Why Open Education Will Become Generative AI Education

Summary

David Wiley discusses the future of open education, asserting that it will evolve into generative AI education. He reflects on the history and goals of the open education movement and how generative AI will significantly impact the creation, revision, and distribution of educational resources. Wiley highlights how the emergence of generative AI can transform teaching and learning by removing barriers and providing tailored educational experiences for diverse learners. He argues for shifting focus from traditional OER to generative AI to enhance access and effectiveness in education while acknowledging the challenges of this transition.

Highlights

• 🎓 Open education's future lies in generative AI, as discussed by David Wiley.
• 🚀 AI-empowered resource creation could widen access to diverse, tailored educational materials.
• 💻 Potential for increased cultural and linguistic localization in educational resources using AI.
• 🗣️ Generative AI tools offer possibilities for innovative pedagogy, removing barriers of expertise access.
• 💼 Parallels drawn between early OER skepticism and current generative AI concerns in education.

Key Takeaways

• 👨‍🏫 Generative AI presents a new frontier for open education, revolutionizing resource creation and customization.
• 🔄 The transition from traditional OER to generative AI can improve educational access and effectiveness.
• 👥 Addressing equity in education can be enhanced through the strategic use of generative AI tools.
• 💡 David Wiley emphasizes the importance of solving problems in education rather than clinging to outdated solutions.
• 📚 Echoing past challenges, objections to AI echo earlier debates on open educational resources.

Overview

David Wiley takes us on a journey from the origins of open education to its potential future driven by generative AI. He reflects on the movement's aim to make education more accessible through open resources and how generative AI can amplify these efforts. Wiley makes a compelling case that generative AI can produce educational resources that are not only tailored to specific learner needs but also easily updated and refined, providing dynamic learning tools for diverse educational contexts.

While open educational resources (OER) have transformed access to educational content, generative AI has the power to revolutionize it further by creating on-demand, customized resources. Wiley's talk explores the idea of letting go of traditional OER in favor of solutions that embrace generative AI's capabilities. This shift could address many of the past criticisms of OER, such as quality and sustainability, by utilizing technology that allows for continual improvement and adaptation of educational materials.

Wiley doesn't shy away from addressing challenges and criticisms associated with this shift, acknowledging parallels with those faced by the open education movement decades ago. He calls for a focus on the underlying goal of equitable educational access rather than being attached to the methods of achieving it. The use of generative AI in education holds promise for creating more inclusive and effective learning experiences, marking an exciting evolution in how we think about and deliver education.

Why Open Education Will Become Generative AI Education Transcription

• 00:00 - 00:30 [Music] all right let's get started it is September 19th 2024 and you are at a special presentation with Dr David Wy at the University of Regina uh before we get started I'd like to provide a land acknowledgement I would first like to acknowledge that I'm here today on the traditional territories of the neak annab Dakota Lota and nakota peoples in the homeland of the mate Nation find
• 00:30 - 01:00 myself extremely fortunate to live work and raise my children on these lands I've been reflecting lately and you know over the years I've been I've learned from the generosity and insights of indigenous teachers and students it has only been recently that I've come to fully appreciate how their influences shaped my views on knowledge as something to be shared and nurtured much like the land itself this understanding has guided my approach to open education which at its core seeks to embrace sharing and reciprocity as we turn today to today's
• 01:00 - 01:30 discussion on open education I offer this acknowledgement with deep respect and a commitment to continue listening and fostering practices to honor these values I want to First acknowledge before we get into the presentation I acknowledge uh our staff I'd like to recognize sha an the ctl's open education and Publishing manager who's developed and organized a stellar series of O related workshops including this one Shauna has done an incredible job supporting open education initiatives across the University of Virgina
• 01:30 - 02:00 I'd also like to acknowledge a technical swort of Sayad musavi who's helped with a great deal with the technical aspects of this presentation including the YouTube stream and just setting everything up and getting over our registration problem which we ended up having a limit of 500 people but uh something like 800 people ended up uh connecting we will have the chat on today we'll also have a Q&A we may not get to the Q&A today um David's mentioned that if if if you do have questions you you could put them in the
• 02:00 - 02:30 Q&A and he if we don't get to them today he may off may be able to answer them on his blog uh so that was very gener generous of him as well so a bit of an introduction to our speaker our speaker today is none other than Dr David Wy David is already known to many of you and we've already included a full bio on the registration page so in the interest of time I'll spare you some of the detail well David is currently the chief academic officer of lumen learning I know him as one of the most important
• 02:30 - 03:00 and early thinkers and doers in the open education movement I've been highly influenced by his work as far back as the 90s he's cited in my dissertation I've used the work hundreds of times in presentations over the years and his work has been highly influential for so many of us in this room today's presentation why open education will become generative AI education uh certainly a provocative title uh it's it's created tremendous interest as witnessed by the nearly 800 regist ations for today's session as well the
• 03:00 - 03:30 topic has certainly produced some push back for instance if you haven't had a chance take a look at the blog post from Heather Ross from the University of Saskatchewan and the rebuttal from Steven DS important points were made by each of these perspectives and for me as I read these posts it felt a little like the early blogas feere debates of the 2000s uh the only thing that was missing to me felt like I just need a Google Reader to make the uh experience complete David has already always had a
• 03:30 - 04:00 knack of getting people to debate on these topics and I'm certain that today will be no different so with great anticipation and admiration I will now turn it off turn it over to Dr David Wy over to you David thank you Al thanks so much thanks for the invitation um this is a super fun topic these are questions I've been asking myself and struggling with uh for a while now and to Alex point about the early
• 04:00 - 04:30 blogosphere you know these ways of writing and blogging and putting ideas out into the world and the feedback and the criticism or the reinforcement that that comes back from that is just so valuable so I'm so grateful that uh so many folks are interested in this topic and really looking forward uh through reviewing the chat and the Q&A and and other things I and hopefully there will be additional blog posts written after this to uh to hearing more about your perspective and what you think about where this is all going
• 04:30 - 05:00 um quick overview I want to start with Origins and the goals of the modern open education Movement by which I don't mean like the open University open education movement in the UK from several decades ago I mean uh the open education movement that we associate with open educational resources talk a little bit about uh this idea of the most powerful emerging technology and use that as a frame to get into talking about the impact of generative AI on traditional Oar the
• 05:00 - 05:30 emergence of what I'm calling generative Oar how that might encourage us to extend and adapt our pedagogies um address a couple of objections and criticisms um I think it's very interesting that the objections and criticisms of generative AI are very similar to the objections and criticisms of Oar 25 years ago and then I'll wrap up with some concluding thoughts
• 05:30 - 06:00 so first talking about goals strategies and tactics for over 25 years for me I I know everyone kind of came into the movement at a different time and in a different way for me I I date myself back here to 1998 being when I started working kind of um really dedicating a lot of focus and energy to open education um but for over 25 years I believe the primary goal has been increasing access to educational
• 06:00 - 06:30 opportunity Educational Opportunity is made up of several things it's access to other uh people to talk to and argue with it's access to an instructor um it's access to a lot of different things but the open education movement's primary strategy for increasing access to Educational Opportunity has been increasing access to educational materials has been the strategy and the specific tactic that we've been implementing for over a quarter of a century now is creating and sharing
• 06:30 - 07:00 Oar so let me tell just a little bit of my personal story how I came into this again I recognize everybody comes into it in their own way um but just sharing a little bit of my story so you get some of the background and understand where I'm coming from um in 1998 I was working at a as the Web Master at Marshall University which is my alma Mo here in West Virginia and I remember that time having this Epiphany and it wasn't something that I was the first person to
• 07:00 - 07:30 understand many other people had already realized this economists even had names for it um but I remember as I was sitting in my office doing my own work having this realization that there's something very different about a physical resource and a digital resource and with a physical resource if my wife is sitting across the table from me at breakfast reading the physical newspaper I can't read the newspaper until she's done with it and lets me have it so there's some sense in which we have to
• 07:30 - 08:00 compete for access to this physical resource however however if you take that same news and instead of publishing it in print if you publish it in a digital format online a million people can all view it and access it and use it at the same time and to me this seemed like a incredibly important thing for me to understand and filled me with a sense of responsibility and obligation to take this realization and try to do something with it particularly in the context of
• 08:00 - 08:30 Education educational materials and educational opportunity um just because it's possible for you to convert a resource into a digital form and share it online so that millions of people can use it all at the same time does not make it legal for you to do that those of you who are old enough will remember Napster and lime wire and uh other tools like that that people use to convert uh physical resources they had music CDs
• 08:30 - 09:00 into digital files which they then shared um you leveraging all the technical power of the internet um in a way that was technically very compelling but definitely illegal um so just because sharing was possible technologically didn't mean that it was legal now our good friend the burn convention here and I say good friend in the most sarcastic way possible
• 09:00 - 09:30 um you know enters into the picture here because not only can you not share things that belong to other people obviously if other people hold the copyrights you can't go online and share them without their permission but the burn convention once as countries around the world signed on to the burn convention copyright law changed so that instead of if you had a work that you thought you might want to commercialize at some point in the future you might want to protect it in some way you would go and register
• 09:30 - 10:00 for your copyright uh as Nations became signatories to burn copyright law changed so that the very moment a creative work was fixed in a tangible form it was copyrighted to the full extent of the law uh whether you had any intention of ever commercializing it or not whether you wanted that copyright protection or not even if you adamantly did not want that copyright protection it's still attached automatically to your works the moment they were fixed in a tangible
• 10:00 - 10:30 form so you have this incredible technical capability of the internet making it possible to share and yet everything that we create is copyrighted to the fullest extent of the law the moment it's created whether we want that or not um and so in 98 I was reading on slash doot which was the place where we all hung out at the time all the Nerds and Geeks back in the day reading about what was happening with the free software movement the open-source software movement and it occurred to me that we maybe we
• 10:30 - 11:00 could do the same thing with educational materials maybe we could take an open- Source style license which makes it legal for people to download the source code make changes to it adapt it use it in whatever way they want to use it maybe we could apply that same approach to educational materials and so in 1998 I launched the open content project which um included the first open license for educational materials and other kinds of content that aren't
• 11:00 - 11:30 software um so that happened in 98 in 1999 I collaborated with Eric Raymond who's the author of the cathedral and the bizaar uh which is an incredibly important paper I think everyone in our movement should read that paper um on the open publication license shortly after that the FSF released the ganu free documentation license or the gfdl and then in in 2002 uh the Creative Commons licenses were first released and as the Creative Commons licenses became
• 11:30 - 12:00 increasingly popular we deprecated the ocl we deprecated the opl and finally kind of the last throws of these alternative open licenses for non-software Creative Works uh it all kind of ended in 2009 when Wikipedia finally switched from the gfdl to the Creative Commons attribution share loock license um but the point of all of this work around open licenses for Content was that then what the internet made possible from a technical perspective
• 12:00 - 12:30 was made legal right I could take my own resource and I could put it online and share it and in a very easy way let you know that you had permission to engage in what we would Now call the 5r activities so what the internet made possible open licenses made legal and 26 years later now from 1998 um huge progress in terms of improving access to educational opportunity right over two and a half billion Works have been licensed under a
• 12:30 - 13:00 cc license I recognize that a large number of those aren't educational materials um and just sharing numbers that I know because they're from my work like lum's Oar have generated over over a billion page views and over 2 million students have been assigned our Oar course where as they required course materials and then you think about the impact of all the other individuals and organizations that create and share Oar 26 years later we've made some progress um this quote from the open syllabus
• 13:00 - 13:30 project here says in 2013 Oar textbooks were barely on the scene in US higher ed used in only about one in 400 classes 10 years later it was one in 80 and Canadian Oar adoption looks very similar both in terms of rate of growth and its heavy focus on text vooks so we're making some progress um I think the thing that's interesting about where the Oar movement started this idea of applying open
• 13:30 - 14:00 licenses to content giving people permission to engage in the 5r activities all happened in the context of what was the most powerful emerging technology of its time right um what we would probably call the information age uh at this point the most powerful emerging technology which was the internet uh could instantaneously make and transmit copies of educational resources at essentially no cost anywhere around the world that there was an internet connection
• 14:00 - 14:30 and in that context the tactic of creating educational materials openly licensing them and then sharing them as Oar so that they could be made and distrib so that copies could be made and distributed freely and legally around the world that was the optimal way to increase access to educational resources uh that was the most powerful technology this was the way that we could utilize its affordances most effectively to try to meet our goal but what I want to suggest
• 14:30 - 15:00 to you today is that we're beginning to transition from an information age where the most powerful emerging resource was about making copies and transmiting those copies around into a generative age which is a time when the most powerful emerging technology can create new resources on demand and these newly created resources can be ordered to spec you can say I need it on this topic I'd like these specific kinds of examples use this specific kind of pedagogy create this
• 15:00 - 15:30 kind of practice along with it with this kind of feedback I want it in this language Etc um so thinking about it in in kind of a fun way if you're not familiar with the Drake equation this is a fun kind of thought exercise thinking about what are the number of civilizations out there in the Milky Way with which we might be able to communic extraterrestrial
• 15:30 - 16:00 civilizations and the Drake equation is kind of a funnel a funneling down or a winnowing down you start with the average rate of star formation in the Galaxy then you think about the fraction of those stars that have planets the fractions of those planets that could potentially support life the fraction of those planets that actually do develop life right you see how this kind of Wis down till finally you get to some uh some number and this isn't meant to be like an exact scientific formula obviously there's tons of estimating
• 16:00 - 16:30 going on all throughout here it's just a way of thinking about the way that things have to filter filter filter down before you reach the point of being able to estimate how many civilizations out there could we potentially talk to I think there's something similar to the Drake equation with Oar when you need an open educational resource on a specific topic you have a very kind of Drake Equation like funnel that you have to run through first you have to hope that someone else
• 16:30 - 17:00 somewhere else has already created the resource that you need then you have to hope that they've openly licensed it then you have to hope that they've either published it online where Google can find it or they've put it in a repository and created sufficient metadata for it then you have to hope that Google actually did index it or that you're able to find your way to the specific repository where it's stored and then you hope that all of your
• 17:00 - 17:30 searching effort actually results in you locating that resource so there's definitely the same kind of funneling winnowing down um when you need a resource on a specific topic for a specific group of Learners you do have to run this kind of Drake Equation sort of gamut whereas with generative AI when you need an educational resource on a specific topic for a specific group of Learners you describe what you need and then generative AI creates it you're not hoping that someone else
• 17:30 - 18:00 created it hoping that they openly license it running that that whole gamut there's something there's something very powerful about living in a time and granted these capabilities are very immature today but they're improving all the time um but living in a time when we can ask for resources on the topics we need for the students we need Etc and have them created essentially in real time now there's a saying that's famous
• 18:00 - 18:30 um in entrepreneurship schools that you should fall in love with your problem and not your solution let me give an example of what it means to fall in love with your problem not your solution say your problem is I need a way to keep food drinks and other items cold at one point the solution to this problem was to harvest ice from frozen lakes and deliver that ice to customers homes then there were some technological
• 18:30 - 19:00 advances that happened and a better solution was to make ice industrially to make it close to where you needed it and then from there deliver it to customers homes and then there are additional technological advances that made it so that we didn't have to have ice delivered to us at all we could all make we could make our own ice in our own homes and kind of famously the people
• 19:00 - 19:30 who harvested ice from Frozen Lakes um did not make the transition into industrial ice making and the people who did industrial ice making did not successfully transition into the business of inhome refrigeration and freezers um the people who had focused on these previous Solutions were not in love with their problem they were in love with their solution and as new technology came along and their solution
• 19:30 - 20:00 became outdated because they didn't see that actually we in the business of helping people keep food drinks and other items cold they thought they were in the business of harvesting ice and delivering it to people's homes they weren't able to survive through that technological transition so if you fall in love with your problem you can always leverage new technology to solve your problem more effectively and efficiently if you know that you're in the business of keeping food drinks and other items cold then it
• 20:00 - 20:30 makes complete sense for you to stop making ice and delivering it to people's homes and instead now get into the business of inhome refrigerators and freezers but if you don't fall in love with your problem if you fall in love with your solution then eventually you'll be left behind so if the problem that we care about in the open education Community is inequitable access Educational Opportunity if that's our problem and
• 20:30 - 21:00 our primary goal has been what I've said and the strategy has been what I've said generative AI has the potential to provide access to dramatically more resources on more topics in more languages in more formats than the current kind of create each Oar by hand through a bespoke process does and if that's true then the optimal tactic or what we might call the best
• 21:00 - 21:30 current solution for accomplishing the primary goal of the open education movement is now generative Ai and a group of people whose primary goal is to increase access to educational opportunities by increasing access to educational materials I believe really should begin shifting their primary focus from thinking about Oar to thinking about generative AI now Oar will still have a role to play in the future just like printed textbooks will continue to have
• 21:30 - 22:00 some role to play uh into the future but that role is going to decrease over time I believe so let's talk for a minute about the impact of generative AI on Oar when I say traditional Oar I'm going to contrast that with what I'm going to call generative Oar in in a minute but by traditional Oar I mean this very common definition the hulet Foundation definition Creative Commons definition
• 22:00 - 22:30 Etc that Oar are teaching learning and research materials that either reside in the public domain or have been released under an open license that permits their free use and repurposing by others so we would typically think of these as being articles chapters textbooks images videos audio recordings simulations etc etc now generative AI profoundly affects the way that traditional Oar can be authored and the ways they can be
• 22:30 - 23:00 revised and remixed and I want to talk about that for just a minute um and and maybe set up this kind of before and after of before generative AI all Oar were handcrafted but now with generative AI they can be AI drafted um and generative AI can reduce the amount of time and resources it takes you to get to your first draft by an order of magnitude by which I mean take the amount of time you used to spend divide it by 10 um and that's
• 23:00 - 23:30 probably still more time than it's going to take you to arrive at a draft in the handcrafted model compared to in the AI drafted model and we certainly see this in our own work at Lumen and other folks working in instructional design that I talked to at companies at nonprofits at universities if you're if you work in the business as an instructional designer whether you're at a for-profit or non profit if you create educational
• 23:30 - 24:00 resources for a living you're probably already using generative Ai and it is reducing the amount of time it takes you to make that first draft by an order of magnitude now philanthropists and others who want to be effective stewards of funding I think are going to see that this is true and I think they'll likely begin requiring new oar that is funded under grants to be AI drafted because all other things being equal you know after that first draft
• 24:00 - 24:30 the Oar should still go through a QA process a peer review process an editorial process like it's just a first draft right but all other things being equal why would you pay 10 times the amount to get to that first draft than you could if people were using an AI drafted approach to that first draft as opposed to a handcrafted approach to getting there I I think there's some interesting questions around stewardship of public funds or philanthropy to be
• 24:30 - 25:00 asked there and revising and remixing traditional o I think there are big implications here from generative AI as well again open licenses make it legal to revise and remix educational materials but open licenses do not give you the skills and expertise that you need to engage in maybe some high demand revise and remix activities like translating a resource into another language CC license might give you permission to translate that textbook
• 25:00 - 25:30 into Spanish but if you don't speak Spanish it doesn't help you it might give you permission to turn that static resource into an interactive resource but if you don't know how to do that permission doesn't help you or adjusting the reading level of a text from college to high school or simplifying it for English as a second language Learners an open license gives you permission to do that but if you don't know how you're not going to be able and this reminds me of Berlin lends two concepts of Liberty if you're familiar
• 25:30 - 26:00 with this I think this is kind of popular in the in the Oar space but um just in case there's he talks about negative Liberty and positive Liberty where negative Liberty is freedom from interference by others the absence of obstacles barriers or constraints there's nothing stopping me from doing what I want to do I'm not being prevented that's negative Liberty because positive Liberty is not just about the obstacles being moved it but it's about you having the capacity and
• 26:00 - 26:30 the ability to act to be able to do the things that you want to do um to make your own choices and achieve your own potential so it seems like there is an interesting point to be made here about Oar generative Ai and the two Liberties right because as I said a minute ago open licensing removes those obstacles and barriers to action that are created by copyright law or in other words open licenses increase our negative Liberty with regard to Oar but open licenses do
• 26:30 - 27:00 not magically endow us with the skills and expertise that we need to be able to take advantage of those opportunities so open licenses leave us still with a deficit of positive Liberty but generative AI gives us access to those skills and expertise which increases our positive Liberty with regard to Oar so if you think about time and effort as obstacles to revising and remixing um we published a paper over a decade ago showing that the more time
• 27:00 - 27:30 and effort are required to engage in a specific Revis and remix activity the less often users will engage in it so for example users are far more likely to delete a chapter from from an open textbook which might involve you know hitting highlight and hitting a delete button or something very very simple very quick like that far more likely to engage in a very easy and quick Revis remix activity like that than they are to write a new chapter or to make
• 27:30 - 28:00 extensive revisions to an existing chapter but in as much as generative AI makes these kinds of tasks faster and easier I think now in the context of generative AI we should expect to see more revising and remixing happening in the future because as the revising and remixing gets easier and faster more people should be likely to engage in it so that's one and a second point is
• 28:00 - 28:30 around the quality of revise and remix so um if you're a follower of Ethan mik on Twitter or on LinkedIn or wherever you might be following him he is fond of highlighting research that shows that productivity gains associated with using generative AI are the highest among lower skilled workers and if you think about that that makes sense for example for a translation task generative AI is going to be a lot more helpful to somebody who doesn't speak the second
• 28:30 - 29:00 language than it is to someone who does speak the second language if I don't speak the language that the resource needs be translated into I can't even get started so it's dramatically more helpful to me than it is to someone who speaks the second language right so in as much as generative AI makes these tasks possible that were either previously impossible or impractical we ought to expect to see an increase both in the variety the kinds of revising and mixing that will happen in the future as well as an increase in
• 29:00 - 29:30 the quality of the revising and remixing that happens in the future and I think those are all positive things um pivoting for a moment from traditional Oar to what I'm going to call generative Oar generative generative Oar are not designed to be used by teachers or students directly for teaching learning or research they're designed specifically to be used
• 29:30 - 30:00 as part of a generative generative AI system to create other Oar and the the two things that I think are the most interesting for us to think about here are openly licensed prompts and openly licensed model weights so let me take a drink here for a second let's talk about each of these and how they interact with the five RS and some of the other things that we've
• 30:00 - 30:30 been talking about open prompts many of the prompts that are written by firsttime users of generative AI are relatively simple something like write me a poem about rainy days all right but when you really get into the work of trying to prompt these systems to do things that are more um going to be more powerful or more effective pedagogically you're trying to elicit more complex
• 30:30 - 31:00 Behavior a prompt like that can be hundreds of words long or thousands of words long and once you're into writing thousands of words guess who shows up our old Nemesis burn and now we're back in the place where these more useful prompts are all automatically copyrighted to the fullest extent of the law the moment they're written which means if we want to be able to share prompts with each other and permit each other to revise them and remix them it's going to require open
• 31:00 - 31:30 licensing of prompts openly licensing prompts matters for a couple of reasons first I think there's broad awareness within the open education community that we need to be able to localize Oar for linguistic reasons to make them more culturally appropriate for for a whole host of reasons I can't possibly sitting here in my home office in West Virginia anticipate every need
• 31:30 - 32:00 that you and your learners who are somewhere else are going to have and so when I put Oar out into the world part of why I put it out into the world under an open license is that so that you can take it and make the changes that you need to make to it so that it will actually work for you and the Learners that you're working with whether you know whether it's you individually as a learner or you as a teacher working with Learners or you're part of a group there are changes that are going to need to be made and you need to have permission to
• 32:00 - 32:30 be able to make them um users of generative AI are going to use are going to choose to use different generative AI models for a range of reasons maybe my campus has a subscription to Microsoft co-pilot or maybe we have a deal with open AI so that everyone has access to chat GPT or maybe I'm an independent learner and I can only use the free tiers of some of
• 32:30 - 33:00 those or maybe I'm using an open model on my own laptop for economic reasons or because I care about privacy or for the point is that different users will choose to use different models for a whole range of reasons and different models will respond differently to the same prompt right so another kind of critically important aspect of of localizing prompts is refining being
• 33:00 - 33:30 able to refine a prompt so that it will work with your model because it was probably given all the models that are out there was probably designed to work with some other model and you're going to need to do some work to it to make it work for you PRS are fairly straightforward they look a lot like traditional Oar they're you know a bunch of words um open weights I think are more interesting and and potentially more powerful certainly
• 33:30 - 34:00 more powerful um and so I want to I want to linger here a minute longer than we did on open prompts now just say right at the beginning I am purposefully skipping over all the debate about what constitutes an an open model OSI is looking into that there are lots of people arguing about that we're not going to go there in this conversation I just want to talk about open weights okay so in a generative AI model
• 34:00 - 34:30 weights are now this may not be a helpful comparison to you and I apologize but the weights in a generative AI model are conceptually similar to the weights that you calculate when you do a linear regression a weight is a number like that number there on the screen and when we say the model weights what we mean is we mean the dozens or hundreds of
• 34:30 - 35:00 matrices that are storing the millions or billions of these numbers within them right this this is the heart of a generative AI model is this collection of Weights which is a bunch of numbers organized into matrices that correspond to the different layers of the model so when I say open weights what I mean are these weights these numbers in all
• 35:00 - 35:30 their matrices all of them all the all the matrices all the numbers that comprise the model those being licensed in a way that we can retain revised remix RS and redistribute them and there are many models there are many I should say there are many projects that have released their model weights under open licenses this isn't like pie in the sky this is already happening today but it's not on the radar of a lot of people in our
• 35:30 - 36:00 community which is what I part of what I'm hoping to accomplish by talking a little bit about it today why would open weights matter and let's let's linger here a little longer uh then we lingered on the conversation about prompts as well foundation models those models that have just that have been through the pre-training process maybe haven't had any fine-tuning at all like the instruct fine tuning even if they've had the instruct fine tuning these Foundation
• 36:00 - 36:30 models like uh claud's Sonet 3.5 or a GPT 40 or even an 01 like as was released recently they are not designed to behave pedagogically most often they're designed to try to make you happy um to try to follow your instructions and give answers to your instructions or to follow your instructions in a way that's going to satisfy you that's that's what they're
• 36:30 - 37:00 primarily designed to do at least the ones that we interact with um they're not designed to behave pedagogically and they can lack disciplinary knowledge they can lack cultural knowledge you can lack a whole range of information that's necessary for a specific teaching or learning situation now with prompting you can temporarily change a model's Behavior just for that that
• 37:00 - 37:30 session during with you're interacting with it or with a technique like like rag retrieval augmented generation you can temporarily change a model's Behavior or change what it knows but if you want to permanently change the way a model behaves or what it knows you have to fine-tune it and fine-tuning means changing the values of the model weights okay and we'll talk about a couple of
• 37:30 - 38:00 things we'll talk about distillation and quantization and other model changes are necessary if we want to be able to run models on consumer Hardware like our own laptops or our own phones and I think that is critically important for a range of reasons too which we'll get into so fine-tuning is is like remixing if you think about the five RS applied to open
• 38:00 - 38:30 weights fine tuning is remixing right and fine-tuning is a process by which you update you change you change the values of those numbers you update the model weights through additional training on curated data and you've curated that data specifically for the purpose of changing the model's behavior in a specific way so for example if you're to fine-tune an open weights model and your fine-tuning data included
• 38:30 - 39:00 10,000 examples of interactions between students and expert tutors then after the fine-tuning process was over that model would behave more like an expert tutor than it had before okay and you can imagine that when you're trying to update millions or billions of these parameters that that is very computationally extensive takes a lot of time takes a lot of resource takes a lot of money so there are techniques like Laura
• 39:00 - 39:30 that make it possible to F tune only a subset of the model weights instead of having to update you know the 405 billion parameters um in the wama 3.1 model so that it becomes possible for people like you and me with the hardware that we have access to to actually do fine tuning it's not just in the realm of you know Google and Facebook and
• 39:30 - 40:00 whoever else but fine-tuning is what remix means in the context of open weights right and distillation is a specific kind of fine-tuning which I think is very interesting for our purposes distillation is a process by which you update the weights of a small model so for example in the recent llama 3.1 release from meta there's an 8 billion parameter model 70 billion parameter model and a 405 billion parameter model
• 40:00 - 40:30 obviously you're not going to run that 405 billion parameter model on your laptop but you can run a version of that 8 billion parameter model on your laptop it doesn't have all of the capabilities of the larger model but this distillation process is a process by which you can update the weights of a smaller model using data that you've curated but this time you're generating the data
• 40:30 - 41:00 with the larger model so that you can transfer some of the knowledge and capability from the larger model into the smaller model so now this fine-tuned smaller model won't have you know the same kind of general purpose capabilities that the larger model has but say you're just you just want to create a smaller model you just want to fine-tune this small model so it's amazing at giving feedback on drafts of student writing that kind of capability you
• 41:00 - 41:30 could transfer from a larger model to a smaller model through this distillation process and again smaller models are faster because they're smaller and they're less expensive to run and might be capable of running on consumer Hardware like the hardware that you and I have I also want to talk about quantization for a minute quantization is it it's more complicated than this but conceptually it's similar to rounding
• 41:30 - 42:00 right so say that in one cell and one of the matrices in your model you have a DOT you know 0.425 63 some value um if you were to round that off say to 0.43 then you reduce the Precision of the model a little bit but you make the model size significantly smaller right if if each of those if each of the numbers stored in each of those matrices
• 42:00 - 42:30 if you've got 405 billion parameters and each of them are 16 characters long and you truncate them to be eight characters long you've cut the memory footprint of the model in half right meaning that it's going to be faster and less expensive to download because it's significantly smaller now than it was before and because it's smaller again I might be able to run it on my consumer Hardware now quantization is not rounding just to be clear I'm
• 42:30 - 43:00 saying it's it's conceptually similar to that the process is more convoluted but it ends up with a higher Precision number being a lower Precision number which is smaller and has these effects so coming back to why open weights matter with permission to engage in the 5r activities we can use a range these techniques I've talked about and others we can use a range of techniques to change the way that models behave to change what they
• 43:00 - 43:30 know to decrease their memory footprint and do other things to them so that they can run on your laptop and eventually on your smartphone and we've talked about revise and we've talked about remix but retain and reuse and redistribute are also very important here this permission to download and run smaller customiz ized open weights models locally I believe is going to be a key
• 43:30 - 44:00 to our being able to use generative AI to improve equity and access um in a way similar to mw's mirror site program does so if you're not familiar with mitw mirror sites there are a couple hundred of them 250 300 something and what MIT does is they take the entire open course forare site copy it onto a hard drive drive and they put that hard drive in the mail and send it
• 44:00 - 44:30 to you because maybe where you are maybe you have a land maybe you've got a local area network you've got Wi-Fi um inside your school building but your school isn't connected to the broader internet so I can't I can't actually use MIT ocw in in this example because I'm not connected to the internet even though the computers inside my building are connected to each other but if mitw puts the the whole open course whereare collection on a hard drive and sends it to me I can plug
• 44:30 - 45:00 it into one of the servers in the school and now everybody in the school can access all of mitw when it's possible to download and locally run these smaller models we'll be able to send them to the same kinds of places that we can currently send uh these kind of static resources like PDFs or websites or things like that so that even access to the broader Internet won't be a requirement to be able to use
• 45:00 - 45:30 them but you'll be able to use them in these offline settings like like the mirror site program makes possible and customizing those models and being able to run them locally does a bunch of things it improves privacy because your conversation's never leaving your own laptop it decreases energy consumption because the model is smaller takes less energy to run and now you're you're running it on your laptop and inad of accessing some giant Cloud where that's happening and if the models are
• 45:30 - 46:00 customized in in ways that are informed by evidence-based teaching and learning practices then we can increase the likelihood of improving learning outcomes through these customized models as well and all of this is only possible if we have permission to engage in these 5H hour activities with regard to the weights so open weights really matter um if you want to dip your toe in here
• 46:00 - 46:30 and play around with running models locally um I've got the names of several um projects here on the slide that will let you do that I think LM studio is kind of easiest to use but your mileage may vary um one of the things I like best about LM studio is that when you open it on your machine and go and start browsing through all the open weights models that are available it will tell you I don't think this one's going to fit on your laptop don't bother
• 46:30 - 47:00 downloading it because it's not going to run but this one looks like it will run go ahead and download it and run it locally and just to be clear when I say run it locally I mean after you've downloaded it and You' started it up in LM Studio you can turn off the Wi-Fi and use this generative AI system right on your own machine without any internet connection right that's what we mean by running it locally okay two more topics and both of them are
• 47:00 - 47:30 shorter than the ones that we've just been through so I'm hoping that we'll be able to get through them first I want to talk about impacts on our pedagogies if you've heard me talk in the past you've probably heard me talk about what I call O enabled pedagogy Oar enabled pedagogy are the things you can do with Oar that you can't do otherwise the the teaching and learning practices that are possible when you have the five R permissions that are not possible POS when you don't have the five commissions right so for
• 47:30 - 48:00 example um if you're using Oar as a core instructional resource you and your students might collaborate to revise and remix and then republish your course materials both as an exercise and improving their own understanding but also to better support the students that will come along next semester obviously you cannot do that with a traditionally copyrighted resource that's a kind of pedagogy of revise and remix that you can only engage in when you have the
• 48:00 - 48:30 five R permissions I think I'm sure there will be something similar although I think there's still a lot of exploring of the frontier yet to be done here but I think there's definitely something analogous to Oar enabled pedagogy that we might call gen AI enabled pedagogy which is the things you can do with generative AI that you can't do otherwise so for example when you're teaching a fully um
• 48:30 - 49:00 asynchronous online course that where you never meet FAO face um we're typically not able to assign synchronous collaborative activities like like a think pair share uh kind of activity in that setting because the students are never together the reason they signed up for asynchronous online is because they have crazy hours and crazy schedules and all these demands and they can only do what they can do when they can do it but if generative AI can play the role
• 49:00 - 49:30 of that partner or that collaborator or um as vigotsky might call it the that more capable other then you can assign students to engage in some of these activities even in this context which you couldn't do before um now that's kind of a trivial case but I hope that one example makes the point that when we can assume that students have access to
• 49:30 - 50:00 generative Ai and if they're taking a fully online fully asynchronous course they're obviously connected to the internet if they have access to a tool like that now there are pedagogies that we can employ that we couldn't employ before so what does that what does that mean and what will that look like I think there's there's a lot there's a lot left to explore there um I'm a big fan of this framing of of thinking about barriers and removing barriers that kind
• 50:00 - 50:30 of negative Liberty perspective that we talked about a minute ago and I would I would suggest that you know 20 30 years ago let's say the internet really removed time and place as barriers to education right now you can do it from anywhere you can do it at midnight etc etc how did our pedagogy evolve in response to those affordances the internet presented to us Oar removes
• 50:30 - 51:00 copyright restrictions as barriers how is our pedagogy evolved to take advantage of that affordance generative AI removes access to expertise as a barrier to education how how is our ped pedagogy going to evolve in that way and when I say access to expertise as an example I mean if I'm doing my homework at midnight I have a question I can't
• 51:00 - 51:30 proceed until I get my question answered prior to generative AI I would just be stuck there's probably nobody else online nobody I feel comfortable calling my fac m is certainly not available for me to talk to to get unstuck so that I can keep moving forward but in the context of generative AI I can have a conversation I can ask a question I can say that explanation didn't make sense to me can you give me another explanation uh actually can you give me
• 51:30 - 52:00 another and uh use Taylor Swift as an example somehow please um that kind of expertise in being being able to answer questions create new examples create new explanations give feedback that's a big barrier removed how will our pedagogy evolve objections and criticisms and I said I I think it's it's so interesting to me having lived through this the first time to see it just playing out
• 52:00 - 52:30 again right before our eyes the number one concern about o has always been quote quality and so I've been asked all of these questions 10,000 times probably if anyone can create Oar how do we know that they're going to be accurate are the authors experts do they teach in the prestigious University who does the peer review what does quality control look like and then the gotcha ah see I found an error on this page you can't trust Oar told you they were going to do
• 52:30 - 53:00 quality problems this whole thing is uh whole thing's a scam and the the second concern has always been sustainability right where is the money going to magically come from to pay people to create new oar once Oar created when there's a new advance in the field who's going to update it when the information gets out of date who's going to update it when somebody finds an error who's going to
• 53:00 - 53:30 update it and why should I go I mean if you if you've taught at the University level why should I go to all the trouble of repr prepping my course from these materials I'm currently using that are working fine for me to start using this Oar when there's no guarantee that it's ever going to be updated and two years from now three years from now I'm going to have to Reep my course again and of course if everyone starts using Oar does that mean that no one's ever going to get paid to write text books what are we doing to authors livelihoods are we destroying you know
• 53:30 - 54:00 authoring as a as a way of making a living if you think about generative AI the concerns are so similar the number one concern is quality if we don't know exactly what data the model was trained on how do we know if its results will be accurate if we don't fully understand the process of inference by which a model generates responses how can we know its results will be accur aha see look it hallucinated it made a mistake right there I told you you can't trust
• 54:00 - 54:30 generative AI this whole thing is a scan and again with generative AI another concern that comes very quickly behind hallucinations and accuracy is sustainability if generative AI is used to automate a lot of work do I still have a job model training both training and inference and inference is what we call it after you submit a prompt and the model generates a response for you that process is called inference training and inference require a lot of
• 54:30 - 55:00 electricity and a lot of water um don't models perpetuate bias isn't the way that human labor is involved in this whole process problematic what do deep fakes and automatically generated misinformation mean for democracy is can this is this whole Enterprise even sustainable right so I I want to give you a framework or I want to suggest to you you you know feel free to W it up and throw the garbage I want to suggest to you though a framework for thinking about these objections first that
• 55:00 - 55:30 generative AI right now is a lot like oar was in 2000 and when the very first open licenses were out there Creative Commons wasn't even a thing yet and a handful of us were stomping around trying to convince people to openly license their educational materials people were somewhere between dismissive and just completely incredulous um at that time we we could see that there were promising Pathways to addressing the real issues that
• 55:30 - 56:00 people had concerned about and we'd already started doing the work um and we knew that it would never be perfect but that it would eventually be good enough right that's where generative AI is today from my perspective most people are somewhere between well I don't think anybody's dismissive anymore but I think a lot of people are incredulous but they might not realize that there are promising paths to addressing issues work's already underway it's never going to be perfect
• 56:00 - 56:30 but arguably now for many use cases it's already good enough and even now 26 years later if we're being completely honest we still don't have great answers for the Oar objections but lots of people use Oar enthusiastically anyway so what I mean is when someone raises a concern about the accuracy or the quality or the sustainability of oar an idealized response to those concerns would be to say let's just take Open
• 56:30 - 57:00 Stacks for an exam they get huge grants they spend as much as a million dollars per book the authors are experts from famous universities they have a whole production process that mimics a big publisher process where the experts get support in technical writing and graphic design and editorial and copyright review and there's a quality control and a peer review process that mimics the ones that the big Publishers use and in terms of sustainability you know open STX generates licensing fees from their
• 57:00 - 57:30 commercial partners and they sell they sell printed books and all of that Revenue means that they'll always be able to update update their materials um and keep them up to date now if there are a million Oar in this world this response speaks to 50 of them right if you were to draw an oar out of a hat and respond to these concerns concerns a more typical response would be well an individual instructor applied
• 57:30 - 58:00 for an oar Grant at their institution and got $3,000 then they wrote an entire textbook in one semester they never received any production support though we did give them a pressbook account and there was no peer review process and there's no ongoing funding to update the resource after this semester long Grant is over that's actually what the majority of R in the world look like and how they work and yet we're all still very
• 58:00 - 58:30 enthusiastic about Oar and using them pretty effectively right um I see we're just about out of time let me share my concluding thoughts if if we're truly in love with our problem which is inequitable access to educational opportunity and we're not just obsessed with our solution then shifting our Focus from traditional Oar to generative Oar seems like the OB obvious path forward generative Oar are capable of
• 58:30 - 59:00 dramatically improving access to different resources on different topics with different examples in different formats in different languages Etc and because generative Oar can be interactive in a way that traditional Oar cannot they have the capability to improve the effectiveness of those opportunities once students get access to those opportunities I want to acknowledge the change from working with
• 59:00 - 59:30 traditional Oar and a wizzywig tool like press books to trying to work with open prompts and particularly with open model weights is pretty similar to the change from working with HTML and CSS to working with JavaScript and apis like on the one hand there are wizzywig tools and like pretty much anybody can do it with minimal training working with generative Oar particularly with open weights is just it's more Technical and it's more complicated I I'm just
• 59:30 - 60:00 acknowledging that but it will also be significantly more powerful in the same way that websites that have databases behind them Etc are significantly more capable useful helpful than websites that are just flat HTML and CSS files final slide call to action so after all these years 26 years later I am still dedicated to this goal of increasing access to educational opportunity and I
• 60:00 - 60:30 believe that many of you are as well um I've also become interested in increasing the effectiveness of those opportunities but that's kind of a separate conversation but I'm still committed to this this first goal and emotionally it's very hard to leave behind or think about beginning to move on from this traditional Oar tactic that has consumed 26 years of my professional life but I think that it's really important for the people that we're doing this work for that we use the best
• 60:30 - 61:00 most effective tools at our disposal as we do the work so my call to action to you is let's use the best most effective tools at our disposal as we do this important work and I know I've taken us two minutes past I apologize but thank you very much well thank you so much uh I'm not sure if I'm on camera there if you can hear me I want to thank you so much David for your uh presentation it was
• 61:00 - 61:30 provocative informative there was a fantastic chat there was some push back of course um I would invite people to save the chat if they want to see because there are plenty of resources that have been put in there uh David's also uh said that he may look at some of the Q&A questions posted and potentially blog them so you may take a look at that as well we'll make the video available if that's okay with David and uh I'm assuming your slides will also be made available somewhere as well perhaps on
• 61:30 - 62:00 your blog uh thanks so much David this was wonderful and I I really want to thank everyone for coming out on this really important topic as this world transitions to something uh quite a bit different but with obviously uh a lot of the ethos from the past uh that's moved into this new uh spans of technology so thank you again for being a part of this thank you David for your wonderful presentation and uh uh by the way if people want to show up for another presentation John wilinsky is uh coming
• 62:00 - 62:30 in on October 22nd we have a number of speakers coming in and you may be interested in John's take on some of this as well uh thanks again uh at the University of Virgina thanks you all and uh take care [Music] [Music]