# Stanford CS221 | Autumn 2025 | Lecture 20: Fireside Chat, Conclusion

## Метаданные

- **Канал:** Stanford Online
- **YouTube:** https://www.youtube.com/watch?v=5u5I5jvWR5k
- **Источник:** https://ekstraktznaniy.ru/video/20901

## Транскрипт

### Segment 1 (00:00 - 05:00) []

So, we're gonna have going to have our uh fireside chat session. Um going to have Percy come up um is our lovely um instructor for this class and um we going to let me change this to this window. So, all of you have submitted a lot of interesting questions and we'll try to hit as much of them as possible. Um but there is a schedule and there's no guarantee on what questions we will ask but please uh there will be live Q&A sessions and you can raise your hand and like shout and I will signal when that is. — Uh try not to you know uh to be too eager on asking your questions. — So the way this will run is that we have three sections. We have uh career life and research advice. We have class Stanford and miscellaneous questions. And then we'll have a section on AI and its outlooks. Okay. and then we'll go through them kind of like one by one. Um and then we'll like it will be more like a guided uh interview per se. — Okay, — sounds good. — I mean we can also go off script if needed. — Okay, so we're doing something different. This is a fireside chat. Um this is Percy. Percy, this is a very uh good way to get the you know personal perspective to you know Percy and ask questions. Um we'll start with the background in introductions. So what's your first AI class uh when you were undergrad? What' you take? What did you learn? Is it similar to 221? Is it not? — Yeah. Oh, you're asking me to remember something from quite a while ago. Um, yeah. So, I would did my undergrad at MIT. So, I did take the undergrad AI class there. Um, I remember it uh I don't remember it being at that time I was much more interested in theory and algorithms. So, it wasn't like I took that class and I was like, "Okay, I'm going to be an AI researcher, per se. " I think part of what I remember is that back then machine learning wasn't as kind of pervasive or important as it is now. And so, a lot of the more um you know, classical AI techniques um I could see kind of limitations in how they could uh scale. Um, so I remember taking a um a class on NLP and doing grammarss and we were just writing grammarss by hand which didn't feel very satisfying. It wasn't until later um during my undergrad when I uh did algorithms and then got into machine learning um and statistical language uh processing which was this was like uh early 2000s. This is when it first started kind of really taking off. Then I found this is like the way to combine really um kind of interesting math with uh kind of more scalable algorithmic approach. — So you mentioned that there were techniques like grammar um and some old techniques that didn't work as much. Were there a turning point where you were like you look at these techniques they don't work that well you could have switched to do something else and why did you stick with uh AI back then? So I mean I got into AI when I was already doing machine learning. In fact the first project I did back in 2005 was essentially what you could call training a language model that back then of course it wasn't a transformer. It was a hit a Markoff model. Um and there was a particular way you train it. Um but the idea was the same. You would take a large amount of text. um back then it was you know I think maybe a hundred million words um instead of you know trillions of words um and then you would um maximum likelihood it's the same thing we do now but with a different architecture and a different learning algorithm and at the end of the day you get this model which had features it would basically cluster the words and I remember seeing how um it would automatically figure out here's a cluster of city names or here's the cluster of days of the week and so on and that was a moment when I thought wow this is really I guess you could call it emergent capabilities and that's why I um decided to go pursue AI — that's very interesting journey um 10 years later or even 20 years later now the world is very different um did you expect any of this to happen like language models to dominate um the AI paradigm or you know were you back then still thinking about other kinds of like techniques and — so it's interesting because even 20 years ago there's this idea that training generative models on uh large amounts of data can be useful but I don't think I or really I think anyone had the imagination to think that if you just kind of certainly well I'll just speak for myself but I think mo most people did not as well have the imagination to think that if you just

### Segment 2 (05:00 - 10:00) [5:00]

really stuck with that idea and carried to its logical conclusion that you will get you know GPT GPT3 and GPT5 and things like that. Um I think that the transition towards interesting word clusters to like zeroing a bunch of different tasks that was a pretty major leap that um even I think if you talk to the open AI researchers I think there was a conviction that this was the path but whether it came would come in 2020 versus 2030 or 2050 I think no one knew. Now is a good time to transition into AI outlook. Um we can ask about uh what we think about AI being happening now, what's going to happen in five to 10 years and we get person thoughts on that. So you know what what's like one concrete thing that changed about all of this uh in your mind when you think about AI systems in the last 3 years seeing all this happen like maybe before you were thinking more perhaps in terms of like principles and techniques and algorithms but right now it's just data and compute um is there anything else that really drastically changed your thinking seeing this progress? Well, I think the big shift I think if you think about AI is that before it was very much a researcher thing, right? AI researchers, they do a bunch of experiments, they write papers. Um, and now you walk on the street and everyone's talking about AI. You drive on the 101, there's like AI billboards and it's become not really about a research thing, but this is about a kind of a global phenomena. I mean this happens like it happened with the internet for example as well you know the I guess you know I'm probably um not many people were around uh when that was happening but um the so that's a big shift so I think a lot of the conversations are have broadened instead of thinking about just the techniques and the principles but you know think about Richish's lecture from last time you think about you know data and energy and you know compute and resources and think about jobs and all these other things. Um because AI has kind of escaped the lab in some sense and has been um is actually having real impact in the world in a very substantial way, right? when you have the entire policy national policies being written about AI or you know you have a AI for strategy or companies calling them like the biggest companies are call themselves AI companies I mean this is like a pretty big you know shift um I think it's important to remember that the research still continues right my group many people at Stanford other universities uh folks in industry there's still a lot of open problems um and things that we don't understand and things that could be a lot better. Um, but that's always been the case and always will be the case. It just happens that there's sort of this off-ramp, if you will, from the research highway where it's like, oh, okay, now we can actually get a bunch of kind of impact here and now as opposed to always thinking about the future. M [snorts] um you mentioned a lot of things there about like the role of like research and then the public perception hang bills on 101. Let's start with the public perception because we're talking about how AI is uh happening today. You know what's something that you think is very correct and very incorrect about the public perception of AI today? um when people think about AI, what's something that you want to tell them that um you know this is the wrong way to think about it or this is the right way to think about it. I think a lot of AI perception unfortunately has been kind of shaped by you know science fiction um you know the Terminator and all those like actually it's interesting because in the US al so it's very cultural like the way that people think about AI is very different in the east than in the west I think there's much more of a kind of a dark bleak outlook um here than whereas in Asia I think it's much more of a optimistic uh well this is a technology that really transform. So that's a kind of a first thing that's um interesting. I personally think that there's a lot of like baggage associated with kind of you know science fiction. Um, and you know, I think things could be either better or worse, but in very different directions than what's people typically think when they're anchored to, you know, the idea of a sentient agent that comes and either, you know, destroys. I think that the way I think about it is

### Segment 3 (10:00 - 15:00) [10:00]

that AI is already everywhere, right? It's making deci decisions um in much in the background. So think about it more as um you know infrastructure right like or what recommendations you get on your phone. Now I think with you know chat GPD it's more kind of userf facing but I think the way that AI will influence our lives is much more pervasive and much of it might be hidden um compared to you know some AI walks through the door right that's really good for movies um because it's hard to have like I don't know some abstract thing kind of walks through the door — yeah working on your spreadsheet you know that doesn't — yeah that doesn't seem very uh kind of Hollywood would um on that note, what's the most overhyped capability and underhyped capability of current AIS? — Well, um I think I'll start with the underhyped one. Um and this is specific to language models, right? I think we language if you think about for much of it their existence, they're probabilistic distributions over the next token given the previous token, right? So they're mathematical objects and you can study them as such and now they've transformed into essentially more like a system that you give in an input you get an output. So one of the things that I think is underappreciated is that how much that kind of probabistic view and how much like minimizing perplexity and having strong foundations it really enables all of this right there's a lot of you know people think about as pre-training and post- training so Ken's lecture I think gave a good overview of that um and a lot of the visible things come from post- trainining like you teach the model how to solve math problems or do coding or give um you know analyze financial documents or what have you. Um but I think the it's maybe obviously hidden that many of the things that make AI tick is the ability to predict the next token. And one way you can see this is if you look at the ability for AI language models to predict the next token as a function of um sequence length. Like a good system or a good model should really be able to kind of uh kind of deeply understand the context and you know lower your loss as you predict the next token. — Um even to you know let's say a million um tokens. um and measuring that um loss is something that really gets at the true underlying intelligence or capability of a system which doesn't really I think you know show up I think when you look at these public leaderboards — that makes sense — um maybe overhyped um thing is um obviously a lot of people are interested in thinking models and reasoning um honestly if you look at least when I look at some of these thinking traces is it just seems like it's a very long sequence of very rambly a lot of inefficiency there and yes you get eventually to the right answer and you get good scores but it's like man this must be a scam to get you to generate more tokens that's what it feels like — in fact even today we don't we can't even be sure that uh the thinking like we don't even understand how the thinking trace precisely how the model is it just more budget or is it just actually guiding the thinking Um sometimes the thinking trace can be just wrong but still get you the right answer. Okay. Now let's look forward a little bit. Um you mentioned that there's still a lot of research to do. Um some people will agree with that, some people may not agree with that. Um in particular some students may think that you know the role of academia is diminishing like there's just not too much uh work to be done at school because we don't have the resources. Uh we don't have compute. um what do you have um what are your thoughts on that and you know is AI already solved really we have the bulk of it already yeah or are there really more breakthroughs needed — yeah so I guess this is where being a bit older helps you can see the over the last you know 20 years I think academia has always been a very small fraction of like what's happening in the world right we're always like this kind of forwardinking weird group of people who do weird things and if you look let's say TW even 10 years ago or maybe let's say 20 years ago and if you were to go at Google and say like oh I have this really great machine learning algorithm and we should use it in search they'll you just got know they'll show you to the exit because that it was sort of understood that well these things aren't reliable they don't work and so

### Segment 4 (15:00 - 20:00) [15:00]

AI I mean research has always been in this position where um it's too it's like barely working right that's the definition and what has happened over the last five years is all the very methods that people are working on with scale of course and data and compute start working and now there's a sort of now if you go to Google of course they'll welcome you in it's like go work on Gemini um and But um there's a sort of we're at a sort of transition period and it's not that academia is being less uh relevant but more that um the techniques were suddenly becoming more relevant like there's a sense in which like many of the techniques graduated um from school I guess that's kind of apt — um but there's always more things like the some of the fundamental questions about how you can really generalize well. Um, and how do you be become more data efficient and all these things are not solved by kind of any u means but that doesn't stop you and it shouldn't stop people from trying to productionize the um the current you know best. Um so one class of things that academia is still really you know should be doing I think and can be doing um is kind of much more longer term kind of blue sky research. Um I think some of the feeling that you might get when you talk to peers is that you know they don't see where that is because I think part of the thing is that the thing that they're everyone has looking at which is these particular transformer models if you follow them well they're graduating industry but if you're looking over there of course you know you're not looking at the new opportunities the other thing else other class of things I think is um is that there are a whole class of problems that academia is very well incentivized to do but industry is just uh not for example we've been doing a bunch of work on understanding you know copyright know how much memorized content can you get out of language models right so clearly labs are not going to go do this sort of work because they're already getting you know sued and so any academia since we have no skin in the game here is uh the only ones that can actually do this work. Or if you think about evaluation, how do you evaluate models in a fair way or expose problems with these evaluation um with these models? This is something again which is um harder to do in certain contextes in industry because there's a conflict of interest like you have to you know always be pushing the capabilities and showing that your model is great rather than finding necessarily flaws uh with it. — Yeah, it's a it's very interesting. Um on the role of academia there's both research and education. So you touched on a lot of research problems that are not yet solved. Um but on education side you know a lot of us are CS students. Uh we have a lot of undergrads here and one of the things that uh students commonly worry about is that I'm going to graduate and go into the software engineering market. Um but today the AI is already very good software engineers. They can implement a lot of things much faster um than the average graduate. — Um what do you think about that and what the role is the university to prepare them into this market? — Yeah. So, so Richishi showed a nice slide from last lecture which is it is true that the number of entry-level software engineers as it's currently defined by the skill set that has been traditionally the case for the last 10 years is going down. So, you can't expect to just do business as usual, learn the same, take the same CS classes assuming that they haven't aren't changing and just get a um a job. Um I think that this is a very much a transition period where you know I as an instructor and others in the department are really trying to figure out um what education and the workplace should look like and it's a sort of a joint um problem. I think at a general level I'm not you know concerned that um you know people won't have jobs. I think that um you know the types of things you might have to do will necessarily change. If you look at previous generation of technologies like you know from um being able to do um you know before you if you could mechanical computation really quickly that earned you a job right and then calculators came out and they're like okay you can't get a job by just like multiplying numbers all day long but

### Segment 5 (20:00 - 25:00) [20:00]

that doesn't mean that there are other things to do and in fact I think there's once you get the primitives where oh yes oh I just have to you know specify what I want and the AI will just write it frees up opportunities to build other things right in the limit if you think about um you know think about the CEO of a kind of a company right he's probably the busiest person it's not that um and there are people kind of underneath which have who have many more at least technical maybe other depth in terms of depth have way more skills but there's still sort of a job and not just the CEO but probably the whole management chain of essentially directing and also um deciding what needs to be done. So I think long story short there's a shift more towards from doing to actually figuring out what to even do in the first in the first place. Um, and that is actually a really maybe that's one kind of takeaway is that you know figuring out what needs to be built like suppose you had a tool that could build any app in like uh five minutes like what would you build — right and thinking about that question uh which is not a I mean that's not a trivial question like what app would actually take off or be useful to people that's a very deep question so focusing on those uh problems s um like the people who can do that well um are people probably going to — you know be able to survive in this new — um environment. — So you mentioned like um finding what to build is a very important question. So from your perspective in the next 5 to 10 years or even next 3 years what are the application areas that you think are undervalued that people that students should think more about building in you know um they have the skill set of uh the CS education but what should they do like what's undervalued right now? — Yeah. Um I mean one perspective is that a lot of AI is almost synonymous with these like frontier models which are essentially um everyone gets exposure to AI systems like Gemini or Chad GBT. But I think if you think about the underlying technology, the idea of training a foundation model on tons of data and unlocking new capabilities, this is an incredibly general purpose um technology, right? And it could you've seen it successfully applied to DNA sequences. Um and if you think about the amount of other data out there, whether it be know climate data or um you know time series or kind of uh um other you know satellite imagery there's many other um data sources where the same type of intuition should be able to carry over and I think there's a lot of uh we're just t I think just uh scratching the surface of um this technology that can be can unlock a lot. So I encourage people to think outside here's yet another um you know an AI assistant um and think more about all the other areas outside CS in other disciplines like you know in the sciences, physical materials um neuroscience and think about where there are opportunities to apply um the same techniques. — All right. Um we now have a slot for live questions if anyone want to ask a question. So the question is uh what is the your projected impact of AI? Is it just a tool or is it going to be like drastically transformed society? — Yeah. So if you look at Richishi's slide from last time is it the technology of the year to technology of the the decade or the century or um I guess or ever? I would say yeah century would be kind of my guess. I mean I do think that this is a pretty like the m the general purpose nature is pretty you know powerful in a way. Um I mean okay I guess century well I think about computing as other thing computing was like a huge deal and that was what I guess 70 years ago. So this is probably a kind of a similar um magnitude of of unlock. Do current frontier LLMs pass the touring test? If so, what's the next goal? Has AI been solved or only scaling and incremental improvements? — Yeah. So, I mean, I guess the first thing to address is whether the touring test is the right way to measure it and how it's defined. Um I mean in terms of um you know I think on the first day of lecture I mentioned the terrain test and

### Segment 6 (25:00 - 30:00) [25:00]

I think it's historically been kind of important in many ways to shape the development. I think right now I don't think it's like a very uh um a good measure of what it means what progress um means both the imitation part and also humans being the benchmark. Um, I think that AI should be we want AI to be kind of way better than humans um at certain things which should be much more reliable for example. So putting humans at a bar um doesn't seem particularly um it seems like uh a bit shortsighted I think. — Would there be a easy to understand test like a new test new touring test that you have in mind? I think in some sense um often tests are static which means that you have to um you know you have a system it enters a room and then you get a score out but I think that obviously has kind of problems because they're gameable and so on. I think a better metric is just um kind of in the online sense though. So the way I think I would propose is you know how many new scientific discoveries are is AI you know producing — and that is not gameable right because if a you know AI cures cancer or does some you know invents some kind of new m light material or figures out you know uh you know fusion or something that's going to be there's no one can disagree like oh okay that was cheating because if you cheated somehow and you figured out how to cure cancer. Well, good for you. — Yeah. — All right. We'll now switch there to talk about Stanford and classes and uh CS21 in particular. So, we'll start with um the class that you're all in right now. So, you know, when CS21 started 10 years ago almost and um you know, this year we're thinking about how to make it better. We're trying to design it better. So, when we design it or when you design it trying to make it better, what were you thinking? what problems were you trying to solve to make this class better as a introduction to AI class? — Yeah. So, there's a few years along the way where we've made large changes to the class. Um, and this was one of them. Um, and there were multiple objectives here and in hindsight it was probably a bit ambitious in terms of how much we wanted to do. Um but um first um thinking more about the social impacts of AI that has always been as I mentioned a few minutes ago hard to kind of incorporate because from the ground up when I designed this class from scratch like 11 years ago it was a technical class we didn't worry about or think about the societal impacts as much and so the whole structure of how the class is built doesn't really naturally incorporate that in but at the same time it's clear that a intro to AI class needs to confront some of these issues. So one of the things we did is in the latter part of the class have dedicated a homework and some lectures to that. You know obviously this is kind of work in progress and we can continue making it better and if you have feedback love to hear it. Um the other major thing we did um was um I've always found it um the disconnect between kind of lectures where you're looking at um abstract diagrams and you know there's sort of like you know you look at something you think you get it and then you go to the homework and you try to code it up and you realize wait actually I don't understand how this connects and one of the things that was successful when I taught the 336 class uh language models from scratch is to use this executable lecture format where we walk through code so there's no ambiguity about what objects we're you know talking about and so this year has been a bit of a pilot in terms of bringing that format to this class now obviously there are um for the first time something happens there's always kind of some rough edges so I appreciate everyone's patience with that. I do strongly believe that you know iterating on this uh will produce just a much better um experience which uh combines the both of a to have intuitive um concepts but also grounded in the code. So um you can kind of see the whole stack. So 11 years ago when this class was designed um to some extent we didn't have quote unquote working AI. So AI is a concept. It seems to be applied in narrow domains. But right now every one of you have this chat GBT this UI thing

### Segment 7 (30:00 - 35:00) [30:00]

this cursor that are actually working and we spend a lot of time on things like search MDPs graphical models and logic and none of that seem to be connected to this uh abstract AI product that everyone's using today. You know um how should students mentally connect these pieces to this large language models paradigm that we're in right now? Yeah, I think as hopefully I've tried to convey that many of these concepts are actually quite uh relevant. So um in search for example um you know we think about states and we think about actions and exploring different possibilities um this is exactly what kind of test time compute with language models is about. It turns out that you can't just train a model and then be done. there's an inference time uh problem of given a model how do you search over different kind of solutions and if you think about solving really hard problems um like uh I guess a problem or just doing some sort of scientific discovery or writing a you know doing some data science problem um it's going to require trial and error right and that's search um so and Markoff decision process processes, MDPs, which is uh it's very much like in language models when you pre-train a language model and then you do reinforcement learning that's basically trying to um learn a policy that can give you good you know results. So that's um clearly um you know related. — Any other ones I should defend? Well, I mean that cover a lot of them, but the other part we want to cover is more on a meta on a more meta level. Uh CS2:1 is a introduction to AI class, but AI is here. Um and AI is like very productive. What should such an introduction to AI class really prepare students for? Apart from just the concepts and knowledge, is there anything else such a class should do for students? — Yeah, I've always been a strong believer that you should understand the fundamentals of how things are, you know, built and kind of peeling over the layers until you get to the bottom. I think there's many things um you can do if let's say you just want to get a job, right? There's many online resources [clears throat] and tutorials. you can basically learn how to do ML engineering or whatever you want and you know be just fine. But in some sense the whole idea of having you know why you're at Stanford and having a kind of a general education is the sort of a longer term thinking of both the breath of topics like things that you would just like never encounter in the short term if you're in a kind of a job on the job learning on the job. Um, but also thinking about kind of a layer above the kind of okay, I need to get this PyTorch program kind of working. Um, and so 221 tries to lay out, you know, some of those abstract ideas and also ground them into code. — But now ask some very quick questions on classes. These are very common ones. How would you compare 2 to1 to other Intel TI classes such as 2 to9? Uh, is 2:1 enough to get students to the next level classes such as 36 um or 2 to 4N? Um, — Yeah. Um, I think by now I think a lot of these classes so the short answer is yes. I mean after you take 221 certainly you can take 229 and 224n. Um in fact a lot of people also take them in different orders. Um uh because some of the topics are sort of or orthogonal. Um I think that um you know the as for 336 um this is probably a kind of a different level of a class where it does require a bit more kind of not so much knowledge but just experience um and um I guess some amount of grit to go through um building all of the whole language modeling stack from scratch but certainly this provides a good foundation. So the question is what's your re approach to research and how do you come up with topics? — Yeah. Um yeah figuring out how to do research. I mean it's a it's definitely not a easy thing and certainly I wouldn't say that I automatically know what research topics are going to it's like kind of betting on the future right and you can only you know to try your best. Um the way I think about it is research is about taking bets. So I think zooming out and looking at the

### Segment 8 (35:00 - 40:00) [35:00]

world and seeing okay what is like a big hole that doesn't um that needs to be filled right now right I think the best research are um ones where there is some sort of vision on what you're trying to do and at the same time there's a concrete first step you can take um so I mean just one example this is you know just the first thing that comes to mind is a few years ago I was working on thinking about like decentralized training. So the idea was that instead of having current language model training which is every um all the compute has to be in one place. Imagine if every one in this room and everyone in the world could plug in their own compute and you could have this peer-to-peer network and could you train a large uh foundation model. So this is obviously uh would very much change the um the kind of the power dynamics and how what it means to even train a model and has implications AI policy and all these things and I think of that as like a very it's interesting research direction and now you can work on the kind of first technical problem which is more of a um kind of a systems problem um and so there's I guess the Another perspective and this is more about research taste is that I often try to think about research problems where I generally don't know the answer right and this is like if I basically like things that are in the boundary of like what could work and what could not work u because research part of research is about information gain it's not about necessarily oh I can do this really quick thing and I can get on, you know, get a 5% improvement because, you know, there's that's in let's say 10 years. You know, no one will care about that unless there's some sort of more generalizable lesson you can take away. — Any other questions on Stanford classes 2 to one and take one from the list. Um, [snorts] so there's some questions about advice in the CSP PhD, which we'll get to in the next section. Career life and research advice. Um, we hit on uh hit this one already. Okay, this is an easy one. How do we increase the chance of be becoming a 2 to1 CA? — Um, at this point — perhaps next cycle, unfortunately. Um well I mean yeah doing well in the class and being uh you know proactive I think is something that we like — or be like these people in the front — or be persist PhD student — um you get signed up you don't apply um okay so let's switch to the next section about career life and research advice um so here uh I think is where we get the most questions Let's just start with like um u uh something simple like today uh what would your first job be if you were undergrad here at Stanford? You take these classes. You learn how to do uh CS uh you know implementations and a little bit of research you know about all this. What would your job be? — What would I want it to be or — what would you do? — Uh so I graduated from Stanford. — Yes. — 256 — six. — Yeah. And then I go and um — apply for jobs. — Kind of a strange uh question. I mean I think um I mean okay so what are the options right? You can go to um kind of a big research lab and do build um you know the next generation of a AI. Um I think you can do a startup — um or um you can go to grad school. — Um I guess you could go to law school or business school or any of these things but that's not a job I guess. Um, I think that um I mean I guess the um I mean this is going to sound a bit cliched but I mean I do think that there's a lot of interesting you know okay so backing up like I think um you know when I graduated I had like zero interest in startups. Maybe that's because I graduated from MIT or something. Um and recently I have been more engaged in the SAR ecosystem

### Segment 9 (40:00 - 45:00) [40:00]

and I do think that there is something kind of really cool about a small set of people um working closely together on some you know problem. So I mean it doesn't have to be a startup but I think the idea of kind of being very missiondriven and trying to work on something that requires a bit more than just kind of an individual level contribution seems would be seems most kind of interesting. Mhm. — Um and then I think the you know I am obviously I'm a researcher so I think you know being able to work on interesting research problems would be something I would want to do — for many students here uh I think students are facing this choice between you know research industry startups all the things you mentioned but uh you know they wouldn't know what to choose these feels very exciting all feels lots of possibility there like what would you do like what's your you know value function there or what would your rule of thumb be? — Yeah, I mean pro probably I'll try to answer the question I think uh people want to um answer which is less about what I would do but um it would be like how what's a framework to think about things. I think when you're graduating, right, this is your first job. Um, it's most definitely or most likely will not be your uh your last job. So, you're not you're not like getting married or something. It's like something that is probably you're going to spend a few day years there. So, the thing I would prioritize most is growth, right? I think you know you graduate here you learned things at Stanford but the next part of your journey is still education it's learning it's just a different type of learning um so I would heavily prioritize I guess maybe the more to use RL analogy exploration rather than exploitation part um as okay so what based on that what would you do like being like with good people who you enjoy working with I think is the most important thing. I think it everything else is I think actually you know secondary in some sense. Um you know it could be in academic setting a startup setting you know group but um in a larger company as long as you feel like you are learning what you want to learn. Now of course I think if you have a mission like you want to go you know do invent and you know solve the world's energy problems or like you have a mission then I think it's actually much clearer right but I think that question probably comes from a point where well there's a lot of interesting things I could do what do I do and then I think uh using um like where would you learn the most would be the my top choice — that makes a lotense sense. Um, commonly students would also feel that, you know, if they're not doing this or this by, you know, sophomore year, they're doing not doing XYZ by their, uh, junior year, they feel they're lagging behind. Um, you know, it's almost like a competition. It's almost like, uh, I have to, you know, try this or else I'm too late. Um, do you have any advice to that kind of like mindset or feeling? Yeah, I mean I get it how there's a lot of, you know, pressures and you always you can it's you look at people around you and they're doing all these, you know, crazy things. Um I think to put things in perspective though um if you look at top PE researchers at labs or people who are like doing or CEOs or people doing very well you realize that I mean many of them took a very securous path and maybe they went to a different field alto together and then transitioned like how many people transitioned into AI um and they were doing something completely unrelated. So I don't think that what you what internship you get in sophomore year is really going to have any sort of meaningful [snorts] you know impact. I think it's um still about kind of you and what your skills are and what you learn and your outlook on the world — that is the most important thing and I think any of these kind of honors so to speak — like it's it becomes very clear I think um when it's just like oh this person has just racked up a bunch of like you know good sounding names but actually isn't as qualified as someone who maybe hasn't done as much but you know they're actually um be in skill-wise better positioned to succeed.

### Segment 10 (45:00 - 50:00) [45:00]

— I have two follow-ups to that. One followup is you mentioned skills and like knowledge and applying them but the world is changing so fast right now. There's so many like the things that AI couldn't do six months ago that they now can do. Um as junior students you would worry that your skills would be outdated. Um so what would you say to that? And second is from your perspective, how do you then val evaluate skills? You're saying that maybe some person had this shiny CV, but that might not be a better person compared to someone who hasn't done all these things. — Yeah. I Well, okay. So maybe the first part — so one skill that will never go away is ability to learn and adapt, right? I mean that just almost by definition, right? In a fast changing world, the only thing that matters is how fast you can quickly adapt. Like if the model can do your job, okay, how can you learn something new? Which is why I emphasize so much like can you learn not just learn deeply but also you know quickly — um because that will be the most kind of important thing I think in a fast u moving world. Um and then there's a second part — same part is how do you evaluate skills and learning? So basically how do you interview people in Sanche? Yeah that is a tough um problem. I think there is um you know there are some like obviously there's the usual tech uh screens and you know getting a sense of how you can do work trials um to evaluate how people work together. Um I mean I think there's a bunch of other you know softer skills for example there's a sense of which I know having you know grit and passion and those type of things really do kind of matter. Um because the difference between someone who like really is into something and cares and someone who's just like kind of doing a job is pretty um day night and day. Um also you know working I mean again this is you know kind of cliche but like working well with others um collaboration is something that um actually I mean it's just like very so important in any sort of context and this is something that I think um you know we don't train students that well I think um at least not explicitly because most work is individual and evaluation is supposed to be individual And sure, there's some group projects, but those are um not really, I think, the main, you know, focus. So, but in the real world, everything is collaborative and you have to be able to interact with other people. — That makes sense. Um, what's a common piece of advice that you hear uh for career that you actually don't agree with? — Um, that people say, you know, you should do this, you should do that, — and but you don't think that's true. Um, oh, I don't get very much advice these days. — Um, I mean, — yeah, I mean, maybe along the lines of like, oh, advice, you have to go apply to this internship otherwise you'll be left behind or something. — Um, I think that or [snorts] — yeah, maybe it's just kind of the same answer as before. — I see. — Um, so, you know, AI is very hot right now. Is there a bubble? And does the bubble, if there is, does the bubble distort student choices? And do you have any thoughts on that? — I mean, surely there's a bubble. — I mean, that's you might think there's not a bubble. Um — I And certainly that just I mean it ds not just student choices, but it distorts like everything in the world, right? Like if you look at the investments that are made in like big companies and decisions that governments take. I mean it's just like very much u shaping how the entire world is operating. — Um I don't think yeah I wouldn't say that's necessarily all kind of bad because I do think that AI is real. Like if you think look think about the internet um revolution right I mean that's I think a prime example where there's clearly a bubble and we know what happened in you know 2099 2000 um and yet internet is like completely transformative like it's like just very much changed the way that we've uh we've do things um you know every aspect of our lives is now kind of — you know digital okay maybe not every aspect but many aspects are most aspects are digital in some way. Um but you know

### Segment 11 (50:00 - 55:00) [50:00]

obviously there was it was a I think there's a period where you just overpromise and you underdel and then you kind of re recover. So I do think that yeah AI a bunch there's a lot of like real things there and things that are you know kind of bubbly and at some point that stuff will get cut out and then the real stuff will you know continue. — Makes sense. — The question is what are you studying undergrad? Um how do you decide to do it? Why does it lead to where you are now? — Um this is going to be a pretty boring answer. Um so I studied computer science in undergrad. Um I did also I guess it was uh I did a math major as well but it was really you had a lot of overlap between um you know CS because I was doing a lot of theory and math. Um and then I did a you know a PhD in computer science when I got into kind of um you know research. and now I'm a CS professor. So I guess that's kind of a kind of boring answer. I mean I think there were um maybe periods where I was you know looking at I was really kind of fascinated by physics at some point but didn't really um uh maybe I didn't really pursue it as as much because um I was kind of um I don't know it just felt very comfortable to um be able to build things in software. The question is about why are these companies all so you know non-transparent and what are the outlook on that around that? — Yeah. So um the amount of transparency over the last five years has really gone down. It used to be that openai would publish on what they're working on release models and now there's no information. Um and um I think that so to answer the question directly why does this happen? Um it happens for a few reasons. One is that um competitive advantage right um this is a usual reason companies have trade secrets. If I reveal how I did this training then my competitors are going to you know find out. The second is uh you know lawsuits. Um if I companies are I mean obviously training on a lot of internet data with some of it might be kind of dubious. Um, and if this is all kind of laid out, then that's just kind of inviting lawsuits. Um, and the third is a more mundane answer is that some things that companies would probably be fine with um, revealing, but there's just it's not a priority, right? It takes actually a fair amount of engineering effort to dig up certain types of information and put in a report and get it past approval. And all these companies are just kind of their first priority is to race um to build the best AI. So no one has time to do that which is unfortunate. Um so the idea behind the transparency index was that at least you call this out and you make it uh people aware of it and that's the first step to at least um stimulating some more uh transparency. Um and over the last two years there have been companies that have you moved quite a bit. Um and we've engaged with a number of companies and they've you know done their best but some areas like data are just there's no incentive. It has to be kind of regulatory to really get them to move. — The question is how do we get end users to care more about the ethical implications of AI? — Yeah that's a really great question. Um well I think the first point is maybe related to the transparency is that end users generally don't even maybe think about um where their AI how it's made and so on. So just being transparent about that allows them to actually point to something. Um I mean I guess there's an analog between how do you get um consumers to care about like ethically grown you know sustainable food and all these organic food and so on. It's kind of a very similar analog and so having new nutrition labels and like I think it takes adv advocacy on behalf of you know other organizations to push for certain things to be in place so that they can be um put in front of you know users um because users most of the time are just trying to get their you know work done or do whatever they and don't aren't

### Segment 12 (55:00 - 58:00) [55:00]

really thinking about like, oh, is this um ethical or not? — Or most many people don't even probably realize that how much of you know what are even the possible concerns like the fact that there's so much human labor that goes into them and wondering about like you know wages and and so on and condition work conditions to how much compute it takes and the strain on the environment. So just making people aware of that I think is a first step. — Question is you won some competitions when you're young. — Does that uh changes or influence things down the road? — Um I haven't won any piano competitions lately. Um I think um so for a so I've always been active in um in piano and I still play uh a bit these days. Um and for a long time it was always a parallel track and it was nice because there was a sort of a hobby aspect of piano. There was a few years ago I had you know two postocs who I really wanted to figure out how um to do combine music and AI. So we had a number of papers on essentially training music foundation models. Um which was uh you know pretty cool because then it for the first time it got me to really um kind of the two parts of my lives which are important kind of crossed. — Any other questions? Last one. Yes. — Um just I guess any general advice about getting involved in the research or Yeah. So, general advice on getting involved in research at Stanford. — Yeah. Um, so in terms of getting uh involved in research, there are definitely official channels. For example, there's a curious program that people can apply to. Um, and then there's unofficial um, you know, channels where, you know, the nice thing about research is that everything is open. So, you can see what research is happening. um who's working on what. Um and one recommendation is to just engage uh with that. You look around at people's websites, you find a paper that you like. You read it and then try to engage with either the professor or the student like you email Ken. It's like, "Oh, I read your paper. This is so, you know, wonderful. " And you know, I have these other ideas. and you can try to um try a more bottom up you know approach. Um the other thing is the final project uh in the class can sometimes be a way to um try to do some research which um can continue into something um you know bigger. Um there's no yeah onesizefits-all you know solution here. Yeah. — All right we're at time. Um thank you so much Percy. We can also read one more questions, but um I think we're out of time. — Okay. Yeah, we should probably stop there. — Yeah. — Um well, — thank you so much for taking the time to answer all these questions. — Great. Thanks, Ken, for moderating. Yeah. And thanks everyone for uh sticking with us through this class. It was a pleasure teaching and you know, also thanks to the teaching team here for um helping out. Um, hope you enjoy the class and um, yeah, enjoy the rest of your good luck with the rest of the quarter. Yeah. [clears throat]