Hariri Institute Director Yannis Paschalidis on “The Role of GenAI in Research”
Hariri Institute Director and Distinguished Professor of Engineering Yannis Paschalidis joined host Peggy Smedley of The Peggy Smedley Show on July 17th to discuss The Role of GenAI in Research.
They discussed:
- How GenAI impacts the world—particularly research and development, education, and business.
- Categories of GenAI systems that are out there.
- Some of the key concerns and negative impacts of GenAI and some of the potential benefits of GenAI
Hear the episode: https://peggysmedleyshow.com/the-role-of-gen-ai-in-research
View the transcription below:
Podcast on The Role of Gen AI:
Air Date: July 17, 2024
The Role of GenAI in Research Transcript
Peggy Smedley:
Hello listeners and welcome back to the Peggy Smedley Show. I’m your host Peggy Smedley. My next guest is here to share how Generative AI is impacting research and development education and businesses. He will also break down some of the key concerns and negative impacts of AI as well as share some of the potential benefits it can offer research specifically. Please welcome Yannis Paschalidis, Director of the Hariri Institute for Computing and Computational Science & Engineering, and Distinguished Professor of Engineering at Boston University. Yannis welcome to the show.
Yannis Paschalidis:
Very glad to join you.
Peggy Smedley:
Yannis, it’s exciting to have you on the show, I appreciate you spending time with me today, so I thought maybe if you wouldn’t mind, I’d love for you to give our listeners a little bit of background on your role at Boston University and share a little bit of some of the research that you’ve been doing.
Yannis Paschalidis:
Sure, definitely I have been at BU for quite some time. I’m a professor in the College of Engineering in the Department of Electrical and Computer Engineering and I also have joint appointments in Biomedical Engineering, Systems Engineering and in a new academic unit we have in Faculty of Computing and Data Sciences, and I’m also the Director of the Hariri Institute. As you mentioned in the introduction, the Hariri Institute is a university-wide research center, and the role of the Institute is to bring together faculty that may come from very different disciplines, working in the general area of computing, computational science and engineering, with AI being a big part of that, particularly these days.
Peggy Smedley:
Yannis, you don’t know how to relax do you, very busy, I’ll say that. So how is AI and all the things that you’re doing impacting AI right now. I know that you’re focused on so many things in the science side of things, and you have to look at right now saying AI, generative AI, is impacting our world, particularly, I would imagine you’re looking at it, how it’s impacting research and development, education businesses, it’s got to be impacting all the things that you’re studying and what you’re teaching today.
Yannis Paschalidis:
We live fascinating times. We can call it the AI revolution, if you want, Generative AI obviously has come to the forefront in the last few years. There are many uses, particularly in research and education. We can discuss a bit later, the many different uses of generative AI in research. Just for starters, it can be extremely helpful in processing data in generating new data in literature review in many different aspects of research, but it can also be extremely useful in software development, in preparing data to be digested by software, in many business functions, dealing with documents, extracting information, producing manuals, even taking notes from meetings and presentations. And of course, it has many different uses in education. Perhaps the one I would highlight is tutoring, so people can adjust these chat boxes so that they can help them with studying a new topic and getting review questions, preparing for exams, and many different things, that in the past we would rely on live tutors to do.
Peggy Smedley:
So, as I hear you talk about this, while there’s positives, my first reaction to some of this is the data and looking at it from software development, how it’s impacting us both positively and negatively, with Generative AI because what I have to think is, are we getting accurate information? I think right now, we’re trying to look at all this data and everything is generating data and as you talk about software development, are we getting the right information? What exactly is happening?
Yannis Paschalidis:
So, the systems that are available today are not 100% accurate. The Generative AI chat boxes out there have a tendency to hallucinate and provide information that is false. I think as these systems and models become better and better, the rate of hallucination will go down, but beyond that, I think that there are a number of limitations, a number of concerns. One major concern with Generative AI is privacy and security. You can think about the Generative AI systems that are available out there and put them in roughly three categories. One is the systems that are owned by a specific company, and where actually even the interaction between the user and the system takes place on a platform that is owned by the company. So, whatever you upload in that platform becomes the company property and the data can actually be used in order to train the next generation of the systems. So there you have significant privacy concerns because if you upload sensitive information, that information may become public through the next generation of the chat box and of the Generative AI system. Now you have other systems where I would call them semi open or semi closed, depending on whether you see the glass half empty or half full. In these types of models, these types of systems, you can actually download the code, you can run the generative AI algorithm on your own hardware, you own the interactions, whatever you upload remains at your local server. However, they are not completely open because we don’t know exactly the type of data that has been used in order to train the system, so there is this opacity, if you will. Then there are other systems that are now emerging, particularly more from academia, that are completely open in the sense that the data sets, a way of the model has been trained, are available and even the system is available, the code is available, you can download it, you can run it locally, and you own every interaction that takes place with that system.
Peggy Smedley:
We’re all getting excited about the opportunities. We talked so much about open AI and people say, oh, I want to develop things and create my own AI. Do companies and individuals understand or do businesses understand better versus then individuals? These are the three types of things you just discussed, but when we think about it, do we really understand peeling back the onion, like you just did? Do individuals and companies really know the fine line that you just described?
Yannis Paschalidis:
Perhaps not everyone, and I think that part of our role is to educate. Even the university setting, it’s not that every faculty is involved in this type of research. In fact, one of the things that the BU Provost did this past year was set up a task force. I was a member of that task force and co-chair of the task force. We were asked to produce a report that in fact highlighted what are some of the benefits and what are some of the limitations of Generative AI. One of the main goals of the report of the task force was to educate the faculty and the students so that they are aware when they go out and they use these tools. I think the general public is probably less aware; companies that are using these tools probably are becoming more and more aware of some of the limitations. We can see from the news reports some instances of individuals who were not aware. For example, there’re known cases of lawyers that used Generative AI in order to produce court briefs; they didn’t check the output and there were court cases that were not existent and were mentioned in the court briefs. That’s one example of potential lack of understanding of how these tools work.
Peggy Smedley:
Isn’t that the biggest concern we have to think about with Generative AI? That companies have individuals who haven’t really, and I use this loosely but making the point, done their homework to really understand, because that’s the concerns of the negative impacts. Then there’s also the positive benefits you just described in doing the research and again, I’ll use the analogy of doing our homework, you really need to understand it because the positive outcomes of really understanding where Generative AI can take us and the data on really understanding, is where we started our conversation, software development, the data, really understanding how we can use it as a tool to develop new things, the sky’s the limit?
Yannis Paschalidis:
That’s correct, indeed the sky’s the limit. I think that the biggest danger is not from lack of understanding, I think our understanding and the society’s understanding will evolve over time and will improve over time. I think the even bigger danger is that there probably are going to be people out there that will use it for the wrong purposes. Like g Generative AI has the ability to turbocharge our ability to do research and to be more effective in our daily lives and in all of the things that we do for our work, and even for entertainment, at the same time, it has the ability to turbocharge malicious uses. Think about misinformation campaigns that can be turbocharged by Generative AI, think about deep fakes, and how they are used in a variety of settings, from high schools to even politics. I think that is the larger danger out there.
Peggy Smedley:
That raises the bigger concern, because when do you stop it? We’ll go back to your security, privacy and hallucinations to even something more serious, even younger kids are being bullied, or deep fakes, as you described, or political campaigns – we’re running into a big election right now in six months. I think when we look at all of these things, it can happen after the fact and it’s too late, the damage is done, there’s not a whole lot we can do. How do we stop this before the damage is done? And what do we do? Do we have to start thinking about these concerns as they affect industries like medical, something so severe, our infrastructure, even more than hundreds of millions of people are impacted by these things that you just described? And now, we can’t reverse it, the damage is so severe, we really need to look at it now.
Yannis Paschalidis:
Yeah, I think we should be concerned, but at the same time I think there is no way to put the genie back in the bottle, so to speak. BU’s stance on this, based on the report of the AI task force I mentioned earlier, is that this is a revolution, this is a leap in technology, it’s here to stay, and we need to critically embrace it. So, embrace it for all of the positive uses and impact that it can have on research and education, on leap frogging in scientific advancements, but at the same time, educate the users and understand the implications and use it with caution. Rather than prohibiting use, which is unlikely to be very effective, a more effective strategy is: campaigns to educate the public, and even have some government regulation in place so that there are proper guardrails that companies that offer these tools need to use, and proper auditing verification mechanisms for appropriate bodies to be able to look at these products and to understand what are some of the dangers and the limitations.
Peggy Smedley:
Are we at the point of looking at AI as severely as we look at privacy and security as two separate entities, or will these now have to be considered as one? We have really strong security professionals who have to look at our security and our firewalls every day in the perimeters and make sure the nefarious characters, those bad guys, aren’t trying to penetrate our systems. Do we have to look at AI the same way now, or will these merge one because when I hear you talk like this and saying, you have to put guardrails and auditing, that’s the same kind of the talk that we use to protect our systems every day in a connected world environment from the bad guys trying to do bad things. Are we saying the same thing is going to have to be we have to educate our students to make sure at every level that they have the right knowledge so these hallucinations, you know, that we’re eliminating that there, that they have the right knowledge to protect what we do as what the bad guys are doing, so it’s an equal playing field as we get smarter and smarter out there.
Yannis Paschalidis:
I think what we can say is that the security challenges increase, but I’m not sure I would call these necessarily new fields for computer security. In many ways, it introduces some new challenges, and there are security professionals out there, I’m sure they are looking into what Generative AI can do and in what way. For instance, it can, maybe, facilitate hackers to develop new methods to attack systems. I think that people are thinking about this or looking into this, so I wouldn’t say that we have immensely new security challenges, but we have definitely new challenges that are being generated by Generative AI.
Peggy Smedley:
Let’s look more on the positive side of this; let’s flip the coin a little bit on this and talk about how the role of universities and industry in the development of GenAI and their role now and into the future. The way I hear you talk is, while we do have to watch about what happens with the negative side of this, and we are paying attention, and we’re keeping a closer eye on it, there’s really great opportunities. You talked about tutoring and education and being able to do more things with students literally at every level. What excites you most about this?
Yannis Paschalidis:
Generative AI is a tool that can be leveraged to greatly improve the way we teach some of the classes. It offers a way for students to comprehend the material better. Imagine years ago, if you had the question, even before the time of Google, if you had the question, you had to go to the library and spend hours finding the relevant information, study the relevant information. Then Google came about and search engines, and you could easily search and find the relevant information, get the right paper, even get the right paragraph from the papers or now you have the ability to go into some of these Generative AI tools and ask a very specific question and get an answer, and of course, you have to verify that is the correct answer and there is no hallucination, but at least it gives you pointers to information that otherwise was much more cumbersome to find. So that’s a big advantage. You can now change the way some of the courses are taught, you can flip the classroom, as we say in academia, you can have students prepare in advance for the material, and then come to the lectures and come to class in order to ask questions and to help them understand the material in greater depth. There are possibilities to design specialized tutors so that students that have trouble with some specific material, they can design a tutor that goes over questions for that specific section of the book or the class and they, in the process, become better and they understand better that material that they may have missed. You can have Generative AI tools be helpful in terms of preparing a large course project, even do some design for you, generate some figures. Of course, you have to be careful not to over rely, and so, the way that we give assignments to students that may have to change so that students don’t over rely on Generative AI, but there are many different positive ways in which the tools can be used today to enhance education.
Peggy Smedley:
Do you see that this will open up the university’s capabilities for students who can see new ideas of coursework that they didn’t think that they were able to study in the past?
Yannis Paschalidis:
I don’t see it necessarily as a replacement to university education or any other form of education. I think that the in-person education is hard to replace, right? It’s not just the raw presentation of material that you learn, it’s the interactions with the faculties, the interactions with your fellow students, and there is a huge value in these interactions and living and being in a university environment where you benefit from the structure, from the setting, from many opportunities that arise in that setting, but something that potentially Generative AI can do is help students prepare better for this type of experience and you can have students that may be missing certain aspects and they may want to register, let’s say in a more advanced course, they could spend their summer interacting with Generative AI tutorials in order to beef up on certain aspects, and be more ready to take the class and fully engage in the in-person educational experience.
Peggy Smedley:
Do you think we’ll need government to take a greater role in how Generative AI advances or we need less government involvement?
Yannis Paschalidis:
I don’t see the role of government as micromanaging the way the systems are evolving and micromanaging academia and what the industry is doing, but I think that there is a need for some regulation and for some sort of higher-level guidance, in terms of some very important questions that have to do with data provenance, with copyrights, with patent policy, with who owns what with whether a certain company, for instance, can use all of the data that is available on the internet and even copyrighted material and books in order to train the system, and also maybe some processes in place for some auditing and verification on how these tools have been trained on how they have been tested, so that there is at least a bit more transparency about some of the limitations and better understanding of what the limitations and what the drawbacks may be and what may be some of the dangers that lie ahead.
Peggy Smedley:
You made the comment that the genie is out of the bottle, are they going to ever be able to go back with some of the copyrights and some of the content that’s already been used and things like that? It’s going to be a very fine line to be able to do that and already we see some of that, is it going to be somewhat challenging?
Yannis Paschalidis:
You can always train a new model in a way that doesn’t use copyrighted information. I think that may depend on what government does and what may be some of the laws and regulations that are put in place, but will also probably depend on court cases. There are court cases pending and lawsuits that have been filed to try to understand at least whether copyrighted material was used and to what extent.
Peggy Smedley:
In going forward, we’ve seen AI exponentially grow in the last two years publicly, we’ve seen what’s happened, how do you see the growth in the next couple of years? Do you have some idea of what your expectations are? Do you see it going the same way, slowing down a little bit or just continuing to blossom at just unbelievable growth rates?
Yannis Paschalidis:
I think that the growth will be there. It’s a bit unclear what the trajectory is going to be. Multi-model AI has the ability to handle at the same time, images, videos, and text – so not only text – that is definitely an exciting development and that is happening. I’m not so sure if we’re going to see much bigger models and the models that are the tentative AI models that are already out there, we may see more personalization. I think we may see more tools that would enable even individuals who are not doing research in this space, to train their own Generative AI tool based on some existing foundational model and adjust that Generative AI tool to their specific needs and they wouldn’t be businesses, they would be universities, they would be research groups that are training and fine tuning these Generative AI models in order to make them better at the specific use that they are interested in.
Peggy Smedley:
Have you seen anything so far in Generative AI that you really impressed you or is there something that really has surprised you?
Yannis Paschalidis:
The ability to process text and interact has been quite surprising, and even the researchers from academia and from industry that were involved in that work probably were surprised at the ability of these models to have an extensive conversation with some human. Maybe two important elements in this where there was a lot of use of so called “human feedback” so those models that were originally produced may have not had the same ability to converse and there were humans that were involved in the training of these models and providing feedback that was used to train better models. The other, I think, interesting aspect is that there was a scale effect, so the models that were originally developed were not that big, and they were not that good. And after a certain point, in terms of the scale of the model, there were emerging properties that nobody was anticipating, and the quality of the ability of the model to converse and enter into a discussion improved drastically after a certain level of scale, and that was a surprising feature. To me another interesting and quite surprising feature is their ability to generate art and draw images based on a text description.
Peggy Smedley:
I have to tell you Yannis, this has really been a wonderful discussion. I’ve learned a lot from you. So, Yannis Paschalidis, Director of the Hariri Institute for Computing and Computational Science & Engineering and Distinguished Professor of Engineering at Boston University. Thank you for your time today. Where can our listeners go to get more information on what we just talked about?
Yannis Paschalidis:
I think an interesting resource may be the report that we produced at Boston University, the BU AI Task Force. And if you look for the Hariri Institute Website, there is a page that talks about the work of the BU AI Task Force and there is a link to the report which we think is quite informative.
Peggy Smedley:
Again, thank you for your time today.
Yannis Paschalidis:
Thank you.