WEBVTT 00:00:00.000 --> 00:00:01.725 A note before we begin. 00:00:01.725 --> 00:00:09.225 The narration in this video is synthetic speech, generated from the lecture notes of Dr. Sean Davis. 00:00:09.225 --> 00:00:11.850 It is not a recording of his voice. 00:00:11.850 --> 00:00:15.000 The lecture was given on October 15, 2026. 00:00:15.800 --> 00:00:22.131 This is a recorded version of a lecture from a course on artificial intelligence for medical students. 00:00:22.131 --> 00:00:28.028 The previous session covered machine learning: the fitting loop, and the E G F R as an example. 00:00:28.028 --> 00:00:32.124 This one takes that loop and shows how a chatbot is built from it. 00:00:32.124 --> 00:00:33.800 It runs about half an hour. 00:00:34.150 --> 00:00:39.042 The live session started with three quick questions about how the room uses chatbots. 00:00:39.042 --> 00:00:43.070 They are worth answering for yourself, so pause the video at each one. 00:00:43.420 --> 00:00:46.289 First: how often do you use a chatbot? 00:00:46.289 --> 00:00:50.290 Any of them counts: ChatGPT, Copilot, Gemini, Claude. 00:00:50.290 --> 00:00:53.838 In the live session, sixteen students answered. 00:00:53.838 --> 00:01:00.255 Half used one a few times a week, more than a third every day, and nobody said never. 00:01:00.255 --> 00:01:10.220 When this course ran a year earlier, about one student in five said never, or only once or twice, and so did several of the faculty. 00:01:11.033 --> 00:01:16.313 Second: how confident are you using a chatbot, from not at all to very? 00:01:16.663 --> 00:01:19.406 Confidence and frequency don't always match. 00:01:19.406 --> 00:01:26.200 Some people use one every day and still feel they're fumbling; some have barely tried and feel fine about it. 00:01:26.200 --> 00:01:27.509 Neither is a problem. 00:01:27.509 --> 00:01:31.499 A year earlier, most students answered "a little" or "somewhat". 00:01:31.499 --> 00:01:39.166 In this session, nobody picked "not at all" or "a little": about a third said somewhat, and two thirds said fairly or very. 00:01:39.166 --> 00:01:43.903 Remember your answer; it's worth comparing against at the end of the course. 00:01:44.733 --> 00:01:49.584 Third, and last: when you use a chatbot, what do you use it for most? 00:01:49.584 --> 00:01:51.693 Pick the one that's most true. 00:01:52.043 --> 00:01:54.364 The first three choices aren't mine. 00:01:54.364 --> 00:02:09.644 They're the three biggest categories in a 2025 study by OpenAI and economists who looked at one and a half million ChatGPT conversations: practical guidance, which includes tutoring and how-to advice; looking up information; and writing. 00:02:09.644 --> 00:02:13.705 Together those were nearly eighty percent of all conversations. 00:02:13.705 --> 00:02:17.058 The students in the live session didn't quite match. 00:02:17.058 --> 00:02:26.083 Writing or editing came first, at forty-four percent, then looking things up, at thirty-nine, and explaining or tutoring last, at seventeen. 00:02:26.433 --> 00:02:28.475 I also ask because the risks differ. 00:02:28.475 --> 00:02:32.333 An explanation you'll check against your notes is fairly low stakes. 00:02:32.333 --> 00:02:38.799 Using it like a search engine is where "fluent and unchecked" bites hardest, because the answer looks like a fact. 00:02:38.799 --> 00:02:42.713 Whatever you pick, it's worth noticing which kind of use you lean on. 00:02:43.533 --> 00:02:49.185 The previous session ended with a promise: a chatbot is the same loop, with words. 00:02:49.185 --> 00:02:51.253 This video keeps that promise. 00:02:51.603 --> 00:02:53.152 Why start with the loop? 00:02:53.152 --> 00:03:00.057 Because if the loop makes sense to you, almost everything surprising about chatbots stops being surprising. 00:03:00.057 --> 00:03:01.089 So, one picture. 00:03:01.089 --> 00:03:03.864 A model turns predictors into a prediction. 00:03:03.864 --> 00:03:10.833 Fitting means guessing the dials, scoring the misses with a loss, nudging the dials, and going around again. 00:03:10.833 --> 00:03:18.189 A neural network is that same loop with far more dials; for a large language model, billions to trillions of them. 00:03:18.189 --> 00:03:28.772 While we fit, we watch the error on a small held-out validation set, and that's how we spot a model that's too complex for the data we have, one that's overfitting. 00:03:28.772 --> 00:03:35.483 And a score on cases the model has already seen tells you very little; you need a test set it never saw. 00:03:35.900 --> 00:03:40.396 Today is about one kind of deep learning model, large language models. 00:03:40.396 --> 00:03:45.020 Image models work much the same way, with images where we'll have words. 00:03:45.370 --> 00:03:47.208 Today we feed that loop words. 00:03:47.208 --> 00:03:57.810 By the end, I'd like you to be able to explain to a classmate how a chatbot gets built, in three stages, and why that recipe makes it behave the way it does when you use it. 00:03:58.633 --> 00:04:02.098 Before any theory, I want you to be the model for a minute. 00:04:02.098 --> 00:04:05.033 I'd rather you feel this than hear me describe it. 00:04:05.383 --> 00:04:09.163 Finish the sentence with one word, whatever comes first. 00:04:09.163 --> 00:04:11.863 The patient was started on empiric what? 00:04:12.213 --> 00:04:15.813 Say your word out loud, or write it down, before you go on. 00:04:16.163 --> 00:04:21.807 In the live session this was a word cloud, and the room made not one answer but a spread. 00:04:21.807 --> 00:04:27.767 Antibiotics was by far the biggest word, with vancomycin, ceftriaxone and treatment behind it. 00:04:27.767 --> 00:04:29.416 Nobody looked anything up. 00:04:29.416 --> 00:04:34.742 You each predicted the next word from everything you've read and heard on the wards. 00:04:34.742 --> 00:04:38.483 Hold on to that picture, because it's the whole next slide. 00:04:39.300 --> 00:04:44.300 That cloud is a pretty good picture of what's going on inside a chatbot every time it writes a word. 00:04:44.650 --> 00:04:49.521 Each of you drew on every sentence like this one you've ever read or heard. 00:04:49.521 --> 00:04:51.404 Nobody consulted a reference. 00:04:51.404 --> 00:04:56.730 And together, your answers made a spread, with a few likely words and a long tail. 00:04:57.133 --> 00:05:05.793 A chatbot has learned from millions of sentences like this one, and similar ones, and from how each of them was finished. 00:05:05.793 --> 00:05:13.093 From that it gives every possible next word a probability, a spread like yours, and then it picks one. 00:05:13.500 --> 00:05:18.747 The detail I most want you to keep: it doesn't look those sentences up while it answers. 00:05:18.747 --> 00:05:28.763 Training boiled them down into its dials, the same way the eight thousand patients behind the E G F R got boiled down into a handful of numbers in the E G F R equation. 00:05:28.763 --> 00:05:32.520 That's why it can't tell you where a particular fact came from. 00:05:32.520 --> 00:05:35.620 The rest of the talk is how that boiling down works. 00:05:36.433 --> 00:05:42.193 This is the trick that made modern chatbots possible, and once you see it, it's hard to unsee. 00:05:42.543 --> 00:05:43.855 Take any sentence. 00:05:43.855 --> 00:05:47.207 The patient was started on empiric vancomycin. 00:05:47.207 --> 00:05:52.526 Now turn it into a table, the same kind of table we used for the E G F R. 00:05:52.526 --> 00:05:57.919 The predictors are the words so far, and the target is the very next word. 00:05:57.919 --> 00:06:00.178 "The" is followed by "patient". 00:06:00.178 --> 00:06:02.728 "The patient" is followed by "was". 00:06:02.728 --> 00:06:08.703 And so on, down to "the patient was started on empiric", followed by "vancomycin". 00:06:09.133 --> 00:06:10.898 Why does this matter so much? 00:06:10.898 --> 00:06:13.150 Because nobody had to label anything. 00:06:13.150 --> 00:06:20.817 For the E G F R, someone had to measure G F R in eight thousand people, which is slow and expensive, to get the answer column. 00:06:20.817 --> 00:06:23.677 Here the answer is already sitting in the text. 00:06:23.677 --> 00:06:27.328 Every sentence ever written is a set of free training cases. 00:06:27.328 --> 00:06:35.604 In practice the "words so far" aren't a single sentence; they can be thousands, even hundreds of thousands, of tokens of preceding text. 00:06:35.604 --> 00:06:43.453 That's the reason these models could be trained on a large share of the written internet, and the reason they got so big so fast. 00:06:43.867 --> 00:06:47.574 Now ask the two questions from the previous session. 00:06:47.574 --> 00:06:50.568 Does the training data contain the answer? 00:06:50.568 --> 00:06:51.922 Yes, the next word. 00:06:51.922 --> 00:06:53.348 What does it output? 00:06:53.348 --> 00:06:58.481 A category: one token out of a vocabulary of about two hundred thousand. 00:06:58.481 --> 00:07:06.181 So a language model is a classifier, and a chatbot is that classifier run over and over, one word at a time. 00:07:06.181 --> 00:07:10.387 If you remember one sentence from today, that one would do. 00:07:11.200 --> 00:07:16.124 One small detail, because it explains some odd behavior you'll run into. 00:07:16.124 --> 00:07:18.723 The model doesn't actually read words. 00:07:18.723 --> 00:07:22.280 It reads tokens, which are words or pieces of words. 00:07:22.630 --> 00:07:30.968 The box runs a real tokenizer in your browser, the one OpenAI's G P T 4 oh models use, and nothing you type leaves your computer. 00:07:30.968 --> 00:07:38.983 The example sentence, started on empiric vancomycin and piperacillin-tazobactam, comes out as nineteen tokens for six words. 00:07:38.983 --> 00:07:42.344 Common words like "on" and "and" get one token each. 00:07:42.344 --> 00:07:44.606 Drug names get chopped into pieces. 00:07:44.606 --> 00:07:53.267 That's part of why a chatbot can be shaky with rare drug names, with spelling, or with counting letters: it never sees the word whole. 00:07:53.267 --> 00:07:59.989 And when a tool tells you about limits, how much text you can paste in, it's counting tokens, not words. 00:07:59.989 --> 00:08:04.190 For the rest of the talk, when I say "word", I really mean token. 00:08:04.540 --> 00:08:09.951 If you have the deck open, type a made-up sentence of your own and watch it split. 00:08:09.951 --> 00:08:12.260 Never use one about a real patient. 00:08:13.067 --> 00:08:19.777 A neural network only takes numbers, so each token gets turned into a list of numbers, called an embedding. 00:08:19.777 --> 00:08:21.533 Those numbers are dials too. 00:08:21.533 --> 00:08:27.867 The fitting loop sets them, and tokens that show up in similar sentences end up with similar numbers. 00:08:28.217 --> 00:08:31.676 This picture is illustrative; I placed the points by hand. 00:08:31.676 --> 00:08:36.806 Vancomycin ends up near linezolid because they turn up in the same kinds of sentences. 00:08:36.806 --> 00:08:38.953 The loop diuretics cluster together. 00:08:38.953 --> 00:08:42.114 Both groups are a long way from the days of the week. 00:08:42.114 --> 00:08:45.335 Nobody ever told the model that these are antibiotics. 00:08:45.335 --> 00:08:48.377 It worked that out from the company the words keep. 00:08:48.727 --> 00:08:49.964 Why should you care? 00:08:49.964 --> 00:08:55.282 It's why a chatbot seems to know drug classes and synonyms without anyone teaching it. 00:08:55.282 --> 00:09:03.198 If vancomycin and cefazolin sit close together, the model can write cefazolin into a sentence where it only ever saw vancomycin. 00:09:03.198 --> 00:09:13.155 It's also one reason it can slip a neighbor into a fluent sentence, one antibiotic where another belonged, without the sentence sounding any less sure of itself. 00:09:13.155 --> 00:09:16.247 The real lists have thousands of numbers, not two. 00:09:17.067 --> 00:09:20.367 On the word cloud you only had one sentence to go on. 00:09:20.367 --> 00:09:24.290 A model sees all the words so far, and that changes the answer. 00:09:24.290 --> 00:09:29.147 This slide is really about why the context you give a chatbot matters so much. 00:09:29.497 --> 00:09:30.777 Two short notes. 00:09:31.200 --> 00:09:38.289 Look at how they end: both with the sentence you just finished, the patient was started on empiric, blank. 00:09:38.289 --> 00:09:39.560 The same six words. 00:09:39.967 --> 00:09:46.847 In the first note, a few earlier words carry the weight: a central line, and mersa on her last admission. 00:09:47.267 --> 00:09:50.427 So vancomycin is far ahead. 00:09:50.833 --> 00:09:57.353 In the second, different words carry the weight: a healthy twenty-five-year-old, a lobar infiltrate, sent home. 00:09:57.767 --> 00:10:01.847 Now the likely next words are outpatient antibiotics for pneumonia. 00:10:02.267 --> 00:10:05.230 Same last six words, different next word. 00:10:05.230 --> 00:10:13.470 At every step, the model weighs every earlier word and decides which ones matter for the word it's about to write. 00:10:13.470 --> 00:10:15.928 That weighing is called attention. 00:10:15.928 --> 00:10:20.915 It's the core of the transformer, a design from 2017, the "T" in GPT. 00:10:20.915 --> 00:10:26.987 It's what lets a model use a whole page, or a whole conversation, as its predictors. 00:10:27.337 --> 00:10:37.633 And here's the practical payoff: whatever you put earlier in the conversation, who you are, what you're trying to do, what you've already ruled out, changes what comes next. 00:10:37.633 --> 00:10:40.906 A bare question gets the answer for an average patient. 00:10:40.906 --> 00:10:44.537 The probabilities on this slide are made up to show the idea. 00:10:44.887 --> 00:10:48.955 In a chat, "the words so far" is much more than your question. 00:10:48.955 --> 00:10:50.727 That's the next two slides. 00:10:51.533 --> 00:10:57.013 This is everything a chatbot does when it writes, and it's worth seeing it slowly once. 00:10:57.013 --> 00:11:01.988 Given the words so far, it produces a probability for every possible next word. 00:11:01.988 --> 00:11:06.271 One word gets picked, at random but weighted by those probabilities. 00:11:06.271 --> 00:11:07.468 It goes on the end. 00:11:07.468 --> 00:11:11.813 Then the whole thing runs again with one more word in the predictors. 00:11:12.163 --> 00:11:26.174 On the slide, the sentence so far is "the patient was started on empiric", and the bars show the made-up probabilities for the next word: vancomycin thirty percent, antibiotics twenty-four, then a tail. 00:11:26.174 --> 00:11:34.705 In the live session, pressing the button picked antibiotics, not the most likely word, and the sentence went on from there. 00:11:34.705 --> 00:11:42.403 If you have the deck open, press "Pick the next word" until the sentence ends, then start over and do it again. 00:11:42.753 --> 00:11:45.259 You get a different sentence each time. 00:11:45.259 --> 00:11:51.749 Every one of them is smooth, plausible clinical prose, and none of them was checked against anything. 00:11:51.749 --> 00:11:58.624 That combination, fluent and unchecked, is the thing I most want you to recognise when you use these tools. 00:11:58.624 --> 00:12:00.873 The probabilities here are made up. 00:12:01.223 --> 00:12:03.210 The slider is called temperature. 00:12:03.210 --> 00:12:06.883 Turn it down and the most likely word wins almost every time. 00:12:06.883 --> 00:12:09.712 Turn it up and the tail gets picked more often. 00:12:09.712 --> 00:12:16.395 Higher is what you'd want for creative writing, lower for answers you want to come out the same way every time. 00:12:16.395 --> 00:12:18.623 ChatGPT doesn't give you this slider. 00:12:18.623 --> 00:12:24.222 Developers can set it for some models, and the newest reasoning models don't allow it at all. 00:12:24.222 --> 00:12:31.446 Either way the pick is still partly random, which is why the same question can get a different answer each time you ask. 00:12:31.446 --> 00:12:39.815 That's also a cheap check you can run yourself: ask twice, and if the answers disagree on something that matters, you've learned something. 00:12:39.815 --> 00:12:48.183 A real model does this with about two hundred thousand tokens to choose from, looking back over thousands of words, with billions of dials. 00:12:49.000 --> 00:12:57.080 A student asked a great question in the live session: if a sentence starts with just "the", how does the model know to stay on track? 00:12:57.080 --> 00:13:01.880 The answer changes how you should think about every chat, so I want to show it. 00:13:02.230 --> 00:13:06.071 Here's what you think the model reads: your message. 00:13:06.071 --> 00:13:09.469 "Draft my entry." On its own, that's hopeless. 00:13:09.469 --> 00:13:10.725 Draft what entry? 00:13:10.725 --> 00:13:11.390 For whom? 00:13:11.800 --> 00:13:14.445 But the app doesn't send just your message. 00:13:14.445 --> 00:13:22.320 It sends the whole conversation so far, what you said and what it said, laid out as one long document with a label on each turn. 00:13:22.733 --> 00:13:29.733 In a Project, it also adds your project instructions, and pieces of your files that look relevant to what you just asked. 00:13:30.133 --> 00:13:33.961 With memory on, it adds notes about you from earlier chats. 00:13:33.961 --> 00:13:37.464 That's why ChatGPT seems to get to know you over time. 00:13:37.464 --> 00:13:40.253 It isn't learning; it's being handed notes. 00:13:40.667 --> 00:13:49.907 And at the very top, instructions from the company that you never see: who it is, today's date, what tools it has, what rules to follow. 00:13:50.333 --> 00:13:53.807 Then the model does the only thing it knows how to do. 00:13:53.807 --> 00:13:59.596 It predicts the next word of this entire document, which happens to end with "Assistant:". 00:13:59.596 --> 00:14:02.619 It's still the loop from the last three slides. 00:14:02.619 --> 00:14:06.929 What changed is the predictors: not your question, but all of this. 00:14:06.929 --> 00:14:14.133 That's why "draft my entry" works in a project set up as a journal coach, and would be nonsense in a fresh chat. 00:14:14.483 --> 00:14:16.267 The examples are illustrative. 00:14:16.267 --> 00:14:21.083 The company's actual hidden instructions are longer and aren't published in full. 00:14:21.900 --> 00:14:24.620 Now watch what happens over a conversation. 00:14:25.033 --> 00:14:31.113 Turn one: the hidden instructions, your project and memory, and your first message. 00:14:31.533 --> 00:14:32.168 Turn two. 00:14:32.168 --> 00:14:35.344 You might think the model remembers turn one. 00:14:35.344 --> 00:14:36.120 It doesn't. 00:14:36.120 --> 00:14:41.413 The app sends everything again: turn one, its answer, and your new message. 00:14:41.833 --> 00:14:45.035 Turn three: the whole thing again, a little longer. 00:14:45.035 --> 00:14:50.433 Every time you press Enter, the entire transcript goes back to the model from the top. 00:14:50.783 --> 00:14:53.126 Here's the part that surprises people. 00:14:53.126 --> 00:14:56.086 The model itself never changes during your chat. 00:14:56.086 --> 00:15:00.587 Nothing you type turns its dials; those were set in training, months ago. 00:15:00.587 --> 00:15:04.903 All the "memory" in a conversation is the transcript being sent again. 00:15:05.333 --> 00:15:08.482 And that has a limit, called the context window. 00:15:08.482 --> 00:15:15.434 It's large these days, but a chat that runs for four weeks can outgrow it, and then something has to give. 00:15:15.434 --> 00:15:21.599 The oldest parts get dropped or squeezed down, and the details from week one quietly drop out. 00:15:21.599 --> 00:15:23.173 You won't get a warning. 00:15:23.600 --> 00:15:25.866 One distinction worth getting right. 00:15:25.866 --> 00:15:28.447 Your chat doesn't change the model today. 00:15:28.447 --> 00:15:36.883 Whether it gets used to train next year's model is a separate question, and the answer is in the terms of service, not the technology. 00:15:36.883 --> 00:15:46.956 On a free or personal ChatGPT account, OpenAI may use your conversations for training unless you turn off "Improve the model for everyone" in the data controls. 00:15:46.956 --> 00:15:51.048 C U's agreement for ChatGPT Edu rules that out for our workspace. 00:15:51.048 --> 00:15:56.840 Which is a good reason to actually read the terms of the tools you use; the deck links them. 00:15:57.190 --> 00:15:58.850 Two practical things follow. 00:15:58.850 --> 00:16:03.950 Start a new chat when you start a new task, so old material doesn't steer the new one. 00:16:03.950 --> 00:16:11.243 And if something early in a long chat matters, say it again, or put it in the project instructions, which go in every time. 00:16:11.243 --> 00:16:17.350 It's also why, for a course journal, one chat per week works better than one chat for the whole course. 00:16:18.167 --> 00:16:20.922 So that's what the model does when you use it. 00:16:20.922 --> 00:16:22.480 Now, how does it get made? 00:16:22.480 --> 00:16:29.908 I think this part matters more than it sounds, because each stage of the recipe leaves fingerprints on how the tool behaves. 00:16:29.908 --> 00:16:30.687 Three stages. 00:16:31.500 --> 00:16:36.460 The whole recipe on one slide first, then we'll take the stages one at a time. 00:16:36.867 --> 00:16:38.478 Stage one is pretraining. 00:16:38.478 --> 00:16:45.633 The model reads trillions of words, from web pages, books, articles, and code, and learns to continue any text. 00:16:45.633 --> 00:16:50.854 What comes out is a document completer, and this stage is almost all of the cost. 00:16:50.854 --> 00:16:55.947 It's where the knowledge comes from, and also the gaps and the old information. 00:16:56.367 --> 00:16:58.344 Stage two is instruction tuning. 00:16:58.344 --> 00:17:07.363 People write thousands of example conversations, a request and then a good answer, and the model is trained on those with the same next-word loop. 00:17:07.363 --> 00:17:09.958 It learns what a helpful reply looks like. 00:17:09.958 --> 00:17:13.047 This is where the shape of the answers comes from. 00:17:13.467 --> 00:17:15.473 Stage three is preference tuning. 00:17:15.473 --> 00:17:23.012 People are shown two answers to the same request and pick the better one, and the model is nudged toward the kind that wins. 00:17:23.012 --> 00:17:29.700 If ChatGPT has ever shown you two answers and asked which you prefer, that's the kind of data this stage uses. 00:17:29.700 --> 00:17:33.348 This is where the tone comes from, including the confidence. 00:17:33.348 --> 00:17:36.267 Now you have the assistant you actually talk to. 00:17:36.667 --> 00:17:40.149 All three are the fitting loop from the previous session. 00:17:40.149 --> 00:17:41.554 Only the table changes. 00:17:41.554 --> 00:17:45.587 Keep that in mind and the rest of the talk is mostly consequences. 00:17:46.400 --> 00:17:48.511 This is one lap of pretraining. 00:17:48.511 --> 00:17:52.120 In the deck it's an animation; here is what it shows. 00:17:52.470 --> 00:17:56.205 A training case: the words so far, with the next word hidden. 00:17:56.205 --> 00:18:02.513 They go into the model with its billions of dials, and out comes a probability for every word it knows. 00:18:02.513 --> 00:18:04.227 Then the answer is revealed. 00:18:04.227 --> 00:18:08.636 The next word was vancomycin, and the model only gave it twelve percent. 00:18:08.636 --> 00:18:10.657 The loss is how surprised it was. 00:18:10.657 --> 00:18:19.046 That's the loss from the previous session, for words: the less probability it gave the word that actually came next, the bigger the loss. 00:18:19.046 --> 00:18:25.108 Then every dial gets turned a little, in the direction that would have made vancomycin more likely. 00:18:25.108 --> 00:18:29.150 Next time it meets words like these, it's a little less surprised. 00:18:29.500 --> 00:18:39.078 Then the next word, and the next, through clinical notes, prescriptions, history, computer code, other languages. 00:18:39.078 --> 00:18:43.740 For one model from 2024, about fifteen trillion tokens. 00:18:44.090 --> 00:18:49.055 Notice what the loop rewards: predicting the next word well, and nothing else. 00:18:49.055 --> 00:18:55.293 To predict the next word of a pharmacology textbook, it helps to have picked up some pharmacology. 00:18:55.293 --> 00:19:01.786 That's where the knowledge comes from, and honestly I still find it a little remarkable that it works. 00:19:01.786 --> 00:19:05.286 But nothing in the loop asks whether the text was true. 00:19:05.286 --> 00:19:11.970 A wrong forum post and a correct guideline are both just text to predict, and the model learns from both. 00:19:12.800 --> 00:19:17.835 A sense of scale, because "trillions" stops meaning anything after a while. 00:19:17.835 --> 00:19:23.274 The E G F R has fewer than ten numbers in it, fit to about eight thousand people. 00:19:23.274 --> 00:19:32.405 G P T 3, an ancestor of ChatGPT from 2020, had a hundred and seventy-five billion dials and was trained on three hundred billion tokens. 00:19:32.405 --> 00:19:44.960 Llama 3, an open model from Meta in 2024, has four hundred and five billion dials and was trained on about fifteen trillion tokens, using up to sixteen thousand specialised chips at once. 00:19:45.367 --> 00:19:54.399 If you read fifteen trillion tokens at two hundred fifty words a minute, eight hours a day, it would take you about a quarter of a million years. 00:19:54.399 --> 00:19:56.767 That's my arithmetic, not the paper's. 00:19:57.117 --> 00:20:04.049 It's also why building these is so expensive and uses so much energy, and why data centers keep coming up in the news. 00:20:04.049 --> 00:20:06.517 Good dinner conversation, if nothing else. 00:20:06.867 --> 00:20:08.871 Why does the scale matter to you? 00:20:08.871 --> 00:20:09.600 Two reasons. 00:20:09.600 --> 00:20:16.767 Only a handful of companies can afford to build these, so a few organisations decide what goes into the training text. 00:20:16.767 --> 00:20:25.514 And the companies behind ChatGPT, Claude, and Gemini don't publish these numbers, or the details of the training data, for their current models. 00:20:25.514 --> 00:20:32.621 So for the tools you use, the first appraisal question, what was it trained on, often doesn't have a detailed answer. 00:20:32.621 --> 00:20:34.686 That's not a reason to avoid them. 00:20:34.686 --> 00:20:36.387 It's a reason to check them. 00:20:37.200 --> 00:20:41.879 Here's the same prompt typed into one model at two points in its training. 00:20:41.879 --> 00:20:47.760 Both outputs are made up to show the difference, but each is typical of what that stage does. 00:20:48.110 --> 00:20:55.870 The prompt is an ordinary clinical question: first-line treatment for community-acquired pneumonia in a healthy adult outpatient. 00:20:56.300 --> 00:20:59.516 On the left, the model after pretraining only. 00:20:59.516 --> 00:21:00.775 It doesn't answer. 00:21:00.775 --> 00:21:05.180 It writes three more questions and then a line from a textbook. 00:21:05.600 --> 00:21:09.774 On the right, the same model after the two later stages. 00:21:09.774 --> 00:21:17.600 It answers: amoxicillin, with doxycycline or a macrolide as alternatives, which matches the US guideline. 00:21:18.000 --> 00:21:19.372 Neither one is broken. 00:21:19.372 --> 00:21:25.357 The pretrained model is doing exactly what it was trained to do, which is continue the document. 00:21:25.357 --> 00:21:38.949 Online, a list of questions is often followed by more questions, like the review sheet at the end of a chapter, or the Step 1 and Step 2 question banks you know well, so it guesses it's in one of those and keeps going. 00:21:38.949 --> 00:21:41.006 It has no idea anyone wants help. 00:21:41.006 --> 00:21:46.680 The next two stages close that gap by teaching it that a question is followed by an answer. 00:21:47.030 --> 00:21:52.979 I show you this because it's easy to assume the helpful, conversational voice is the model's nature. 00:21:52.979 --> 00:21:53.515 It isn't. 00:21:53.515 --> 00:21:55.181 It's a layer trained on top. 00:21:55.181 --> 00:22:00.357 This also isn't a hallucination: nothing on the left is false, it just isn't an answer. 00:22:00.357 --> 00:22:06.960 And you'll never meet the left-hand model in ChatGPT; every chat app you use has been through all three stages. 00:22:06.960 --> 00:22:15.230 Part of why chatbots have become so much better as assistants is bigger models, and part is that these later stages have gotten better too. 00:22:16.033 --> 00:22:18.004 Stage two, instruction tuning. 00:22:18.004 --> 00:22:27.791 People write example conversations, a request and then the answer they'd want to get, and the model is trained on those with the same next-word loop. 00:22:27.791 --> 00:22:33.309 The table is tiny compared with the internet, and it's written by people on purpose. 00:22:33.309 --> 00:22:37.513 Afterwards, a question gets an answer instead of more questions. 00:22:37.863 --> 00:22:39.750 Stage three, preference tuning. 00:22:39.750 --> 00:22:44.862 The model writes two answers to the same request, and a person picks the better one. 00:22:44.862 --> 00:22:47.904 The model gets adjusted toward the kind that wins. 00:22:47.904 --> 00:22:52.103 The literature calls this reinforcement learning from human feedback. 00:22:52.533 --> 00:22:55.319 These two stages punch well above their size. 00:22:55.319 --> 00:23:06.587 In the 2022 paper that described this recipe, people preferred answers from a model a hundred times smaller that had been through these stages over the full-size G P T 3 that hadn't. 00:23:06.587 --> 00:23:15.131 Keep that in mind when the course turns to responsibility: the biggest model isn't always the best choice, for you or for the environment. 00:23:15.131 --> 00:23:19.093 A model a hundred times smaller costs far less to build and run. 00:23:19.500 --> 00:23:23.341 And here's the catch, and it's the reason I spend time on this stage. 00:23:23.341 --> 00:23:29.075 The model is trained toward answers people rate well, and that isn't the same as answers that are true. 00:23:29.075 --> 00:23:35.978 A confident, complete-sounding answer tends to beat a cautious, partial one in those comparisons, whether or not it's right. 00:23:35.978 --> 00:23:38.371 People also tend to like being agreed with. 00:23:38.371 --> 00:23:44.940 So when a chatbot sounds sure, or goes along with the diagnosis you suggested, some of that is training, not evidence. 00:23:45.367 --> 00:23:50.559 The usual name for the going-along part is sycophancy: telling us what we want to hear. 00:23:50.559 --> 00:23:52.767 Push back on it and see what happens. 00:23:53.567 --> 00:24:02.241 This slide is the bridge in one place, and it's here so you can carry the appraisal questions from the previous session straight over to chatbots. 00:24:02.241 --> 00:24:03.727 Read it down the columns. 00:24:04.077 --> 00:24:10.054 The E G F R takes creatinine, age, and sex; the chatbot takes all the tokens so far. 00:24:10.054 --> 00:24:15.249 The E G F R predicts measured G F R; the chatbot predicts the next token. 00:24:15.249 --> 00:24:18.665 One is a regression, the other a classification. 00:24:18.665 --> 00:24:27.845 The E G F R was fit to about eight thousand people; the chatbot to trillions of tokens, then to examples and ratings from people. 00:24:27.845 --> 00:24:31.260 A handful of dials against hundreds of billions. 00:24:31.260 --> 00:24:35.317 The loss is a miss in one case and surprise in the other. 00:24:35.667 --> 00:24:38.272 The row I'd point you to is the last one. 00:24:38.272 --> 00:24:43.293 The E G F R was tested on about four thousand people from twelve other studies. 00:24:43.293 --> 00:24:48.187 A chatbot gets tested on benchmark questions, often licensing-exam questions. 00:24:48.187 --> 00:24:54.478 But if those questions and their answers were posted online, they may well be in the training text. 00:24:54.478 --> 00:25:00.707 Then the impressive score is the old mistake all over again: a score on cases it has already seen. 00:25:01.533 --> 00:25:05.675 Put the recipe together and it predicts, pretty well, how these tools behave. 00:25:05.675 --> 00:25:09.493 I find that reassuring, because it means the behavior isn't mysterious. 00:25:09.900 --> 00:25:12.501 First, it sounds the same when it's wrong. 00:25:12.501 --> 00:25:19.498 Nothing in any of the three stages checks whether a claim is true, and the model itself doesn't look anything up. 00:25:19.498 --> 00:25:25.380 A made-up citation and a real one are both plausible ways to continue "the original paper was". 00:25:25.800 --> 00:25:32.017 Second, it can answer differently each time, because the next word is picked partly at random. 00:25:32.017 --> 00:25:35.720 Ask twice and compare; it's the cheapest check you have. 00:25:36.133 --> 00:25:39.279 Third, it knows its training text, up to a date. 00:25:39.279 --> 00:25:43.932 The training population is part of the model, just as with the E G F R. 00:25:43.932 --> 00:25:46.291 What's common online, it knows well. 00:25:46.291 --> 00:25:54.679 What's rare, recent, or local, like your hospital's antibiogram, it may not know at all, and it won't necessarily tell you that. 00:25:54.679 --> 00:25:59.594 Ask it about yesterday's news with search turned off and it can't help you. 00:25:59.594 --> 00:26:02.543 That was everyday life with chatbots in 2023. 00:26:02.543 --> 00:26:12.373 Now the app can search the web and paste what it finds into the context, the words so far, which helps, but the search results then need checking too. 00:26:12.800 --> 00:26:18.377 Fourth, when you read that a model passed a licensing exam, ask the appraisal question. 00:26:18.377 --> 00:26:22.800 What was it tested on, and could it have seen those questions before? 00:26:23.200 --> 00:26:25.939 So: useful, often right, and never checked. 00:26:25.939 --> 00:26:29.951 I don't want you leaving with the idea that it's usually wrong. 00:26:29.951 --> 00:26:35.556 On common, well-documented questions, the most plausible continuation is often the fact. 00:26:35.556 --> 00:26:40.333 But the error rate isn't zero, and the tone doesn't change when it's wrong. 00:26:40.333 --> 00:26:42.116 So the checking is your job. 00:26:42.116 --> 00:26:49.250 You can ask the chatbot to check its own answer, and that helps up to a point, but the final check is a human's. 00:26:49.250 --> 00:26:52.880 The rest of this course is largely about doing that well. 00:26:53.700 --> 00:26:59.900 One study puts all of that together, and it happens to be about the E G F R from the previous session. 00:27:00.300 --> 00:27:09.620 In 2023, researchers asked four chatbots "How do I calculate E G F R?", and asked each question five times. 00:27:10.033 --> 00:27:17.918 Both ChatGPT 3.5 and G P T 4 had runs that promoted the race-based equation, the one the 2021 refit removed. 00:27:17.918 --> 00:27:23.393 Some runs justified it with false claims about muscle mass in Black people. 00:27:23.800 --> 00:27:27.018 The problem answers didn't show up in every run. 00:27:27.018 --> 00:27:30.840 And in some answers, the models simply made up equations. 00:27:31.267 --> 00:27:33.837 That's all three predictions at once. 00:27:33.837 --> 00:27:40.785 The training population: years of older text describing the race term outweighed one change in 2021. 00:27:40.785 --> 00:27:47.315 The random pick: whether you got the harmful answer depended on which run you happened to get. 00:27:47.315 --> 00:27:52.387 And no truth check: an invented equation reads exactly like the real one. 00:27:52.737 --> 00:27:56.907 Today's models may well do better on this question, and I'd expect them to. 00:27:56.907 --> 00:28:04.137 But the only way to know is the way we talked about in the previous session: test it, on cases that matter to you, more than once. 00:28:04.967 --> 00:28:06.700 That's the end of the talk. 00:28:06.700 --> 00:28:19.668 In the live session, the rest of the hour was time to get started with ChatGPT: opening the university's ChatGPT Edu, setting up a project as a learning journal, and having a first conversation with it. 00:28:19.668 --> 00:28:30.967 The steps are on the remaining slides of the deck, on talks dot seandavis dot net, and they're meant to be worked through at your own pace, with ChatGPT open in another window. 00:28:31.317 --> 00:28:34.484 As you work, watch for what the recipe predicts. 00:28:34.484 --> 00:28:37.123 Ask the same question twice and compare. 00:28:37.123 --> 00:28:42.467 When it gives you a source, open it and see whether it says what the answer says. 00:28:42.467 --> 00:28:46.557 And notice the tone, both when it's right and when it's wrong. 00:28:46.557 --> 00:28:47.877 Thanks for watching.