Ahead in the Cloud: Lifelong Learning in the Age of AI with John Domingue of Open University

Chad Watt: Welcome to a Head in the Cloud where business leaders share what they've learned on their cloud journey. I'm Chad Watt, Infosys Knowledge Institute, researcher and writer, here today with John Domingue, Computer Science professor with the Knowledge Media Institute at Open University. Open University is the largest university in the United Kingdom by number of students. Open University began teaching in 1971 using television and radio to broadcast educational content, which was a bit of an innovation at the time. 50 plus years on, Open University and the Knowledge Media Institute are continuing to apply new technology to its mission to spread educational opportunity. We're talking education, cloud, AI, and automation on this episode of Ahead in the Cloud. Welcome, John.

John Domingue: Good to be here, Chad.

Chad Watt: For listeners who may be unfamiliar, tell us a little bit about Open University and its mission.

John Domingue: So, I've only ever had one job and I've always been proud to work at the Open University. It's an institution that has social justice at its center. So, we believe in social justice through education, and we believe at teaching at scale. So, in fact, we have this mission to be open to people, places, methods, and ideas.

Chad Watt: That's terrific. So, you arrived at Open University in the 1980s. What was the state of the art in education technology at the time?

John Domingue: So, I was a bright young Ph.D. student. I think that maybe just reflecting to think back on what technology was like at the time. So, I remember going for an interview at a famous university, I won't name them, to do my undergrad degree, and I had a long lecture from the interviewer who was, I guess an assistant professor, about why I should learn to program through punch cards because that would really teach me something fundamental about the nature of software. So then finish my degree, I go to the Open University and there was an AI group there already using AI to teach students on a dedicated course. It was an AI in Psychology course, so cognitive psychology course, and I kid you not, but we were using AI over these teletype printers if you remember those. So, students would go to a study center, put in a program in a language that my supervisor had designed and got back a response as teletyped out. And in the middle of that, we were trying to introduce AI, be understanding about what the student programmers were like, what bugs they had and how to fix those. In fact, my Ph.D. was focused on building an AI tutor for teaching LISP, if people remember the LISP programming language. So, students would type in their program, if there was a bug, my assistant would start up, diagnose the underlying concept related to the bug and then give them some advice. But it was a different world to the one we were involved in now.

Chad Watt: So, I guess the big point here is that AI is not new, but I would argue the reason you, me and even my mother is talking about AI right now is that thanks to internet, cloud, fast networks, more powerful computers, AI is newly democratized and experiencing very rapid adoption rates. Is that generally something you agree with and what are the implications of this new very accessible AI?

John Domingue: So, there's a lot to unpack there. So, if you go back to the ancient world we were talking about earlier, so one of the big problems in AI at that time, which everybody's forgotten about now, was called the knowledge acquisition bottleneck. And what that was we were building these systems, for example, to mimic expert doctors, expert engineers, expert people in finance. And the problem was, well, how on earth do we get the knowledge out of their brain into the machine? So, there's a variety of psychological techniques, interviewing them, filming them while they work, et cetera. And now we just don't need that because everything is out there on the web, all the data is out there, so the knowledge acquisition bottleneck is gone.

So, you have somehow, that democratization of data that the web provides at huge scale. And then you have the democratization of access on any machine, that people can have, or even a mobile phone, you are accessing all of that data in various forms. If you think about Wikipedia, if you think about the big search engines, and then on top of this now recently we are giving access to any sort of intelligence that we have. I guess the excitement is it's more exposed because people have been, every time someone's done a search or play a song on their favorite player or watched a movie, some sort of AI has been going on in the background, whether it's in delivery or design. But now the AIs are somehow coming out from behind the shield.

Chad Watt: You have a machine learning system that helps you identify students at risk. Tell me a little bit about that. Its OU Analyze I believe.

John Domingue: So, the word “open” in Open University means that we have no constraints on who can study with us. You do not need any qualifications. We have over 200,000 students, and I think a third of our students do not have the qualifications to study at a standard university. And that means that a number of them will struggle early on. So back in 2012, we built, OU Analyze, which is always tracking what the students do when they're interacting with our online learning environment, the VLE, the virtual learning environment, and then uses that data plus the demographic data of the students to predict is this student in danger of failing the next assignment? Is this student in danger of failing the course as well?

Now the great thing about OU Analyse is it's tied to university data. So we have 200,000 students. That means that on some of our larger courses, we have 3000 students plus looking at exactly the same materials over the same time period. You machine learn over last year's students and that provides your decision-making model for this year's students. And then every tutor, 4,000 tutors has access to this, and then the various signals that come up saying, oh, student X or student Y hasn't been engaging so much and they're in danger of dropping out. And then you can target your resources accordingly.

Chad Watt: That's fascinating. So, OU Analyse is 11 years old or so, can you give me a little bit of a description of how the capabilities have advanced, is it better now than it was in 2012?

John Domingue: So, we've done a lot of studies around this, so over 14,000 students, and we can predict that the second biggest predictor of success for an OU student is the extent to which your tutor is using OU Analyze. The first predictor is how well you've done in the past. There's on average a 7% increase in pass rates along this and it's 10% for students from an ethnic minority. So, the latest variations of this tool are in fact looking at equality and diversity issues. So, seeing are there specific elements of our course materials or assignments where particular categories of students struggle? And then relaying that back.

Chad Watt: When I hear university, I think of the manicure greens and the quads, but it's helpful to kind of reset again, Open University is massive open online, and your course catalog is similarly massive and similarly diverse.

John Domingue: Yeah, I think we have around 300 courses.

Chad Watt: 300 courses. And you're constantly creating new coursework.

John Domingue: Exactly.

Chad Watt: How do you do that in the human context and how do you see AI, generative AI, ChatGPT, and the like supplementing that effort?

John Domingue: The people that designed the OU were real geniuses and there were several innovations that they created. One of them was, and this is back in the seventies, remember, was the module team. So, the idea that a course isn't just an academic, sitting in their office writing on the back of an envelope. This is a team that will have academics, software engineers, graphic designers, program producers, now web designers, and they come together for a couple of years to make these high-quality materials. There are different instances. So, one place is in the creation of a new course. So, from the design, the outlines, the materials supporting the academic, the other is downstream. So maybe some of the roles of editors, grammar checkers, making sure everything's in the right voice, they might be taken over.

We adapt our materials to different contexts. So, we have a platform called Open Learn where we give away around 10 to 15% of our materials for three, we've had tens of millions of learners from there, imagine having a box that you say, here's my course, adapt it for two weeks instead of a full term. We're also thinking about instantaneous adaptation to the student context. So, if you think in a standard university setting, you need to personalize the learning. Everyone knows that. And you have, what's not personalized is the book. The book is the same for everyone. And then the lecture to 200 people online is the same for everyone. And then you personalize through your local tutor. But imagine if you could do that with the written material or even a video.

So, the student is reading something online and says, "I'm sorry, I just don't get this page. I don't get this concept. Explain it to me like I'm 14 years old," and you just hit a button and, presto, the new material comes out.

Or we know that with the GBT technologies, you can automatically generate scripts that could be fed to an AI avatar. So, say, "Okay, just explain this to me, maybe through my home speaker or maybe through a human avatar in different ways around that." So it's thinking of materials not as text anymore, but as data, data that can be massaged and re-massaged to be personalized to the student contact.

Chad Watt: This is data that is almost ready to use. And we've been talking kind of the possibilities and prospects but let me talk about the guardrails now a little bit. Everyone, I think who's used a generative AI has been first amazed. And then when they probe into their subject matter that they are truly deep in, they see how basic the information is, and sometimes these generative AIs will just deliver incorrect information. How do you put the guardrails in place as you go through this process?

John Domingue: I should say at the beginning, I am just like lots of people just impressed with the performance of these tools, but they do suffer from this flaw. What we are doing in the pilots we are running right now is we are limiting the answers to come from existing materials that we have. So, imagine that we have been around for 50 years, and we have 50 years’ worth of course materials, any answer you give will be scoped around materials which have been blessed by an academic. The other is we have just actually created a new tool, which I call Core GPT. So, we have a research group in my lab who have produced the automatic semantic indexing of all online research papers. So, from libraries. So, they have an index of three hundred million research papers as an index, and they've linked that to GPT.

So, then what happens is you say, "Okay, write me some material on quantum mechanics or an introduction to AI, and I want it to be up to date." So, it does a search, a semantic search in Core, finds the relevant materials, then you give that to GPT and then GPT creates a summary of those results with all of the references in between. There is an extra step because GBT is so clever, we use GBT to create the query. So, you have a human query, which is a paragraph of text. GPT understands our specific semantic query language to Core, so it writes a query for you. Then it summarizes the results. So again, there is just two checks there. One is you're limiting it to research papers, and then you have the references in between.

Chad Watt: In fairness to the AI, I should finish the cycle there. First, it's amazing, then it's a little bit disappointing, then you learn that it's about iterating, it's about revising your prompt using the tool to help revise the prompt to get at a better question. To ask your question better.

John Domingue: Exactly. And that can be automated. That whole area of what is called prompt engineering, and my colleagues told me that there are more and more automated tools coming online where you put in your poor prompt, some machine improves your prompt, and then that goes to the system.

Chad Watt: Let's talk a little bit about academic accreditations, certifications. This really comes back to the mission of Open University. You guys are also an advocate of using distributed ledgers to track accreditations and certifications for adult education, which I mean, at the end of the day, what is education for? From the student perspective, they are seeking a better job, seeking an opportunity for a new job, that sort of thing. What are you guys doing there and what makes it better than the incumbent ways of doing these things?

John Domingue: Okay, great question. So again, I have to step back a little bit. So we believe in lifelong learning. So learning does not finish when you are 21 or 24. And also where you start in life shouldn't determine where you end up. Now, if you believe in lifelong learning, that means that you are not going to be learning at one institution.

You're going to be learning in many contexts, from school, specific university, maybe at work, maybe on these MOOC, massive open online course platforms, and you're going to be accumulating credentials from all of them, hopefully. So then, where are you going to store them? Are you going to print them all out and put them up in your loft and then forget where they are? People ask me, do you really have a Ph.D.? And then I prove it. I'm not sure if I really could now.

So then you need to store that. So then the best place to store this is in some place that's controlled by the student. So we really think about empowering the students to be self-sovereign with respect to their credentials because maybe you don't want to share all your credentials with everyone all the time. So if you imagine a place where students are in charge, maybe they store it on their phone, maybe they store it in their favorite place, then we need a way to verify those. And again, how are you going to verify across somebody's whole life? Then once you start writing out those requirements, a distributed ledger, a blockchain seems like an obvious choice, where you have locally controlled credentials and you have a place where different education institutions can sign a regional or a national or international ledger saying, yes, John really did get a Ph.D. Yes, he really did get his swimming certificate and they can all be put together.

So we've been working on a variety of projects supporting us, some funded by the European Union, one larger one, the Institute of Coding, which had a £40 million budget and was launched by a Prime Minister in the UK. So once you think about this, you can then think about different forms of credentials. Another thing you learn is fraud is very big. So there are people at big institutions that have been called for fraud, I have vice presidents of large companies, even prime ministers of countries that I won't name that have shut down their alma mater, so no one can have any queries. Also, there's the cost. So I know students, for example, who've come from Asia and they come to Europe and people don't believe that the university is a real university, and many weeks of effort goes into proving that, and then they move somewhere else, and that has to be replicated. I know for example in the UK, if a nurse moves from one hospital in the UK to another hospital, nobody trusts their credential, which is a sad state of affairs.

Chad Watt: That's interesting, and I'm thinking about moving across from one continent to the next. If you showed me a credential from Japan, I would have to take it on faith or-

John Domingue: Exactly. So then you could have, maybe somebody else in your country has taken a student from Japan and they said, yes, not just for this particular student, but for this whole class of accreditation. I can verify that. So that you're not reinventing that accreditation verification wheel every time.

Chad Watt: That's very fascinating. If you could hop back in time for five minutes, what's the one piece of education advice you would offer up to 18-year-old John?

John Domingue: Today is the slowest that technology will progress for the rest of your life. Be confident, you are really going to make it and take care of your network. The size and quality of your network will be important to you.

Chad Watt: I think that's useful to 18-year-old John and 50-year-old Chad. Thank you, John, very much for your time and insights. This podcast is produced by the Infosys Knowledge Institute as part of our collaboration with MIT Tech Review in partnership with Infosys Cobalt. Visit our content hub on technologyreview.com to learn more about how businesses across the globe are moving from cloud chaos to cloud clarity. Be sure to follow Ahead in the Cloud wherever you get your podcasts. You can find more details and our show notes and transcripts at infosys.com/IKI. Thanks to our producers, Catherine Burdette, Christine Calhoun, and Yulia De Bari. Dode Bigley is our audio technician. And I'm Chad Watt with the Infosys Knowledge Institute. Until next time, keep learning and keep sharing.