Latest / Practice Ground / Sanmi Koyejo: Stanford Professor '50% of AI Is Data — 10% Is the Algorithm We Actually Teach' | Ep 9
Transcript
- Nifemi Aluko: ejut, thank you for joining me today. you know, just to give a background, you're assistant professor computer science at Stanford University. you know, you also lead the Stanford Trustworthy AI Research There, founder. Chief AI officer at Virtue AI. A lot, you know, there's so many things we could talk about. And I think you're also writing a book on machine learning from human preferences. So thank you. Thank you for joining today. Cool. Cool, cool, cool. Yeah, I mean, before we get into you know some of the deep topics, you know, like I just want to share a story of, you know, Sami and I were in Austin. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: That's great. excited, looking forward to it. Nifemi Aluko: We we spent more we spent more time together in a boxing gym in Austin, Texas. You remember that? Bruce, Bruce boxing. Yeah, that man. That gym, that gym really is really part of like my Austin story. Like it was literally like this garage. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: god, you're right. I do. Yeah, yeah, my. Yeah, it's been a very long time, I wish I kept that up. Nifemi Aluko: turned Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yep. Nifemi Aluko: into a gym that we all used to go to after work. You probably have to doing research on campus. Bruce, many hours in that gym, and my god. Yeah. And I remember to be honest, I remember like just moving to Austin from you know, I just finished my master's at Texas AM, moved to Austin, looking for things to do, joined this gym. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yes, Many, many, many hours there. This is a character, I think, a unique, unique person. He's incredible, but you know, it's just a fun character. Nifemi Aluko: I first went to like one of those fancy boxing gyms, you know, like and I was like, and then I came to this one and you know, no AC, grimy, like the the Bruce just yelling at people, right? And Yep. Exactly. Yeah, and and you know, I remember you in particular, like I think I saw you on this like speedback. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: It's like the textbook video movie version of a voxing meme. All the stereotypes are there. Nifemi Aluko: Right, boom, boom, boom, and you were and you you were going fast. I was like, damn, this guy can box, right? And and later on, I was like, wait, wait, did you go to ISL? And then just to go reverse a bit, we actually went to the same. We were in the same high school in Lagos, International School of Lagos. I was in your brother's class, Simi, Simi's class. so we go way back, man. Austin. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: It's so good. I remember it. Yes. way, Nifemi Aluko: You go. Yeah, yeah, yeah, yeah, yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: back. Your memory is better than mine. Now, once you say it, I'm like, oh, all these memories are flooding back very quickly. But yeah, I haven't talked to Bruce in a while. We used to be actually quite close, and I quite a little amount of time in that gym. it was really like escape for me while I was... Nifemi Aluko: Yeah. man. No, that that gym was like it was definitely an escape. It broke at some point, Bruce even gave me the the key. I'm sure you had the key also. You can go anytime. It's like come in anytime, just make sure you lock the door. Yeah. So okay, so tell me. So you you finished from Austin, Austin, you're doing your PhD out there. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yeah, I had a key or something. Yeah, and it was incredible. He really treated us like, you know. Yes. Nifemi Aluko: postdoc in Stanford. We just talked about it a bit. You did something, you kind of had different things between industry going into research or banish and pain, some Google in the mix, but now you're back in Stanford. nice, nice, nice. So, you know, I was you know, as was getting ready for this conversation, I was looking at some of the stuff you've already talked about. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: That's correct. Yeah. It has been about three years and changed now. Nifemi Aluko: and one thing that I spent some time listening to was this talk that you did about like AI measurements, right? Benchmarks and building out effective systems. And you know, even for myself, as I've gone through this process, being that we know that AI is something that is is based on probability, right? So you kind of just want to If you're ever building it, you want to kind of have like this effective way of making sure it's giving you good and better results. can you tell me a little bit more about like the importance of like benchmarks, measurements in AI? Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Mm-hmm. Yeah, AI is a quantitative field. It's mostly experimental. Most of the work, I mean, there are different ways to distill what the field has done and achieved and the guiding light of the work. If you talk to people, it's about building intelligence and matching or exceeding human intelligence. If you talk to some other people, it's about building tools and software that's useful for people that can automate different kinds of tasks. All these things are true to different extents and also differ across people. But I think the actual thing that the field has been good at is, you know, give me some measure metric of performance of something and I'll figure out how make that number get bigger. I mean, in some ways, a lot of the world is built around this. There's an adage in business and what you can't measure, you can't improve people. So I think measurement shows up everywhere as a big part of any kind of a Nifemi Aluko: Yep. Yep. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: I mean, least definitely automated systems, engineering systems, but often even human systems. think it really anchors strongly for any field that's highly experimental that also has meaningful resource investments. People care about the numbers. Measurements in the field used to play a very narrow role. used to be people were trying to figure out better deep learning architectures, better data sets. And so most of the work on measurement was Nifemi Aluko: Mm-hmm. Yeah. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Can we get some signal about whether we're making good choices, how to make things better? So very famously, there was a big revolution in the field when one of my colleagues, Fei-Fei Li, developed this data set called ImageNet that allowed for new ways of capturing complicated data distribution. So understanding all the different ways that we could look at computer vision data sets for this task of recognizing objects. so a lot of the work there was Nifemi Aluko: Mm-hmm. Yeah. Yeah. No. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: compiling these data sets and actually getting some meaningful signal about what good image recognition could look like. But really, it led to this revolution where people could then come together and build tools to do a much better job than anything that had come before for improving this number of how well are you doing image recognition. This is this benchmark thing, this thing that's on it. Nifemi Aluko: Got it. Got it. Yeah, yeah, yeah, yeah. And you know, in your work, right? I think one thing I've learned is a lot of people they get so carried away by, you know, AI and, you know, the top level and everything. But from what I've seen working with other people in the field, there's a lot of work that just goes into like, as you said, putting the data set together and the data engineering. How what percentage of your work is doing that? work that nobody really sees where you're just kind of gathering data versus the actual analytics and Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yeah. I mean, percentage is hard. Let me just finish a brief thought from the last thing. I think one of the things that has changed on measurement is that the world, AI has become much more important. It used to be again, this insular academic discipline. And over time, it's just become a bigger and bigger component of resource investment and policymaking and things like that. So these numbers that used to be really for internal consumption have become important numbers for people to make decisions in the field. Nifemi Aluko: Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: OpenAI or Google gets a new model. They trumpet benchmark numbers. They trumpet numbers in the arena. And I think in that talk and in a lot of my work, the gap that we're seeing is that the reason we were building measurement systems before and the kinds of work that we did don't actually answer the call for what measurement is required to do now, which is help us make decisions in a real world where we're interpreting the actual number. Nifemi Aluko: Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: those numbers were not built to be interpreted. They were built to just roughly tell us how we're doing better or worse. So there's a big gap there that a lot of my technical research work is trying to figure out how to. So your question on the data, huge, extremely important. Arguably, academia is very biased towards the algorithmic side of things. Some of this is changing, but in many ways, the sexy part of Nifemi Aluko: Figure out. Yeah, yeah, yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: the field is for a long time was I built some new algorithm. It shifted a little bit. think now it's a, maybe less I build an algorithm, but I can solve some interesting task that you want to consider before. But the core of modern machine learning and AI is data curation, way underappreciated for any project that's successful. Nifemi Aluko: Mm-hmm. Yeah. huh. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: In terms of resource investment and outcome, I mean, it's hard to put numbers, but easily over 80%. In fact, I'd argue, okay, again, in real time, you're putting me on the spot a little bit. The thing I love is the algorithmic, like that's the thing that, you I'm most excited about. confess as an academic, this is the thing that I'm most excited about. The things that have led to meaningful, significant outcome. Nifemi Aluko: Yeah. Yeah, yeah, yeah. It's all good. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: our combination of data and engineering, engineering in a sense of like, how do we scale up systems? So a lot of the success of Google OpenAI in these places is that, you know, it's one thing to have compute, it's another to figure out how to compute, can scale at world level, solve these really like incredibly complicated, incredibly resource intensive kinds of tasks. Nifemi Aluko: Mm-hmm. Yep. Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: part of the solution is just throw money at the problem. But you also need incredible engineers who are nitty-gritty on details that we often don't teach actually, because they feel like details. But if you talk to anyone who's built a real system, and any other work that we've done in building real systems, those details on the engineering side on how to get a place like the Big Roll. So if you put me in a corner to put a number, I'd say, Nifemi Aluko: Mm-hmm. Yeah. No sé? Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Despite what I love, the actual thing is maybe 50 % data, 40 % engineering, 10 % the stuff that we actually teach. And it's not that it's not important. We have the rough rules of the world on how to make progress and what has actually led to meaningful advancement. But lot of the nitty gritty has mattered a lot in engineering and data components. Nifemi Aluko: Yes. Yeah, yeah, yeah. Yeah. Yeah. Yeah, yeah, yeah. Okay. And so, you know, when you when you think about like debugging a complex system, which is what you literally do day to day, right? What does your intuition tell you that you can't necessarily see in the data? Like how much of your intuition do you bring into your work? Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yeah. So I think one of the interesting things, I'll flip the question a little bit because actually, because of, mean, I'm going to ascribe intention to mechanism where I'm not completely sure. The thing that I observed though, is that people actually don't look at the data nearly enough and actually rely a little bit too much on intuition to debug systems. Nifemi Aluko: Mm-hmm. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Again, this is maybe a good idea for many things, but for systems that are primarily engineering empirical systems, which is again, most of what the AI world actually is, it really does matter that you take a meaningful look. We trumpet algorithmic automation, but really it's people behind these things. It's people that are part of the data labeling and data curation. It's people that are part of the engineering building. Nifemi Aluko: Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: and carefully looking at data artifacts to make sure that things make sense. There's an incredible amount of humanness behind advances. I think the story, the narrative is not nearly as cool when you say it's people, human decision making. But the automation narrative is much more compelling and I think is what we hear the most. you really do need, if you go to any successful system that's out there, Nifemi Aluko: Mm-hmm. Mm-hmm. Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: things break, we don't yet have good completely automated tools to know what might be broken. To give you a simple, like one example. Whenever you're training any of these really large neural network systems, the thing you're typically tracking is some notion of error. So like how often the models are making mistakes for the classic task of next word prediction. as some of the audience might know, or as you might know. Nifemi Aluko: Mm-hmm. Any more? Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: a lot of the core of large, successful, modern AI systems, this thing's called transformers, is this task of trying to predict the next word in a sequence of words. That's sort of the key underlying technology that makes things work. So yes, the token is the real thing. are easier to talk about. But if you build a real system, what you'll see is that, you Nifemi Aluko: Uh-huh. Next word, yeah. Autokium. Yep. Mm-hmm. Yep. Yeah, no, it's easier to say the word. Yeah, exactly. Yeah, yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: As you apply the optimization tools that we have, the error gets better, better, better over time. And then suddenly it'll just spike. And there's no explanation for it. And the only way to actually figure out what's going on. So this is actually very common. I'll even, so much of building, the reason why people pay so much for top engineers is actually you need to babysit these systems quite a bit. They're appreciated. Nifemi Aluko: Yes, sir. Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: But it's not, know, turnkey, set something, walk away. You're spending a lot of time looking at the system behavior over time. You're like tracking logs, you're tracking, sort of, know, outcomes, outputs, all that stuff. If you don't do that, your system will be crap and you don't yet know a way to avoid you doing this right now. So much of the process is human. So your point about intuition. Yes. Intuition plays a huge role. Nifemi Aluko: Uh-huh. Mm-hmm. Yeah. Uh-huh. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Often for success, the way intuition comes in is actually a lot in this mix of intuition and data analysis. So again, we haven't found ways to fully automate this process of building complex systems. So actually looking at logs, looking at data, forming intuition for why there might be mistakes and errors. So one example that people have, know, over time, we talk to engineers, something that happens, one of the reasons why the error might spike is that you might be looking at data from Reddit. Nifemi Aluko: Yeah. Like blue. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: and then someone put up, I don't know, like a garbage... blog post that's all S's or all A's, information content free spam, essentially. Maybe they put it up as a joke or a placeholder, but if you're just grabbing data from the internet, you missed this and this ended up in your data set. Or it's a list of emails, something that's very far from conversational text. And the system has to try to figure out how to train and predict the next word on something that's garbage. And it just throws the system out of whack. Nifemi Aluko: Mm-hmm. Mm-hmm. Yep. Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: And if you're not looking at your data, all you see is a system, the training process is broken. You have no idea what happened. Over time, you get experience in looking at logs to get intuition for like some of the common failure modes and what might break. And so a lot of debugging in modern ML AI ends up looking a lot like, you know, data debugging things. When you build intuition is really like you find common patterns. Nifemi Aluko: Mm-hmm. Got it. Yeah. Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: over cycles of things that are happening. I think a lot of my work, what we're hoping to do is maybe come up with automated evaluation systems that can, you know, again, I think it's okay. And for a while that these things are much more manual than advertised. I think we'd like to democratize what it might look like to build world class systems. And I think that path there includes some tooling Nifemi Aluko: Mm-hmm. Yeah. Mm-hmm. Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: for measurement of out of behavior things. Nifemi Aluko: Tooling, yeah. Mm-hmm. And you know, it's interesting, you said I was just I had a conversation, similar conversation with my professor from Texas AM, right? And he's at RPI now and doing data science particularly for systems, bio biological systems, right? And I I gave him this example, like when I used to work at Rockwell, where we used to implement like neural networks. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Okay. Nifemi Aluko: At factories, right? We we had this thing that when everything is at steady state, everything is all good, right? Everything is going well, you know, everybody's relaxing until some random stuff happens, and then you see the spike, and everybody's panicking, and the the model is now just going crazy. At this point, we have to turn it off, and then that experienced operator just runs out and says, you know. This valve was open and we need to change it every other three months. So they call maintenance and maintenance comes in and but it's the experience of that person that kind of jumps in and just kind of mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yes, exactly this. Yes. It matters so much. you know, again, we tell the story about automation, but experience plays such a big role in complex systems, because we haven't yet figured out a better way to build in these kinds of intuitions. Of course you could, you know, in your example, you could figure out and nail the issue, but that's six months of work or something. We're testing every single thing, figure out what's going on. Nifemi Aluko: Yeah. Mm-hmm. Yeah, yeah. Yep. Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: way if you have someone who has built experience over time and they have the right, they get it right away or it can point you in the right direction. So yeah, this is an important point. Nifemi Aluko: Mm-hmm. Yeah. Yeah. Do you think yeah? Do you think there's there's always going to be a world where we still have to babysit the automated systems? Do you think 50 years from now, 60, 70, 80, regardless of what we build, there's still going to be some form of babysitting, even if it's just a Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yeah. Nifemi Aluko: I'm just sitting here and making sure every if it's all even if it's one person, maybe sitting the entire world said there's that one guy that's like, hey, I just want to make sure it doesn't go crazy. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yeah. Right, right. Yeah, you're asking a hard question. This this long term prediction thing. I mean, for sure, I think a lot of folks in the field would like the world where we don't have to. I'm yet to see a complex system where you don't need babysitting. I'm always worried about saying never about anything because the world is what it is. But just for an example, the answer to your question, I don't know. Nifemi Aluko: Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: other than, don't know of an example of a complex system that doesn't need people for edge cases. I'm thinking of the canonical examples. I'm like airplane flying airplanes, mostly fly themselves now. but you still need pilots because there's going to be edge cases of something that goes wrong or like some parts of the system where something about the human way of thinking and ability to problem solve and experience to know failure modes. Nifemi Aluko: Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: is key to finding solutions in reasonable amounts of time. So is it our failure of innovation to come up with some automated approach that gets around this maybe? Or is it just a fundamental component of complex systems that once the system gets too complicated, are automated tools generally insufficient? I don't know where to put the bets. It's a good question. Nifemi Aluko: Yeah. Yeah. Yeah. Yeah, yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: But I'm yet to see a counter example for humans in the loop, even modern, say, AI systems. Nifemi Aluko: Mm-hmm. Yeah. Mm-hmm. And you know, that makes me think about this next question where I feel a lot of parts, like those edge cases or whatever it is, sometimes it's experience, sometimes it's just tinkering, you know, people just playing around and just realize this whole thing. There's some play involved sometimes, even while you're going through this scientific approach of trying to solve a problem. So what role for you? Does getting lost in play, right, in your research, how does that impact the way you think about a project? Do you kind of say, hey, let's ask more questions, let's investigate, or do you kind of go more into that mode of like Mm, just a little investigation and then let's just kind of we'll found something, let's not will I say exploit it, but let's move forward. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yeah. Yeah, I think that's a reasonable, explore exploits is a reasonable analogy for how to think about this, at least to my thinking. So I'll start off by saying I think research is, I think, a quite privileged place in the world where the incentive is so well aligned with exploration. Nifemi Aluko: Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: I think it's harder in business because I get to see both sides now where I say, you know, for the research side, a lot of the long-term useful outcome comes from being able to take risks that are not obvious, that require you to do a bunch more exploration. You know, if something is as straightforward as pure iteration and grinding, it can still be useful research, but often companies are better at executing at that than most. Like they have more resources. They can just... Nifemi Aluko: Yeah. Yeah. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: throw at the problem and scale that more efficiently. And in fact, a lot of work that we've been doing around, say even a lot of startup work is because we see a gap where there's some insights and the thing to do is execute on that gap. And execution is actually much, more better done in a company environments. Things are one those reasons why companies also have a love-hate relationship with their research teams. They like the innovation, but it's not always well aligned with company outcomes. Nifemi Aluko: Uh-huh. Mm-hmm, mm-hmm. Yep. Yep. What quarterly goals, yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yes. So research is really, I mean, if that's something that one values, and not everyone values the same way, but I do. And for someone that values this ability to think broadly and just spend time exploring, think research is this magical, this position that I and some small fraction of the world has that this is actually the place where you get, Nifemi Aluko: Mm-hmm. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: most of the bent or at least the career positive signal comes from helping our problems that might take a long time to solve, that requires you to explore. And so I say this a lot to a bunch of my PhD students. mean, in terms of making money, in terms of like most of the obvious signals for career success, Nifemi Aluko: Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: you're probably better off actually not doing a PhD and not being a researcher. I think you should do research if you value the exploration process, because I think that's you do a lot of that and you do a lot of that in a way where you won't get, there'll be long periods where you don't get signal that you're going in a good direction. And you to be resilient to the fact that there's no positive feedback for long Nifemi Aluko: Uh-huh. Yeah. Yep. Yep. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: You can call it resilience, can call it enjoying that part of the exploration. Again, it's not for everybody. But if that's not something that you enjoy, I think you'd be frustrated and reason. Nifemi Aluko: Okay. Yeah. So like for you, how has it been? Like now you're you know you're you have that research background, right? Now you're also doing virtual AI, right? And I'm of the opinion that researchers can be great at build building businesses, particularly for myself as an entrepreneur. Like I've taken this approach of like, okay, you're gonna do stuff, it's gonna fail. You're gonna learn as long as you have that learning mindset of like what else can we do and reiterate, right? So there's a part of me that thinks like, man, there should be more researchers building businesses, right? But that doesn't pan out the way it is, right? When you look at the numbers sometimes. What do you what have you brought from the research field that you see is very aligned as you're building your business? Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yeah. Should there be more researchers building companies? all need learn that a little bit. mean, high level, yes. do think for, there's a very like obvious failure mode for researchers that choose to build companies that one has to be aware of, which is you don't. Research rewards so much the exploration. So we talked earlier about this explorer thing. I think it's a mistake to throw away exploration if you switch to building a business. But it's also a mistake to not change the fraction of efforts on exploration versus execution. So if you don't like exploit, execution is a good alternative word. And I think a failure mode I've seen some of my colleagues fall into. Nifemi Aluko: Exactly. Yep. Yep. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: is they forget that the business has different incentives and outcomes tied to it. And you're running it like you're running your lab and you're doing pure research, which can be very valuable. And in fact, just to really anchor on this, there are many multi-billion dollar startups that are purely researched in AI. So this is a very unique thing about the AI ecosystem. But if the thing you're building is a standard business in the sense that you need customers and you need to make money and all that stuff, then I think there's value Nifemi Aluko: Mm-hmm. Record, yep. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Researchers can be very good at this because they have the research exploration mindset. But my early sense, some of this is my experience, some of this is talking to mentors and folks who have done a bunch of this research slash industry in the past, is that the thing you have to do though is find a different balance in a company execution setting, assuming what you're doing is a standard business. Nifemi Aluko: Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: where you have to find, you are open to exploration and learning, but you care about good execution, making customers happy, which is not something that a researcher is doing. The customers are your community and they care about your work, but there's no execution in that same way. Software artifacts stop at maybe the version that makes sense in the paper. So it doesn't need to scale, doesn't need to productionize. The UI parts of it can be sloppy because... Nifemi Aluko: Mm-hmm. The research Mm-hmm. Mm-hmm, mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Often the win is he has a good idea. Not necessarily here's how it serves a millionaire billion people in a different place. I would say I also feel extremely fortunate. There are a few universities that are really good about encouraging entrepreneurship. Either explicitly or just like it's just such an ether that it's not unusual. And I think Stanford is one of those places where, you know. Nifemi Aluko: People, yeah, yeah, yeah. Mm-hmm. For sure. Yeah, yeah. I mean, that's why you know when I when I was applying to business school, I like Stanford was number one. Per particularly because there was all just the brand like of entrepreneurship, right? And and there's not an ad for for Stanford by the way. But yes, I do agree that there are just certain places where either b a lot of yeah is in the air. Exactly. A lot of bills businesses have been built. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: It's in the air. everywhere. Students, faculty, how's that? Nifemi Aluko: Exactly. Alumni, all that kind of stuff. so some can you tell me more about trustworthy AI? What does that mean? Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yeah, no, trust is a tricky, I mean, it's one of those things where you're trying to operationalize something about language and really something that's a feeling, is like, so, so trust is, I find is valuable to separate. So trust is really human opinion and trust in technology broadly is about human opinion about technology. The place where this matters a lot is trying to build systems where humans have calibrated trust. So what I mean by this is there's a tendency to either over trust where you imagine a system and solve many more problems that engage in many more things than it can actually do. And so you apply it in places where it's not ready and you have some disaster happen or under trust where you just imagine. So over trusting. I know we talked a bit about AI being very happy right now. Nifemi Aluko: No? Yep. Yep. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: The consequences when people just assume they can solve every problem. So there's a bit of over trust issue right now, I think, in AI systems. The under trust version, example maybe. this is really like, assuming system cannot solve many other problems that it can actually solve. Many examples of this, maybe some parts of self-driving, though that one's tricky depending on the specifics. I like to use the example of some of my doctor colleagues. Nifemi Aluko: Yeah. Yep. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: I work quite a bit in AI for healthcare. My experience is that my clinician colleagues are extremely on the, like they're very much on the extreme. So there'll be a couple of people, some fraction of the community that believes AI will solve everything about medicine and immediately is like, you know, bought into everything that you say or anyone says. And then there's some fraction that believes that there is nothing useful about technology. Nifemi Aluko: Uh-huh. Mm-hmm. About the system, yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: you know, we'll jump on every failure and there will be some failures. And so, you know, that becomes the whole narrative. I mean, the truth is in the middle, not in a generic, in the middle sense, but really, I think the goal is trying to find this right calibration. And in the ideal world, the way that society engages technology will be in this calibrated middle where we understand where technology, where it can work, where it doesn't work. Nifemi Aluko: Mm-hmm. Hello? Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: where it's useful and recalibrate accordingly and where we apply it. So my group tries to engage in this question. The way that we do this work is we pick a few different aspects that we think will engender trust by people. And we try to build research conceptual applied foundations and deployments that can help close this gap. Nifemi Aluko: Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: So these directions look like robustness. What is happening when something changes in the world? How does the system react? This looks like fairness issues. How does the system behave across different groups, different graphics? Does it work the same for everybody? This looks like privacy issues. Is this data? Are people able to get information from it? This looks like security issues. What is, what is the threat of your system from a bad actor third party? Nifemi Aluko: Mm-hmm. Yeah. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: So these kinds of, these are very specific directions and this broad notion of what it will take for society to trust AI systems. And what we like to do, the kinds of work research likes to do is thinking through building both understanding and tooling to help close the gap across these directions with the hope that this engenders more calibrated trust in technology by society, by people. Nifemi Aluko: Yeah. Trust, yeah. Mm-hmm. And i is the outcome just moving society forward? Because you know, I think in general, like trust is amongst humans, trust is like this huge currency, right? Like you trust your family members and then mean you trust your neighborhood, and then you trust the bigger community, or if you're in the same tribe, all that kind of stuff. and Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Mm-hmm. Mm-hmm. Nifemi Aluko: I feel like the technology we've been using for that is something like storytelling. We tell all these stories, you know, whether you trust that guy or that guy, you know. yeah. So is the outcome here that if we build this trustworthy AI, human trusts the intelligent system, then we can use it better. Is that the outcome? Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Strongly agreed. Yeah. so the, you you mentioned, and I think the way we're discussing these things is there are two parts of the story, which is the human part, human society, that whole bit, which includes things like governance and policymaking, right? So part of trust is people use to some extent, trust in institutions that decide for them. can walk into any grocery store now, in most places in the world and buy food off the shelf and have some reasonable confidence that you're not going to die by eating that food. We don't think about it now because it's ubiquitous, but this is a consequence of regulation that has ensured food safety to where it's now invisible. Because there's a whole process that just made measuring potential for risk in food and processes across the industry to make sure that food systems are safer. Nifemi Aluko: Mm-hmm. Yep. Yep. Mm-hmm. T. Yep. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: So similar thing with technology systems in general. So does the human part, this is the technology part. We don't have agency necessarily, or at least the mandate of my group is not to try to solve the human part. What we are hoping to do is to make progress on the technology part so that it's easier to, for us maybe as partners, but really like for our society to solve the technology part. So we wanna build tooling. Nifemi Aluko: Mm-hmm. Got it. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: into AI systems so that it's easier for society to trust them. What are you looking for? You're worried about mistakes that systems make. So, you know, the easiest way to lose trust of a technological system is that you expect it to solve something. You have some prediction of it can solve this task and then it does something wrong. Immediately you lose trust and that trust can be hard to get back. In policy making, governance cares about risks. They care about negative outcomes. They care about Nifemi Aluko: Mm-hmm. Mm-hmm. Something else. Yeah. Mm-hmm. Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: different behavior across different demographics of this fairness issues. And so again, ways to lose trust there is that you build systems that discriminate in various ways. And so the way that technology can engage in this is that we can operationalize what it looks like to build systems that don't have that issue. can both understand the risk of that kind of problem. We can build into technological tools, mitigations for those kinds of issues. Nifemi Aluko: Mm-hmm. Yeah. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: and we research the outcomes on society. And so this is a way to close some of the technological gap to engender trust. But again, it's a big social technical system. The place where we have agencies on, or at least where my group focuses, there are other people that do some other parts of this, but my group focuses on what can we do on the technological side to make systems better so that they're worthy of societal trust, society's trust. And we try to provide tools for Nifemi Aluko: Okay. Yeah, yeah, yeah, yeah. Huge huge system. Yeah. Mm-hmm. Yeah. So that's it. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Technologists have done a bunch of work trying to build tools that end up being deployed in clinical systems. And we talked about companies and industrial systems. We provide tools for policymakers to engage and understand the systems that are being deployed in the world, provide tools for folks in ethics and civil society broadly to understand what's going on in these systems. So our goal is to close the technical gap so that Nifemi Aluko: Mm-hmm. Hmm. Yeah. And yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: people can engage. Nifemi Aluko: And for you you know, we've talked about like your journey, right? You know, from Austin to, you know, research to industry and now you know collaborating with industry. How do you how how did you get a sense of like, okay, this is what I want to do? Like or these are the things I want to do. Like where did you find when did things start become when did they start becoming meaningful to Okay, I'm gonna keep working on this and Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yeah, it's a great question. It's a mix of things. A few directions. One, it's meaningful. I think I was fortunate to have good foundations, at least for the kinds of work I'm doing on the technical side. I think, you you can only, you can only sort of prepare yourself so much. There's so much of outcomes that depend on serendipity, luck, whatever you want to call it, that no matter how Nifemi Aluko: huh Yep. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: prepared you are and how you set up yourself. Things happen and it can go either way really. And I think I'm, I feel grateful that many of these things, at least the outcomes have been such that they've my way. you know, along the way I always leave space for unexpected conversations, unexpected ideas. Lots of my success have been tied to Nifemi Aluko: Mm-hmm. That's all. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: People I know who are advocating for me in front of me, but also people I know from the outcomes that people are advocating behind the scenes. Examples of this, you win academic prizes every once in while. No one wins an academic prize if there isn't someone behind the scenes that wins academic And these things play such a big role in career success because they're also the world that someone thinks you're doing good work, they should be valued. No one gets a job without an advocate somewhere. Nifemi Aluko: Uh-huh. yeah, vouching and mm hmm. Yeah. Yep. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: good, right? I mean, at different levels of depending on the import or at least societal perspective on how important the job is, but no one's becoming faculty of Stanford without significant. I've seen how the been on hiring committees in various. You need strong advocates that really believe in what you're So I say all this to say, I guess I want to be very aware of how much of this is other people. Nifemi Aluko: Yep. Mm. Yeah. Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: I mean, to that point, you can do a lot to put yourself in as good a position as possible. A lot of this is preparation. Others is sometimes serendipity of being open to some good ideas. So specifically in this direction and some of my work. So I started off my, when I started my sort of technical career, I thought I would be doing, I've actually shifted a fair amount, mostly in engineering, but all of my degrees are electrical engineering. I thought I'd be doing control systems because that was Nifemi Aluko: Mm-hmm. Yeah. Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: where I got initial research exposure in undergrad. Nifemi Aluko: Mm-hmm. We would have been doing like similar things that I did a wrong one. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Yeah, that's true. And then I took a couple of classes in communication systems in my first year at University of Texas. And I got really excited about, I think I just had just good instructors as well. They got me excited about the area. So I ended up thinking I would do communication systems. I'm going to tell the story, but you know, to give example of all these switches that I think would have been hard to predict from the outset, but you know, I think I'm quite happy where they've landed. Nifemi Aluko: Well. You heard it? Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Um, in the middle in communication systems, I got excited about what used to be called cognitive radios. It's still a thing is this idea that we can build intelligence into the way communication systems work. Uh, in particular, as you might know, you're running out of spectrum radio. And so, because, you know, so many things use wifi and over the air communication. And at some point, the FCC had this. initiative to allow some shared bandwidth. And so you need intelligence to allow you to switch automatically between different bands so that you're not interfering with other people. And so it was wireless com, but also a bit of AI systems, like some predictive systems, some intelligence switching, things like that. And weirdly, I got much more excited about the intelligence part than the radio part. And so then I started to research on machine learning and AI systems and Nifemi Aluko: Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Many things happen by end of switching advisors and mostly focusing on that part of the work. was doing the thing in vogue in the field at the time was various kinds of probabilistic machine learning. So spending a lot of time thinking about uncertainty and modeling uncertainty and things like that. So technical that kind of stuff. I walked into a random talk at some point late in my close to the end of my PhD where a lot of it really was random. No, so Nifemi Aluko: Yeah. I love the random random talks. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: And machine learning, you know, half a generation ago. So right now, all of the software tools, the most popular software tool are variations on Python tools. PyTorch is the main thing. Before PyTorch, which is for deep learning kinds of things, the main tool that people use was something called scikit-learn. So this is a, actually, it's actually the place that we, I still send most people to start their machine learning journey. It does pretty much everything but the deep learning stuff. Nifemi Aluko: Mm-hmm. Mm, mm-hmm. Yep. Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: we mostly do in this. It used to be called SciLearn. I forget, it had a different name, but at some point it became called Scikit-learn, which is this whole package that does a bunch of the basic machine learning, so predictive modeling, k-means, bunch of it's like nuts and bolts of the software tools for machine learning systems. And the lead researcher, the lead of the Scikit-learn happened to be visiting Texas. It turned out that, so I went to go see his talk. Nifemi Aluko: Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: because I worked in machine learning, lead of this tool that so much impact is around. It weirdly turns out that the people who build scikit-learn are mostly neuroscientists. It's weird because they're machine learning as well. neuroscience and machine learning used to be so much more coupled. In fact, the conference Europe, the neuro part is neuroscience. Exactly. So it used to be much more coupled. It's not that surprising in retrospect, but Nifemi Aluko: Yeah. Yep, no one doing that just happened. Yep. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: It does seem weird, I think, on the outset. And so I went to this talk, and I stayed later to chat with a guy and ended up chatting with this neuroscience professor and got really fascinated about machine learning in neuroscience because they were talking about these open problems. that ended up becoming this professor end up being on my PhD committee. That became my postdoc advice. So I went with Stanford. That's how I got into sort of this area. Nifemi Aluko: Yeah, yep. Interesting. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: human AI things. Nifemi Aluko: Yeah, yeah, yeah, yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: And it became the core part of my research agenda for a few years. I went to Illinois, started my research career, you know, students, random conversation with a colleague, we said, Oh, let's get the students, the student turned out to be excellent in helping us think about things like distributed systems and a bunch of that became a lot of what got me tenure at Illinois at the time was it's sort of, you Nifemi Aluko: Mm-hmm, mm-hmm. Interesting. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: opening my Nifemi Aluko: Yeah. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: ability, this, I don't know, willingness to try some of these ideas. I got into healthcare. A lot of this happened because people just randomly pinged. Because I did a lot of neuroscience and if you're far enough away from, from the computer side and you're looking for someone who knows learning machine learning things, it's interested in humans. Neuroscience and healthcare look the same from the outside. ended up getting random pings. Nifemi Aluko: Mm-hmm. Yep. Mm-hmm. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: And at some point, I got this really fascinating reach out from a doctor. There's an article somewhere about this, actually. They called it Shut in the Dark. It really was. The doctor randomly emailed. We were interested in something about chests, x-rays, and some radiology questions. I'm like, this sounds cool. We end up collaborating. There's a patent with my name on it now that comes from this random thing. There's a of these. mean, again, there were foundations, but. a huge impact from events that I could not have planned. Nifemi Aluko: Yeah, no pr not predicted. There's no model prediction for this one. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Oh, the company again, you know, random colleague reach out. We're working on this problem. We think you could be helpful. Like, okay, let me learn more about it. We ended up writing this paper that was probing trust and safety questions in AI systems. It ended up being the core of the company. The company itself, you know, pivoted. We thought we'd be doing more like measurement and evaluation thing. And it turned out the market was much more on guardrails and mitigation. Nifemi Aluko: Mm-hmm. Address, Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: to protect systems as opposed to finding Nifemi Aluko: yeah. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: the gaps. so, you know, at least for me, the serendipity was this colleague reaching out and working on the sort of the original paper. Paper got some really good reception, won a NeurIPS, this conference I talked about, the best paper award there. We got a best paper. I learned from the first time that the National Security Agency has a paper award. I had no idea that they had this. They have a security paper award. Nifemi Aluko: Okay. huh. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: I learned about this by getting an email that we won the cybersecurity paper award. So they read this paper and thought it was exciting. And it was exciting enough for them. like, hey, this is... So we were the first to try to operationalize what trust and safety could look like for modern LLMs. And at the time, was a huge gap. People were worried, but didn't know how to measure, how to think about this. so it became a big deal. And it's on hugging face. Nifemi Aluko: Hello, okay. Thanks. Mm-hmm. Mm-hmm. Yeah. Yeah, yeah, yeah, yeah. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: thousands of downloads, people are using it. So it became a meaningful resource for the community. But again, what I have predicted that this is the thing I'd be doing, I'm not sure. Nifemi Aluko: nice. Yeah. Yeah, yeah, yeah. I hear you. I hear you. Hey, you know, y you're telling this story and it just reminds me of how I ended up in Austin. Like I I just finished my master's at Texas AM and my professor was like, go present this thing that you did at this process safety conference. And it was a Hilton in College Station, the most random Hilton ever. And I was giving a presentation and I Practice so hard for this presentation, right? Two slides in the fire alarm goes off. Right? And then so I'm like, man. And you know, like I was nervous, but as soon as I the fire alarm goes off, like I literally like slump my head and everybody starts laughing, right? Everybody starts laughing. We kind of walk outside and I'm like standing in the corner, like practicing, going over my slides in my head. And this guy just comes up to me and is like, I saw your two slides. That's exactly what we do at Rockwell. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Wow. Yeah, because he's the first to do that. Nifemi Aluko: Well, it's a company called Pavilion. You know, we we have this neural software, da da da da da da. It was like, let's talk after the presentation. Dude, that's how I got my first job. Yeah. The most random it was a fire alarm. Yeah. Yep. Exactly. Yep. Exactly. Yep. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Wow. ads. Wow. Yeah, from the outside, Like, I mean, but you had to do the preparation. You had to, you know, create it done the work. You had to have prepared all this time for the presentation. Yeah. Nifemi Aluko: right, two more questions, right? All right, so as an educator, right, when a student's now can get answers instantly from Gen AI, what becomes the job of the educator? Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: You picked some of the hardest questions for these five minutes. And now this is a huge one. I think we've been thinking about this quite a bit. I Stanford, any institution that thinks of itself as training people thinks a lot about this. I think even the best-meaning students, the best students actually, usually the best students because they take on, I mean, okay. Maybe not because they take on so much, but one correlate of a good student is they tend to do too much stuff. And so even if in the ideal world, they would be very thoughtful and very careful about exactly what they work on. know, 11 PM at night, homework is due tomorrow. have 10 other things I'm looking to do. I can push this button on this AI system and it will just solve. It will do something for me that looks plausible. It's just very hard to avoid that temptation and just do it that way. Nifemi Aluko: Yeah. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: In the short term, what has happened, think assessment has changed. So yeah, I your question is different, which is like, what is the role? But I can say very briefly, like, I think the things that are actually happening in education right now in the short term is... Nifemi Aluko: Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: So we think of our role as training students to be useful in the world, either in companies or in research, so innovation, and as a high level. so particularly at a place like Stanford, in addition, we literally think of our students as changing the world. And I don't mean this in any, like we just actually literally think that our students can change the world. And so we're thinking about what do we train them for for this? What do we teach them for this to be the output? Nifemi Aluko: Mm-hmm. Yep. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Um, so, um, I think the dynamics, yeah, it's a tricky space. think the dynamics will go in, um, it's unclear. Like how the role of, uh, education might shift. would say though, it's never been purely content. Um, you know, if you go, you talked a bit about why you went to Stanford earlier. Like, you know, the actual thing you learned in a classroom, you could have learned at 50 other. Nifemi Aluko: Yep. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: place, hundreds of other places, you learn it online. You don't need to be in a Stanford. The thing you get from Stanford is this community and the social learning and the network and a bunch of other extraneous things. doubt, mean, my sense is that that part of the educational process is not, you know, it'll take longer for that to, and I think probably be a loss in that loss. There's another conversation to have about, you this tends to end up being really concentrated. Nifemi Aluko: Yep. Mm-hmm. Mm-hmm. Yeah. Yeah. Thank you. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: at quote unquote elite school. So not everyone has the opportunity when they go to different universities to get the thing about education that's not the literal in class learning. But so much of the experience, in fact, even before AI, the argument is always, the argument that I found most compelling is that a lot of education really is the social process. it's work with your colleagues, learn with your colleagues, figure out how other people think, figure out how to explain things to other people. Nifemi Aluko: Mm-hmm. Yeah. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: a bunch of that thing that's not really captured and what goes on in the classroom is often in some of the most critical part of education process. I don't think that part will change. I do think there's a lot of worry that the execution part of it, so the pain and suffering of not knowing something and having to work hard to learn that thing, might go away. I worry that matters and both getting the muscle to do hard things and fail at them. I think that's important. But also, Nifemi Aluko: Yeah, sure. Yeah, yeah. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Even if AI automates, we talked at the beginning about complex systems and experience. Even AI automates a bunch of what it means to do valuable work. If you don't know how things work, when something breaks, you just won't know what to do. maybe we can automate our, we talked about earlier, know, do I think there's a world where humans are not in the loop somewhere? Maybe I'm yet to see an example. we're training students and we're not training in such a way that they're building this muscle, I think. Nifemi Aluko: Mm-hmm. Yep. Yep. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: I think it's a net loss for the world. So I don't have a good prediction, but I think that's still an important role. And it's always been more than the literal classroom. For sure, when MOOCs became a thing, you always get the content online. Nifemi Aluko: Yeah. Yeah. Mm-hmm. Exactly. Exactly. All right. I'm gonna get you off on this last very you can make it very brief, right? So I've written a few science fiction books, right? there's a playful one. If the entire world went dystopian, right? Dystopian, we you know, the machines are taking control, you know, we're we're we're we're slaves to the machines, or utopian, you know, they're doing our work, trustworthy AI, friendly AI, they're you know. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Hmm Nifemi Aluko: taking care of all our needs and we're just like, what else is there to do? Right? There are these two worlds. You could choose either world that you want to live in. What would you do on a daily basis in either of the two worlds? So choose a world and then tell me what you do with your time. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: Wow. I so I'll choose Utopia because I think that seems better. I think at least to the extent, mean, for whatever it's worth, so much of our thinking is shaped by movies and narratives and kudos to authors for helping us visualize what both of these things could look like. But I think I'll take the lazy route tonight. I think I prefer to live in a Utopia versus. Nifemi Aluko: Okay. Yep. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: I make no predictions about where the world will go, by the way. I don't think we know, and I think it's too complex to make easy predictions about. That said, I hope that if I was in no utopia, this would free up my intellectual time to do parts of what I love doing, which is, again, I think being a researcher, the parts of my time that a researcher, it's such a lucky thing to have as part of my job that I'm expected to think broadly and think creatively and, you Nifemi Aluko: Mm-hmm. Yeah. Mm-hmm. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: explore as a way, this is the thing I'm paid for. And I can help other people and train other people to do this as a way of life. I think it's one of those things that may be proof, like, what's called future proof in a utopia anyway. Actually, I still in a dystopia. Maybe research is no longer that, know, open-ended thinking is not that interesting. But in a utopia, I think it will be a good fit. I, yeah, I think. Nifemi Aluko: Yep. Mm-hmm. Mm-hmm. Yeah, my dystopian looking, you're trying to find food. Yeah. Yeah. Nice, nice, nice. So Miko, thank you very much for your time. I really appreciate you. It was very nice talking to you and hope you have a good day. Enjoy the rest of your day. Sanmi Koyejo Sanmi Koyejo webcam 00h 00m 00s 272ms StreamYard: That's what I'd hope to do. It's pleasure. Thank you.