AI+Education Summit: Generative AI for Education

all right good morning everybody my name is Rob Rish I'm a faculty member here at the University uh I teach political philosophy in the political science department I have an appointment here at the institute for human-centered AI where we're sitting and perhaps a completely appropriate for uh the the day's events uh my own degree PhD is from The Graduate School of Education I was a sixth grade teacher before I went to graduate school so the topic of AI and education is something I'm keenly interested in I'm going to introduce the panel in just a second but I want to share something following Dean Schwartz about my own experience at The Graduate School of Education now 25 years ago apropos the conversation today so number one anyone who thinks about various questions about American Education knows immediately that it is an unbelievably fragmented or decentralized or localized system which makes the adoption of any new innovation across the entire system incredibly fraught um an historian of Education I took courses with 25 years ago at The Graduate School of Education named Larry Cuban um said frequently in classes that the last Innovation adopted at scale and that's persisted over the course of decades in American education is the Blackboard and shock and the enormous promise of artificial intelligence to revolutionize enhance education accelerate learning teaching and learning um offers us great promise also some Peril which we'll talk about today and I'm really curious to try to learn more about the various ways in which not merely we can find demonstration projects that are incredibly important but also ways to navigate the fragmented policy and governance system of American Education including of course higher education at the same time so it with that as a brief introduction to the opening panel I want to introduce uh our our speakers um uh first is Percy the Yang a colleague of mine here at the institute for human centered AI he directs the center for research on Foundation models and a professor in the computer science department um Noah Goodman who is a professor at Stanford as well in the department of psychology and in the department of computer science and Dora demsky I'm I'm a professor in The Graduate School of Education we're going to do short presentations from from each and then a moderated discussion with some questions from the audience as well thanks for joining us and um Percy why don't you kick us off if you're willing okay I thought I was going third but I guess we can reorder um so I'm going to try to do this without slides because I've been inspired by philosophers such as Rob who are so effective and I'm sure you'll see a lot of slides but we'll see how it goes um so as Rob mentioned I'm the director of the center for research on Foundation models and I used to have to explain what a foundation model is and now I could just say chat GPT and I think everyone understands what I mean by that um so the center started two years ago and when we realized there was about to be a seismic shift in the way that these models are going to have a huge impact in our society we assembled a fantastic team including Rob and many others 30 faculty across 10 different departments from computer science law philosophy education economics and so on and since then we've been really interested in understanding and exploring all the different facets of foundation models so what exactly is a foundation model so chatgpt is one example um Dali is another example that some of you might have played with on the image side but more generally a foundation model is a model that's trained at immense scale on internet scale data um so for example a language model is trained to predict the next word given a previous word so it's basically doing Mad Libs for example Stanford University was founded in blank and tried if you complete the sentence so that the neural network is trying to do this over and over again and then the process it actually learns a lot about language about reasoning and knowledge and learns abstractions not just uh you know superficial uh correlations and what can these models do so that's the thing that's really interesting is that these models are emergent so that April you can't predict what they're going to do only after you train them you start you know basically talking to them you realize they can do much more than we initially thought and so in fact they're actually very general purpose engines that can be adapted to a wide range of Downstream tasks answering questions summarizing documents writing emails generating poetry and so on so if you play with chat EBT you kind of get the sense of what I'm talking about okay so that's the context of foundation models which is the backdrop and now I want to explore its ramifications on education so I'm not an educational researcher although I aspire to be one maybe one day um but I I do teach so I'm an educator so I spend a lot of time thinking about what to teach and how to teach it uh to to my students and so I'm going to talk about um how we teach how Foundation models can change how we teach and also what we teach and the mental model you should have is that we have these models and they're always present whether it be night or day they're fast they can generate a whole document in just a few seconds they have absorbed broad Knowledge from the internet and finally at present they're still unreliable so it doesn't have it has a very tenuous grasp of what truth is and maybe limited social awareness although these things will perhaps improve over time so that's something to keep in mind as we explore how the interaction between education and Foundation also will happen okay so um how how we teach so there's students and there's teachers and I think Foundation models can be uh useful for for both so on the student side the the bread and butter of these models is ability to answer questions so sort of an obvious place where these kind of apply to student is trying to learn about something they want to ask questions a foundation model can deliver um an answer in probably a very succinct and specific way you can ask follow-up conversation uh questions because these are conversational and so on um but remember these models are unreliable so the answer might be right or wrong um so I don't think this is a fundamental issue um in the sense that many other sources talking to your call peers even teachers I've I've been wrong Sometimes using the internet everything has a source of uh you know there's errors everywhere and I think the key thing is to manage expectations and know when you're using a tool what can you trust and what you can't trust and um part of that is onus on the user and part of that is owner us on the model developer the second thing I think is maybe more fundamental is that when you're answering questions all these models have been trained to basically deliver the most accurate answer as fast as possible and this is clearly not the same as delivering the answer that would enhance student learning so you want some more pedagogically based reward system I think this is an interesting and fascinating research question although the existing models are not optimized for this but there's no fundamental reason I think this uh this Camp can't be done besides answering questions models can also ask questions so ideally you want the model to ask the right type of questions not too easy not too hard based on the students response the model can maybe give feedback and give hints and of course this has to be done in a pedagogically uh um sane way so on the teacher side I think there's some interesting um parallels so one thing that I think is really fascinating and maybe more uh immediate is the ability to simulate uh students so you can imagine a model behaving like a student that answers or asks questions and simulates student confusions um and different motivation levels you know you have it and this could be useful for teacher training to test out different teacher strategies um it could also be useful for students because the best way to learn something might be to teach it to someone else and hear unreliability is maybe not as big of an issue because um you know students are learning something and they're not they're only can be expected to know everything so maybe for now this is like kind of a a good place place to start um a second application is generating problems um so I spend a lot of time when I'm teaching generating problems starting with a core I idea and then creating some context that's somewhat realistic around it and one could imagine having problems generated uh automatically or or assisted at least assisting me and generally a problem um both in terms of kind of the the skin like whether it's about Harry Potter or football or whatever um and also but maybe more substantively the the type of situation that's being presented given a core idea and here we're really tapping into the generative creative capabilities of these models that it's able to generate an infinite set of possibilities given appropriate prompt um a third thing where a foundation models could be helpful is for assessment and providing feedback um so grading is you know is important but it's the probably the the least scalable part of the whole uh Enterprise but it's um you know again very important the these models aren't reliable enough so I wouldn't want these models just to automatically give a grade and then that grade be determining whether someone gets a job or not I don't think maybe this uh is quite there but I think a lot of cases are routine right if anyone who's a theater for a class knows that you kind of see the same thing over and over again um so a model could perhaps cover those and flag the interesting cases for a human to a teacher to look at and um you know providing feedback I think is maybe even more interesting than just grading because providing um looking at a Student Response um suggesting things to study explaining why they might be confused sending even motivational messages of some sort this is all sort of within scope okay so now so we talked about how we teach what about um what we teach so I think there's this kind of growing fear that Foundation models will just do everything and why bother learn um um if uh you you can just ask a model to you know not just multiply numbers but to some code and and I think you know if we look historically the set of Technologies from calculators for arithmetic wolf from alpha for solving calculus problems spell Checkers um even stack Overflow I think really our help people do tests uh faster and maybe lessen the need to develop proficiency at some uh underlying skill you know meant not to mention the slide rule that Dan mentioned earlier so my personal take on this is that I still think students should learn everything about how the world Works bottom up from first principles but understand enough just on the principles but they don't need to be proficient I think there's a difference between understanding in principle how to multiply large numbers from actually being able to do it really quickly you don't need to do that but you need to know the former um and after you've kind of understand something in principle now if a foundation model is able to just do that for you then great you move on to the the next thing and I think maybe a fundamental belief I have is always a next thing we'll never run out of things to to learn and and teach um so for example instead of learning how to um you know write a code line by line maybe you should focus on writing good specifications and and tests which is what people should be doing anyway when they're they're testing uh writing code or instead of just uh focusing on the word by word essay developing kind of ideas in the form of an outline or something I think there's many possibilities that um if you can get the the basic level of skills sort of automated in some sense for you then there's much more you can do but but one thing that's interesting is now the the type of work that you're doing shifts a little bit more to kind of oversight and quality control right and and this already happens in management right you have um you know let's say programmers and then you have team leads managers executive Executives and there's it's not like the people at the top are like have nothing to do there's uh there's still a tremendous amount of skill there and but the skill is very different than what it is at the lowest level and importantly there's an element of responsibility at each point in time a person should be responsible for what their they or their AI or their uh you know direct reports are producing so in terms of skills I I think that using Foundation models like chat gbt is going to be an integral skill in fact there's already I think Ethan Malik who is this professor from Warren actually has a class where he mandates the use of Chachi BT rather than prohibits it um I think like I said before oversight and thinking about um Quality Control is going to be a much more important skill than just focusing on generating the the answer um and then finally adaptability is it's incredibly important more than so and ever because these Technologies are evolving over time and there's no one thing narrow thing that I think is safe as if you in the sense that if you learn it then you'll you'll be able to do it because if it's it's sort of repetitive then it can be automated so I think the meta skill that we should you know students should be really focusing on a sort of General uh the problem solving skills and adaptivity okay so just to wrap up I think overall um I'm fairly excited and optimistic about Foundation models in the role in human learning and education they can ask questions answer questions generate problems simulate students provide feedback in a very a context-specific way the technology does have to improve it has to get more reliable and more pedagogical and and there's a lot of work to be done and a lot of people are trying to make this technology better um I think it's interesting to discuss and maybe we'll have a little bit of chance to dive more deeply into it what will we teach so I believe strongly in still teaching the fundamentals and I think we can do it better but we have to go also beyond beyond that to higher levels of skills like ideation and oversight and the final thing is that it's really important I think when we're having these conversations to think about where things are headed right so five years ago we could barely generate fluent text and then we had gpt3 in 2020 which showed that these models are extremely multi-purpose and then we have chat GPT only a two years later um I guess that's maybe two months ago and then this year supposedly gpd4 is supposed to come out and it's going to be another uh big change who knows but but I think that there's just this progression so uh so I think when we're thinking about how to you know think about um Foundation models in the role of Education that's important to look at the the trajectory so that we're not always playing a catch-up but but I think it's a really exciting time and I look forward to discussing this more with you thanks hey everyone um I'm very humbled to be on this panel when I first got invited I I had a gut reaction of oh my God although I have many projects in this space I still feel like I'm not an expert in generative Ai and I'm still learning but I decided to adopt a growth mindset and take the challenge of just um taking this opportunity to imagine what generative AI could do and also raise awareness of one important application which I have been working towards specifically um how we can use General models to empower teachers as as I think den and also Percy pointed out now generative AI has kind of become synonymous to with chat GPT to most people and when you when it comes to education most of newspapers and headlines have been about how Chad gbt is going to change student writing and how it's going to raise issues with respect to plagiarism or um or other other things like the college essay and how teachers need to transform the way they teach writing to begin with but but a lot of these discussions that focus on stream writing have overlooked is is another important application which is how we can use Ted GPT support to support teachers directly and this relates to realizing that actually most of the language that happens to instruction is not in student writing but instead it's in the teacher student discourse it's a conversations in which teachers um immediate knowledge and many other things to students so for example and this image was generated by Dolly I I see that some other people were doing the same um as I mentioned it um it is a medium for uh for new knowledge it is a vehicle for for um creating fostering a culture of curiosity with Jennifer Langer or sooner I don't know if she's here but it's her research that shows that depending on the type of talk talk moves um and discourse that the teachers use they can foster a culture in which students are really excited to learn um and that is really like the basis and the conditions under which learning it happened to adapt dense words teacher student discourse also um Can can help students adapt different mindsets like my colleague David Yeager has shown through various studies several nature papers that just a very brief intervention to students and teachers around synergistic mindsets which is a combination of growth mindset and stress can be enhancing mindset that shows that when um that students can Embrace stress is an opportunity to grow for growth and learn and not see see it as a negative thing that hinders um their learning it can have a lot of positive outcomes even 10 years down the line and related to the conditions of learning um one thing that Dan mentioned is many students don't have the same access to learning opportunities and that is rooted in the classroom environment oftentimes so if students don't feel like they belong they don't feel like um this is a place where they should be or where they can learn um no learning is going to take place and that's Again part I mostly on the teacher to create a culture in which every student feels like they belong and in all of these areas we know that classroom discourse is important in many ways there's a lot of room for improvement for teachers through the many data sets that I've looked at between on student teacher interactions they found that the average 80 and 90 of the talk is dominated by the teachers and also I'm working with my colleague David on trying to use generative AI to help teachers translate what they're saying to growth mindset and trying to identify good examples from what teachers have said it's extremely difficult to find language in in teacher talk that's growth mindset supportive in another project where we are trying to analyze different types of classroom management practices um differentiating punitive versus disciplining practices we found that teachers use punitive disciplining practices 35 of the time and um disproportionately so in classrooms with more African-American students and related related to equity lots of social psychology and research has shown that African-American and latinx students feel much less like they belong in the classroom than white students so this is not to say that teachers are bad it's the opposite that they have a very important role to play and there's a lot of room for them to grow like for each of us in our job and if they had the tools and the skills um to grow and um then they could transform their students lives so how do we Empower teachers the experts the experts who know their students best the experts you know their material best in what they already do most of the conventional ways of training teachers or giving them feedback and professional development are not scalable as they require expertise you can put Carol dwark in each classroom to teach teachers about growth mindset and also most effective forms of feedback require consistency and regularity so that teachers can sort of continue to reflect improve and adopt like that's sort of how formative feedback works it's also oftentimes not personalized there's not enough resources to adapt um train you know a teacher training curriculum to every single individual differently and also it's not it's oftentimes not evidence driven or adaptive to the teacher's needs so a question that I've asked in my research is how when can we complement existing form of teacher training and feedback using AI um and I've developed an application that provides teachers with feedback using um Foundation models but not generative models these are measurement models that measure the advanced the the extent to which teachers build on student ideas and it also measures other things like they're you know questioning practices and talk time and we found that this super simple approach can not only improve instruction just within two occasions of getting feedback but it also improves student outcomes so we ran a randomized control trial in two different online learning contexts and when it's code in place which is the um the course that Chris speech runs uh for programming and we found that um what that's teachers who got this feedback are take up student idea 16 more they ask 14 more questions gave higher course ratings and students in a different context so one-on-one learning also talked more um and um I guess I'm just combining the results from these two studies so in the coding place in the in the other study we found that students the the students gave much higher course readings and also felt more optimistic about their academic future so basically we have a lot of evidence that a very simple automated intervention can improve both student both instruction and student learning and so far I mentioned that I mostly used um Foundation models to measure teaching practices but not necessarily to generate suggestions because I was really worried about the risks that might rank I felt it was not reliable enough and I was worried that it will generate distrust and also potentially cause harm if you get best suggestions to teachers but with the Advent of these new models I'm not rethinking this question and I'm thinking about how we could enhance the automated feedback to include include generative models so there are a few ways um I I want to illustrate um through a few examples what that might look like and these are all inspired by real issues and problems that um that I've gathered around me so this is actually a question that a t High School teacher asked in a conference that was hosted here at Stanford about how teachers imagine using chat GPT to improve their instruction and um one question like as teacher asks is how do like churches how do I explain a slope in a way that's most engaging to students and I tried inputting this to GPT Ted GPT and it gave very cool ideas it gave like Hands-On activities like use blogs that you know to build their Ram um like use real life examples like an architect drawing um to you know plan a building and all of this stuff which is probably curated from the internet so with like some search maybe you could find these ideas in other ways but it saves a lot of time it if you just have like I don't know five paragraphs of different things you could do to explain this concept to students so that's really cool and the teacher can decide to pick whatever is like most relevant and like doable for their case um in another context um where we are working with a tutoring provider to improve the suggestions that the tutors get and how they respond to students we found that a lot of the responses that tutors give to students this is to elementary students trying to learn math is a very templated generic and not very helpful so let's say A student makes a mistake like um will teach us how would you how could you write for 400 the student says 400 and then the tutor just said incorrect this is really prominent in a data set um and I tried asking Chad GPT what else you could say um that's like you know more growth mindset oriented and GPT said you're on the right track but when we write 40 100 we actually mean 40 times 100 do you want to try again giving the student the opportunity to improve on their answer and learn and understand where their misconception so that's pretty cool and in another context where we're working with the College counseling provider similarly immediated through text um students come with a lot of different problems and many of them actually like Express just their worry or like fundamental like concern that they're not fit for college that they shouldn't go there or that you know if they're first generation college student they're gonna just feel lonely and not don't like their good luck and in this case too the counselor counselors oftentimes just say oh don't worry it will be okay here are some resources to help you and send them a bunch of links that's not very helpful so I asked GPT to introduce a different type like you know how we could support a student better and it generated this response which actually Echoes a lot of what the blank social belongings social psychology research I would say is the right is the core treatment uh message that can make students feel like they belong first of all recognizing that they're not alone that link feeling like they don't belong is a very common issue that students face when they're going through a transition it also gets better with time as they get more familiar with the environment and meet new people and everything so 10gpt also suggests different strategies like um involving being involved in activities that interest them Etc I'm not saying it can be improved but it's still a pretty good start and much better what the actual tutor said or the counselor said and again um I think just want to emphasize that what Chief charging PT does of the chef might not be the best um but it's sort of 70 of the way so in a different context where a student is stressed out Chad GPT gave a response that tells them oh yeah quitting is not the right solution remember that a lot of people go through this stress is a normal part of life take a step back so for example this response is okay but it does not actually Echo the core um message from synergistic mindset which is stress is not only a normal part of life but it's a good part of like life in many cases stress can actually like help your brain from new connections if you can think of your brain like muscle it's a very helpful way to think about stress as an enhancing thing that happens to you and not just something that you have to like avoid and step away from so there are a lot of ways in which you know if you can fine tune this model it could be become even better and I mentioned a few examples of what GPT can say but how it can be used as another interesting question I think that Percy mentioned already like you know practicing teaching with simulated students um I know that Chris and Julia are working on something like that for code in place so um that's a pretty cool idea it could also give real-time suggestions like what I mentioned before and um related to my previous work it can also give post-teaching feedback so let's say if you're talking face to face to a student it might be really distracting to get like real time suggestions oh you should say this should say that but instead you could like aggregate like different um uh like you know thoughts like results or um about what you did and how could you could improve and what's really critical um which I think that a lot of the today's conversation might also talk about how can we use it ethically this is a very sensitive space and we want to make sure that we don't propagate the existing inequities that happen um in terms of like teacher professional development or like student learning so in my work um it is a highest priority for me to involve expert teachers at every step of the research process so for example um we are working for the for the growth mindset or synergistic mindset project we're working with teachers who are primarily serving um low-income students across the country and have taught for several years to develop coding protocols for what what kind of strategies they would recommend in their classroom to foster a better mindset of their students um and also very importantly like any data that used for fine tuning should be representative of your target teacher and student population and one important thing which I think just sort of speaks to the fundamental ways in which many of these AI models are evaluated usually like you know comparing to a grand truth I think in education many times that just does not exist it's really ambiguous most people have used observation protocols know that the integrator reliability very low um much much lower like than to like you know what you would expect in like most AI um models like you know most like sentiment analysis or other AI tests like 80 90 in in observation protocols it's like maybe 30 or something that so um I think it was cool it we shouldn't like Smooth over these disagreements but we should try to model them model the ambiguity model the inter-subjectivity that maybe to some teachers some things are helpful but not to everyone so kind of can we make it even more adaptive can we incorporate can we incorporate some sort of confidence in the prediction should be really interesting and relatedly what if we just rely more heavily on extrinsic forms of evaluation not intrinsic brand truth evaluations like what we really care about is like instructional practice and student outcomes what if we like measure um use that as an evaluation rather than a grand truth which may not exist and importantly a lot of people are concerned about the Privacy um issues when it comes to using these models and so for example in my work because I mostly use transcripts um it is actually quite easy to de-identify data and I think that's sort of the safest way to um you know fuse these models and also in mitigates biases um I also really advocate for these tools to be private to the teacher and um because otherwise you know they we don't want to be used for like you know environing and potentially firing them um or all of those things and it's crucial that um the teachers that other users have oversight over how their data is being used um so yeah this slide just includes some of the people I'm working with doesn't include everyone and some of my partners um and really excited to talk with you thank you [Applause] we'll conclude the short presentations with Noah Goodman great thank you ads slides okay so uh this is a story that for me starts in deepest darkest pandemic when I suddenly found myself a homeschool teacher um I was teaching math to these kids um and I only managed to teach them about 45 minutes a day and yet they made really remarkable progress much more than when I sent them off to school for you know six hours and I was pretty sure this was not because I was an amazing teacher um so you know I dug around a little bit and eventually discovered that this was a well-known uh phenomenon not uncontroversial but well known um Bloom's two Sigma Paradox basically is the finding that in general one-on-one tutoring is much more effective than other more other means of teaching um across all sorts of different configurations um and so this was really interesting and I puzzled about it um unfortunately um Dean Dan uh sent me a good paper um which was uh ironically by uh our colleague in a psychology department uh Mark leper um they had studied uh tutors and the first observation is that Bloom's two Sigma Paradox has an important caveat which is that good tutors are much better than not good tutors might be obvious but it's important um and so Mark uh then spent a lot of time doing uh qualitative psychological studies of what do expert tutors do um and it's summarized in this beautiful uh beautiful paper um and good tutors um they argue basically they have a a constellation of things so the first is that they have a particular set of stages of tutoring they they choose the problems well they introduce the problem in a motivating way um they encourage reflection and so on um and maybe most usefully um really great tutors uh have this this constellation of different kind of tutoring uh stances and actions they they have this Inspire model I've sort of compressed it to one slide and therefore completely messed it up but Inspire is something like uh the tutor needs to know their subject material they need to be intelligent they need to establish a rapport and display warmth um they they need to be Socratic asking questions um and giving hints not giving answers uh they need to be systematic in their their process they need to be indirect um I slightly changed this from their original paper to emphasize process praise as opposed to no praise because my other colleagues in Psychology have pointed this out in the meantime but in particular um good tutors don't directly point out errors basically ever they just ask leading questions um they strongly encourage students to reflect and provide explanations for what they did when they succeed and they're encouraging in a bunch of ways now when I first uh read this this was a couple of years ago I got really excited about making AI automated tutors that would reflect these Inspire principles better um and we did a bunch of specific projects that had some some interesting local successes but that's like very 2021 this is 2023 and so um what I want to show you is uh kind of inspired by our you know new knowledge that Bing chat is Sydney and Sydney has a bunch of instructions so I just um I tried to distill the Inspire model into some operating instructions for Alfred who is a highly effective math tutor this is basically the things on the last slide Alfred provides systematic help understanding errors Alfred asked questions rather than giving directions Etc and then I said okay so what happens when we use Alfred and uh have conversations with students um so here are a couple of example dialogues of uh a student which initially is me having a conversation with Alfred which is basically uh open AIS gpt3 model using the Alfred prompt um and you know at first it's it's pretty good looking so I'm stuck on solving equations that's okay let's take a look at an example problem together here's an example problem um I don't know what to do okay let's try again um student that to me says what if we subtract something and Alfred says that's right if we subtract 5 from both sides we can solve for x it's going pretty well so far then the student says okay what if I just say X plus 5 minus 5 equals ten so I only subtracted on one side right now unfortunately Alfred says that's correct and keeps going right Alfred really loves this so there's a lot of good kind of interpersonal stuff going on with this chat um but it's a little bit worrying now here's another example where I I started with a student me who was stuck on a word problem and now this is problematic for two reasons the first is that it turns out the answer is actually wrong there should be parentheses around the X plus 600 don't you don't need to read it but you know that gives actually like critically the wrong information here and the second which is sort of more worrying for me is that Alfred really is not following his instructions right he's supposed to give hints and not just give the whole answer but sometimes he decides to just give the whole answer um okay so this is another example um which is another example of Alfred not doing what he's supposed to in a sort of puzzling way there's this long dialogue I won't read it to you um but at some point Alfred starts talking about let's look at the equation but there is no equation that's been introduced so I'm like what equation and then Alfred says let's take a look at the equation together shows student the equation there's still no equation okay um so somehow Alfred is not doing what I asked him to do um in addition to not knowing math then I thought maybe the problem is me I'm a bad student so I asked gpt3 to be both people so I have the Alfred instructions and here's the the Timmy instructions Timmy is a student struggling with algebra and Timmy hasn't been able to solve this problem and so then I I just generated whole dialogues between Alfred and Timmy which you can do and it's actually super cool I highly recommend it um I won't read through this for you but it goes into these like dada-esque extreme weirdness cases um in particular here there's this moment where Timmy so first of all just to put it in common ground I'm pretty sure that the answer to this equation is x equals three um I did it and I checked it uh and and so first of all there's this moment where Timmy says oh I see if the right side is zero then the left side has to be zero two now that's fine and that's true but then Alfred says exactly what does this tell us about the value of x and Timmy says it must be two uh so for some reason and Alfred says that's right um very encouraging uh and then Timmy explains his answer in this amazing way I started subtracting 4 from both sides and I got 3x minus 2 equals x so basically that just drops the 4 on one side not both sides then I subtracted X from both sides which gave me 3x minus 2 equals zero that means X has to be two that makes no sense and so the answer is is two okay so what this really shows is that there's some basic misunderstanding of logic and math it here now you might say as Percy did actually that this is okay on Timmy's side because Timmy is a student right he makes mistakes I I'm actually a little bit worried after looking at a bunch more of these dialogues that Timmy is a very bad teachable agent because he's not teachable um like even if if I'm this if I'm Alfred here Timmy doesn't learn from me the way a human would and I think that is really problematic from the point of view of having teachers interacting with these as if they you know are students so just we can talk about that more um okay so poor Alfred um this raises some really interesting questions actually the first is how do we encourage Alfred to follow his rules and fortunately I think there are answers to this um although we need to try them I think reinforcement learning from Human feedback if we gather data from humans about where Alfred is and isn't following his his Constitution his rules um also possibly reinforcement learning from AI feedback if it turns out that the GPT models are good at saying when Alfred is doing the right thing which I don't know yet um more interestingly um it's unclear to me whether Alfred actually understands the motivational effects of what he's doing um I'm not sure how to figure that out and I think it's really important for us to figure out how to evaluate the motivational effects of these kind of Agents it's much harder than the the specific content um also here's a here's an invitation I think we can crowdsource a much better Alfred OS so I want all the education experts to get together and uh talk to the you know the AI people and figure out how we make the prompt that if the model followed it would be a good tutor I think that's something that would be incredibly helpful and and important to put into into common uh a public domain um okay I just want to tell you a little bit about this last point which I think is the biggest failure mode how in the world can we teach Alfred math Alfred was very bad at math um so something that my student Gabriel posia uh and I have been doing for a couple years now is basically looking at the Khan Academy curriculum trying to formalize it um in a in a kind of axiomatic way so that we can understand what does it take to get models like these to learn how to solve algebra problems um I have time to tell you a whole lot of the details here um but we have the system called Piano which is on the one hand a kind of logical system for formalizing pieces of math and on the other hand an agent for that you can train to to or that can train itself to learn how to solve math um so we formalized um these sections of Khan Academy in the tiny little letters are the kind of the axioms that we need there um and then we tried uh training an agent now the the key idea here is that we're doing expert iteration we have a current agent the agent tries to solve a bunch of problems um using their current you know configuration um critically here we constrain the steps to be logically consistent and this is important because uh it turns out that you know logic axioms are worth a tremendous amount of data one Axiom is like I don't know infinite data basically um and then any successful Solutions we go ahead and throw in the training set and retrain the agent to do better um if we do this naively it kind of works so this is uh this piano system trying to solve the the five sections from Khan Academy and it gets the first two pretty well and then it kind of Falls flat um it gets stuck on the next three and when we look into what's going on basically it's not learning the right abstractions that allow it to for instance take combining like terms and use that as a a procedure that it can use for solving equations um so we then explored something where we added the ability to learn new abstractions or sometimes they're called tactics in in logic um and these are basically just oh I used this sequence of operations three different times in my successful Solutions maybe I should wrap that up as a sort of action I can take a macro action um so then the the training Loop there is okay let's just do that let's uh get the successful Solutions and then let's compress them by using abstraction in this way and then let's try to retrain it keep going and it turns out this works really well um that now the the piano agent is able to solve uh all five of the Khan Academy algebra sections um and furthermore I will show this to you but if we look into the details of the abstractions that are learned they're very interpretable they're they're pretty easy to map back onto specific sections for instance of the Khan Academy curriculum uh and to kind of use therefore for interesting uh teaching uh tutoring things um which is in the paper it's good um okay so let me summarize um so the the this is sort of a series of examples which I really intend to be illustrative examples of kind of the exciting possibilities and also the issues with using generative Ai and these emerging AI techniques in education um I'm actually kind of optimistic about this sort of Alfred type thing even though it was like really comically bad in a you know a lot of the cases because I think it has the basic skeleton there and I think we can make it better and it's a very concrete it's a very concrete research question what to do um in the last section um you know I think the encouraging thing to me about the piano uh project is that we really can teach these AI systems how to know the logic how to solve the math problems um it's an open question that my lab is working on how do we combine the best of basically piano and Alfred the best of the sort of open domain language models with the specific logically constrained solvers um ask me later um this in turn raises something that I think is a very important question that I don't have any idea what the answer is which is even if I teach a language model how to solve problems it's not clear to me that that's going to lead to a good Alfred so it's not clear that solving problems automatically transfers to systematic teaching of problems so I kind of want to raise that last question as a a challenge for this this room in the next year or two to figure out thanks I have at least a few questions and exchange and from the audience as well on slido um so I want to ask a quick question of all of you and I want to preface the question by saying you know I said I was a philosopher by by training hanging out in in on this side of Campus as a as a philosopher is a a phenomenon for me that elicits the following emotional response sometime which is something along the lines of technologists are inventing the future and there's reason for enormous optimism we heard optimism on a couple of occasions about the great promise of generative Ai and the kinds of uh extraordinary things that seem possible you know person you mentioned GPT coming out just two years ago and then chat EPT two years later like in philosophy there's no such thing as massive progress within a two-year span the result for me is I often feel like you know Eeyore amongst a whole bunch of Tiggers and um in Eeyore orientation is a bit of a drag I want to admit but nevertheless I wanna I wanna use that as a way of framing question for all of you ewers if you accept that um you or me you guys are the Tiggers if we all you Tiggers if you accept that which is let's take the analogy or the idea that these generative AI models are kind of like the calculator um there's a whole bunch of anxiety at the moments of its adoption and then it comes to be seen as a kind of ordinary thing that's in the classroom for very useful things and no one really has qualms these days about the incorporation of calculators so do you view generative AI as something akin to a calculator we'll have all kinds of hand ringing about it for a while it'll get better more reliable we'll find ways to incorporate it that are safe and then it'll just seem like another wonderful tool along the lines of the calculator or is it unlike the calculator and as the eor let me give you at least a reason to think it's unlike the calculator two reasons in fact number one um when it's writing rather than doing math um the act of writing I want to think is continuous with the act of thinking there's no such thing as the thoughts are clear in your head I just can't put them on paper in words if that's the case then the thoughts are not clear in your head writing is a way of learning how to think so if we allow machines to do the writing for us we're effectively learning allowing machines to do the thinking for us and no longer teaching people to think that's thought number one thought number two is that as you've said these generative AI models can be agential they can be even Incorporated in some sort of embodied presence in principle and in that respect I don't think of the calculator as an agent in the classroom but I do feel like Alfred or piano or other things have the phenomenon of agency and that makes them feel fundamentally different to me lurking in the background there is in the commercial moment that's just beginning with generative AI aren't me on the path now to displacing humans in the classroom in some respect because the agency of these models will allow displacing human beings all right who wants to start with calculators no I'll go ahead get started yeah I love that that's great um I think it's more like the invention of the printing press in that um you know lots of hand ringing it's going to unfold and we won't worry about it so much but I think the implications are fundamental restructuring to our culture and they're going to unfold over a lot of time as our culture co-adapts with this change um and so I think it's not so much that I think you know oh The Singularity is near um I don't even think that like as an agent we're going to replace humans um I think like having more agents is fine you know um they'll join the culture and we'll get on with it it's just that we're going to have a different culture that's what education has always done um so you know I'm I'm optimistic because I think human culture is very adaptable but I do think we're going to have much more profound effect than you know going from the slide rule to the calculator Dora yeah um good question I think so just I'm really interested in that what you said about like the thinking because I don't actually think that's the case that um that these models will um sort of do the thinking for the students in fact I think that what they do is they raise the bar so you have something pretty crappy you use not crappy in the sense of like bad but usually generic um sometimes the argument like is coherent on the surface but not in the very deep level and what students will get to do is actually learn to edit through editing and through thinking about why it might be not perfect um to do do the harder work I think as a you know I'm I suffer from writers anxiety all the time getting from like a blank paper to something and maybe Chad GPT or these other models could get you that that get you that anxiety away and then give room for more learning so that's how I see it in terms of the calculator question I don't know if um that um I think I have a lot of questions about access and like barriers to accessing these models which might not come with calculators as much so it's more about the actual Logistics but um but I do think that if there's access then it could become like before I get you into the conversation personally let me just ask you I mean for anyone in the class and who's teaching now um I wonder if you've had to put on the syllabus a generative AI policy about whether it's permissible person you mentioned Ethan Malik requires it to use other people ban its use um if at the moment it seems like it's undetectable that's the spirit in which I think thinking and writing or coterminous and if we offer you know students of any age the opportunity to do the machine writing especially if there's no structured way of then allowing the student to interact with whatever the output is but it's just a cheating machine undetectable by ordinary plagiarism detectors um that sounds to me like um the the demolition of critical thinking not the augmentation of it [Music] so I have a lot of thoughts here um just to be answer the question directly I don't think it's fundamentally different than the calculator or any of these Technologies because I think these are there's always going to be higher level things for for humans to do and and you know I think even when the calculator was invented it was also very profound in many ways and I think a lot of what we're experiencing is the the transition our curriculum has has not had time to update in the last two months as a result um I think the point around agency and all this I feel like it's it's really a function of how we view these Technologies and how we develop and use them I don't think of these as static object I guess beyond the technology side I feel like there's AI researchers and more broadly conversations with others have a responsibility to develop them in a way that um they are going to augment uh human capabilities as opposed to squash them and I see a lot of opportunities for these models to enhance the um in the collaborative process of brainstorming and I I think that at some level the human being is has to take responsibility and if the system is able to generate exactly what it wants maybe the the challenges can you go farther than what this system can generate and and I think that um that that's kind of where I see the future is is going is that these Technologies are coming and they're sort of getting just air dropped on us right now and I think we have to figure out how to assimilate and incorporate them in a way that makes uh sense all right augmentation rather than displacement or some type of thing no let me go back to you just for a minute here because the examples of Alfred and piano and the way you started the you know homeschooling by necessity and you know you pointed to this idea that you know high quality one-to-one instruction is extremely powerful I'd say goes back to Aristotle who is also a one one-to-one tutor for a variety of people so um ancient wisdom um still present today but that's the spirit in which I want like if Alfred and piano or whatever the things are called commercialized non-commercialized in whatever way they get much more powerful um won't the Temptation be to get rid of the 35 person classroom with a single human being in front of it because you'll have shown that it's a far more effective means to instruct students it's a more powerful learning setting in a one-on-one setting so how do you think about the augmentation versus displacement like okay as a psychologist I think about education as part of our our channel for cumulative culture right we have this thing called the cultural ratchet which is the most unique aspect of humanity in my view where like you know humans are kind of weak individual learners but collectively over Generations we're incredibly powerful can you pass information forward the thing about that Communication channel for the cultural ratchet is it hasn't been static it's been a it's been undergoing cultural Evolution and what we now call education is the most recent version of that cumulative culture Channel and like it works a lot better than than what we had you know before but it's not the best as Chris Peach who's here somewhere always says there's always a better way and I think that this is the better way like it's not gonna it is going to get rid of that classroom with one teacher and 35 students but it's not going to replace that that teacher it's going to have you know one human teacher and you know 35 AI assistant teachers for those 35 kids and then they'll get to spend their time in more productive ways they'll be able to educate much better I'm Tigger here I think it's going to be great all right I want to get one one or two questions from from slido um uh how about the following this is a good one it's connected to what we've just been talking about in the 21st century economy and the labor force that will evolve in order to update along the lines of what generative AI within educational settings makes possible will writing continue to be a valuable skill or will familiarity and expedience expediency with the machine output of writing be the valuable skill I suppose it depends on what you mean by writing if it means generating every single word of document I don't think that is going to be a valuable skill in itself just like being able to do algorithm of the calculations to solve some engineering problem is valuable but I think a lot of writing is not just the mechanics of putting out words but the ideation and I think uh being able to think about what is the craft the argument and to revise I think you can just imagine a world where you can do like 20 drafts uh whereas before you might do only two just because you have this deadline that you need to meet so I think you can do maybe much more than but I think I would still call that writing yeah or you know we're gathered today um um to talk specifically about um the intersection of generative Ai and education and let me let me ask either any of you the following in my in my mind we should make an important distinction between the deployment of generative AI for people who have learned to write let's call it in in some professional setting in which we can imagine just what you described there Percy it's an augmentation um you put some prompt into the the machine you get some written output and then you can Tinker with it for your own purposes and that requires some higher order skills yeah let's imagine for Middle School where no one's yet really learned how to write the whole point is we're not augmenting an existing writer we're trying to cultivate the capacity to write in the sixth grade classroom and to me this feels more dangerous than than than something to celebrate at the moment absent finding ways to you know maybe go along the lines of what Noah was describing one-to-one personalized instruction and learning to write better but right now it's a dream for any sixth grader to just turn in an essay in a 35 person classroom that'll probably get a decent grade yeah I guess I want to turn to the motivation of my talk which is I think too much of the discussion is focused on this question of writing which is an important piece I think of learning or instruction and thinking but I don't think it's it's to be a be-all end-all of of Education I think there are so many other pieces and maybe when the calculator came came out people are having the same discussion like will people be able to like cap themselves like you know all this thing can they do mental math anymore if you have a calculator all the time similar with smartphones like you know not that you don't have to remember anything like what's the yeah like does that how does that affect affect your memory like there are these questions but I think I think what we should be asking instead is what are um yeah like how do we transform the way we think about learning and I do think that there's a lot of room now that as actually like also enabled by generative AI um form different ways of learning yeah yeah kind of similar to that I I think something that we sort of have to do is disentangle the cultural conflation of writing with analytical thinking um because you know I think YouTube and all of this stuff has already started to sort of displace the role of writing in our culture generative AI will just demolish you know the the standard role and and I don't think that's terrible we have to think okay why do we spend so much effort teaching kids to write I think it's actually because we're trying to teach them analytical thinking and so if we disentangle those we might find like super exciting new ways to teach analytical thinking and you know separate that focus on that maybe the ai's help with that right and to me that's the important skill writing is important but you know right the service of analytical things yes yeah go ahead I would say that a sixth grader should be able to write a paragraph all by themselves you would or would not I would say they should just like they should not live there Percy yeah just like they should be able to multiply two-digit numbers all by themselves maybe not the as fast they can't do it in their heads but but I think that is a fundamental skill to understand the mechanics of writing full sentences and um but but I think as soon as they can do that they don't need to spend you know years and years unaided right because once they know how to do that then they can go on and focus more on the analytical you know abilities and I think I see in you know in my classes people get just so dragged down by a lot of the details I'll speak not so much writing uh yeah English but writing code people spend a lot of time figuring out the syntax and you know debugging in us in a way that they lose sight of the bigger picture of what they're actually building and what properties this is um they're focused too much on the how rather than what and I think Foundation models have the ability to lift the discussion to thinking about what we should build which is perhaps even more important than How We Do It All right well I think that's a good place to end this first panel conversation would you please join me in thanking our distinguished guests

Channel: Stanford HAI

Views: 7,541

Keywords: Stanford HAI, Stanford AI, artificial intelligence

Id: Ks7enkKuZIo

Length: 64min 54sec (3894 seconds)

Published: Wed Mar 08 2023