00:02:32.893 In our demonstration, we do. 00:02:32.893 --> 00:02:35.784 And we need to be clear about what we mean. 00:02:35.784 --> 00:02:38.993 Consider the same newsletter reaching two people. 00:02:38.993 --> 00:02:40.353 One signed up for it. 00:02:40.353 --> 00:02:44.593 The other didn't. The words alone won't explain that difference. 00:02:45.000 --> 00:02:49.777 Before judging our filter, we define the job: which messages should this particular 00:02:49.777 --> 00:02:55.080 person receive? Then we choose examples and labels that fit that job. 00:02:55.360 --> 00:02:58.600 A pile of examples doesn't choose our purpose for us. 00:02:59.000 --> 00:03:02.164 So the human choices are part of the story, even when the 00:03:02.164 --> 00:03:03.560 prediction is automatic. 00:03:04.092 --> 00:03:08.345 Your turn. A subject says, Free ticket enclosed. 00:03:08.345 --> 00:03:11.670 Under our original rule, which tray gets it? 00:03:11.670 --> 00:03:13.958 Take a moment before I answer. 00:03:16.958 --> 00:03:19.918 Spam, because the rule only checks that word. 00:03:19.978 --> 00:03:23.625 Now a harder question: is the message unwanted? 00:03:23.625 --> 00:03:26.318 We can't settle that from the subject alone. 00:03:26.708 --> 00:03:29.828 I'd check who sent it, and whether I expected a ticket. 00:03:30.208 --> 00:03:34.704 Good. You separated the system's output from the truth of the situation. 00:03:34.704 --> 00:03:37.061 That's a small move with a long reach. 00:03:37.061 --> 00:03:40.768 We will use it again when an answer arrives as a beautiful paragraph. 00:03:41.303 --> 00:03:45.700 You will hear the phrase narrow AI for systems focused on one task or a small 00:03:45.700 --> 00:03:50.207 group of tasks. Other tools work across many kinds of tasks. 00:03:50.583 --> 00:03:54.956 Researchers still debate what should count as artificial general intelligence, or 00:03:54.956 --> 00:04:00.103 AGI. You don't need to settle that debate to ask useful questions about a tool. 00:04:00.500 --> 00:04:02.140 Can it do the job you need? 00:04:02.140 --> 00:04:03.600 Under what conditions? 00:04:03.600 --> 00:04:06.740 How often does it fail in examples like yours? 00:04:06.900 --> 00:04:11.540 A spectacular demonstration answers fewer questions than it first appears to. 00:04:11.917 --> 00:04:16.237 If it says it loves my hat, should I take the compliment? 00:04:16.829 --> 00:04:18.685 You may enjoy the compliment. 00:04:18.685 --> 00:04:23.265 But a sentence about a feeling isn't evidence that the system experiences that 00:04:23.265 --> 00:04:29.185 feeling. What an AI can do is different from whether it has feelings or awareness. 00:04:29.583 --> 00:04:33.103 In our classroom, characters have stories and personalities. 00:04:33.156 --> 00:04:36.847 In the lesson, we'll focus on what we can test and observe. 00:04:36.847 --> 00:04:40.223 Your hat remains beyond the scope of this investigation. 00:04:40.832 --> 00:04:43.772 These ideas did not begin with a chat window. 00:04:43.772 --> 00:04:48.047 In nineteen fifty-five, a research proposal used the name artificial 00:04:48.047 --> 00:04:53.025 intelligence and proposed a summer meeting at Dartmouth in nineteen fifty-six. 00:04:53.583 --> 00:04:58.063 Notice the two dates: the proposal, then the planned project. 00:04:58.183 --> 00:05:02.863 A timeline should help us understand what happened, not turn different events into 00:05:02.863 --> 00:05:04.463 one tidy birthday. 00:05:04.875 --> 00:05:08.235 I like seeing the work before the famous moment. 00:05:08.625 --> 00:05:12.625 Me too. A proposal is a beginning, not a finished machine. 00:05:12.645 --> 00:05:17.425 Let’s follow a few concrete milestones, and ask what each one demonstrated. 00:05:17.902 --> 00:05:22.713 In nineteen ninety-seven, IBM's Deep Blue defeated reigning world chess champion 00:05:22.713 --> 00:05:25.726 Garry Kasparov in a six-game match. 00:05:25.726 --> 00:05:29.193 It evaluated possible chess positions to choose moves. 00:05:29.917 --> 00:05:31.597 That was an impressive achievement. 00:05:31.624 --> 00:05:36.677 It was also a specific test, with a board, legal moves, and a clear result. 00:05:36.677 --> 00:05:39.117 Winning it was evidence about chess performance. 00:05:39.500 --> 00:05:43.180 So the headline doesn't tell me whether it can organize my desk. 00:05:43.583 --> 00:05:48.415 Exactly. Whenever a system succeeds, ask which task the evidence covers. 00:05:48.415 --> 00:05:49.727 Keep the achievement. 00:05:49.727 --> 00:05:51.343 Keep its boundaries too. 00:05:51.889 --> 00:05:56.372 In twenty seventeen, the paper Attention Is All You Need introduced the Transformer 00:05:56.372 --> 00:06:00.790 architecture. Attention helps connect relevant parts of an input. 00:06:01.167 --> 00:06:04.820 Look at this little example: The ticket is blue. 00:06:04.820 --> 00:06:10.627 It is on the desk. Understanding what it refers to involves a relationship across 00:06:10.627 --> 00:06:15.207 the words. Our picture illustrates that idea; it isn't a recording of 00:06:15.207 --> 00:06:16.447 a model's thoughts. 00:06:16.833 --> 00:06:20.113 The Transformer became an important design for language models. 00:06:20.353 --> 00:06:23.353 AI can also use other designs. 00:06:23.353 --> 00:06:27.393 One diagram cannot explain every system with AI on its label. 00:06:28.128 --> 00:06:31.152 ChatGPT launched in November twenty twenty-two. 00:06:31.312 --> 00:06:35.952 A conversational interface gave people a way to explore language-model capabilities 00:06:35.952 --> 00:06:37.312 by asking questions. 00:06:37.708 --> 00:06:40.428 Our timeline is a selection, not the whole history. 00:06:40.428 --> 00:06:43.536 The useful pattern is that different designs demonstrated different 00:06:43.536 --> 00:06:47.674 capabilities. So when a new tool appears, we look at what changed, and how 00:06:47.674 --> 00:06:48.588 anyone checked it. 00:06:49.068 --> 00:06:53.460 Now put a confident paragraph beside our misrouted invitation. 00:06:53.460 --> 00:06:57.712 Generative AI can produce false or unsupported information that sounds 00:06:57.712 --> 00:07:01.160 plausible. This is often called hallucination. 00:07:01.542 --> 00:07:04.982 Then how am I supposed to build anything on top of it? 00:07:05.375 --> 00:07:07.475 By choosing the job and the checks together. 00:07:07.475 --> 00:07:10.995 Imagine asking for three cheerful invitation drafts. 00:07:10.995 --> 00:07:12.975 You can judge the tone yourself. 00:07:12.986 --> 00:07:18.148 Asking for the library's actual opening time creates a different obligation: 00:07:18.148 --> 00:07:19.055 check the time. 00:07:19.458 --> 00:07:22.231 The same fluent voice can appear in both answers. 00:07:22.231 --> 00:07:25.218 Style doesn't tell you which facts were verified. 00:07:25.761 --> 00:07:30.075 Let's do the check. Our invented draft says the event starts at one. 00:07:30.075 --> 00:07:33.945 The matching entry in our invented library schedule says two. 00:07:34.025 --> 00:07:38.665 We check the event name and date as well, so we don't compare two different events. 00:07:39.042 --> 00:07:43.442 Then I correct the invitation to two, and keep the schedule with it. 00:07:44.333 --> 00:07:46.996 Yes. We have evidence for the correction. 00:07:46.996 --> 00:07:51.053 Asking the same chatbot, Are you sure, would give us another response. 00:07:51.080 --> 00:07:53.373 It wouldn't replace opening the source. 00:07:53.750 --> 00:07:57.930 For something consequential, choose checks strong enough for the consequence. 00:07:57.930 --> 00:08:02.390 A source must exist, be relevant, and support the particular claim. 00:08:02.510 --> 00:08:06.150 A link that merely looks official is only a starting point. 00:08:06.542 --> 00:08:10.942 And if I tell it the correct time, will it remember tomorrow? 00:08:11.467 --> 00:08:13.693 That depends on the app and its settings. 00:08:13.693 --> 00:08:16.901 A new chat may not include your earlier conversation. 00:08:16.901 --> 00:08:19.973 Some apps can save details or use past chats. 00:08:20.375 --> 00:08:24.645 Supplying saved context is different from teaching the underlying model a new fact 00:08:24.645 --> 00:08:29.855 through training. Check what the app says it saves, what you can review, and what 00:08:29.855 --> 00:08:30.935 you can remove. 00:08:31.333 --> 00:08:35.493 So I should bring the schedule again if the next task needs it. 00:08:35.875 --> 00:08:40.733 A good habit. Give the task the evidence it needs, instead of assuming yesterday's 00:08:40.733 --> 00:08:42.835 correction followed you into the room. 00:08:43.389 --> 00:08:46.950 It also helps to distinguish a model from the app around it. 00:08:46.950 --> 00:08:48.770 Think of our classroom project. 00:08:48.770 --> 00:08:52.983 We could use a model to draft wording, while the app supplies the schedule and 00:08:52.983 --> 00:08:54.690 keeps the draft in a document. 00:08:55.083 --> 00:08:58.043 If that schedule is missing, the task has changed. 00:08:58.150 --> 00:09:02.683 If the app adds a search tool, the available actions have changed. 00:09:02.740 --> 00:09:07.563 Comparing two answers without noticing those differences can be misleading. 00:09:07.958 --> 00:09:13.718 So when I ask what an AI can do, I should say which setup I'm using. 00:09:14.125 --> 00:09:18.091 Yes. The model, the information it receives, and the tools around it all 00:09:18.091 --> 00:09:19.405 belong in the description. 00:09:19.445 --> 00:09:21.965 Now we can make sense of the word agent. 00:09:22.582 --> 00:09:23.868 There's another distinction. 00:09:23.868 --> 00:09:26.748 An answer can describe what to do. 00:09:26.748 --> 00:09:32.095 An agentic system can use tools, inspect results, and decide what step to take next 00:09:32.095 --> 00:09:33.815 while working toward a goal. 00:09:34.208 --> 00:09:37.888 In our example, the goal is to prepare an invitation. 00:09:38.008 --> 00:09:40.648 A permitted tool reads the schedule. 00:09:40.648 --> 00:09:45.628 The system drafts the message, compares the details, and stops for your review. 00:09:45.628 --> 00:09:48.208 Sending isn't part of this task. 00:09:48.583 --> 00:09:52.663 I want that boundary written down before it touches my inbox. 00:09:53.042 --> 00:09:58.050 Yes. The tools, permissions, stopping point, and checks belong in the design. 00:09:58.050 --> 00:10:02.434 More steps can create more opportunities to help, and more places to 00:10:02.434 --> 00:10:03.362 make a mistake. 00:10:03.960 --> 00:10:05.430 Let's try one decision. 00:10:05.430 --> 00:10:09.238 The draft is ready, but its time disagrees with the schedule. 00:10:09.238 --> 00:10:14.760 What should the system do next: send it, inspect the schedule, or make the 00:10:14.760 --> 00:10:16.230 wording more impressive? 00:10:19.250 --> 00:10:20.754 Inspect the schedule. 00:10:20.754 --> 00:10:24.450 Beautiful words won't repair the wrong time. 00:10:24.833 --> 00:10:28.833 Right. The next useful action is the one that resolves the uncertainty. 00:10:28.873 --> 00:10:34.113 After the repair, we still review the message and its recipient before sending. 00:10:34.193 --> 00:10:38.513 Finishing a draft and approving an action are different checkpoints. 00:10:38.917 --> 00:10:41.477 What makes you excited about where this could go? 00:10:41.943 --> 00:10:45.365 The possibility of more people making the things they imagine. 00:10:45.365 --> 00:10:49.443 Someone getting help with a first draft, testing an idea, or finding another way 00:10:49.443 --> 00:10:51.107 into a difficult lesson. 00:10:51.107 --> 00:10:52.755 That's the future I'm rooting for. 00:10:53.167 --> 00:10:54.447 It isn't a guarantee. 00:10:54.447 --> 00:10:59.567 What people build, who can use it, and how carefully it is checked all matter. 00:10:59.627 --> 00:11:03.927 You can be curious about the possibilities and demanding about the evidence at 00:11:03.927 --> 00:11:04.847 the same time. 00:11:05.449 --> 00:11:08.797 Before we leave, close the notes for three questions. 00:11:08.797 --> 00:11:12.610 First: does every AI system learn from data? 00:11:15.625 --> 00:11:20.393 No. Some use rules and knowledge written by people, some learn patterns, and 00:11:20.393 --> 00:11:21.785 some combine approaches. 00:11:22.167 --> 00:11:25.447 Second: does a confident answer prove it is correct? 00:11:28.458 --> 00:11:31.498 No. Check the claim against relevant evidence. 00:11:31.498 --> 00:11:34.698 Our invitation sounded fine while its time was wrong. 00:11:35.083 --> 00:11:39.243 Third: what extra question matters when an AI can use tools to act? 00:11:42.250 --> 00:11:45.810 Ask what it is allowed to do, where it must stop, and how you'll 00:11:45.810 --> 00:11:47.050 check the result. 00:11:47.290 --> 00:11:51.530 If you remembered part of that, you have something solid to build on. 00:11:52.061 --> 00:11:56.737 Next time, we'll look more closely at how computers learn: examples, adjustments, 00:11:56.737 --> 00:11:58.397 and the test of something new. 00:11:58.437 --> 00:12:03.507 For now, choose one AI tool and name its task, its evidence, and one 00:12:03.507 --> 00:12:04.717 mistake you'd check. 00:12:05.125 --> 00:12:07.138 Rain doesn't fall the same way twice. 00:12:07.138 --> 00:12:08.805 Bring your questions. 00:12:08.845 --> 00:12:10.245 I'll see you next time.