00:44:53 · Kevin B. Boyd
[00:44:53] S1: Okay, so you got a lot of questions in there. So let me try and tackle them sort of one by one. So if you think about AI more broadly across the university, like I can answer, you know, what we're doing with our central systems. But you know, as I say, we're very distributed. So there's a lot of other technology that available within, in particular research, teaching as well. So, you know, as I've said, like, we've gone big with the Anthropic, making that available to everyone. Previously, as you mentioned, we had built our own own AI tool. It was called Phoenix AI. It was using the OpenAI models, but we put our own front end on it to, at the time, provide the level of security, privacy, and digital accessibility that we wanted. But the reality is that we also have every one of the frontier models is in use in different ways, whether that's Google's Gemini, OpenAI's ChatGPT, the Meta models, and then open source models as well, or open weight models. It really depends on the application. And there are some things where it's high sensitivity. Like, I'll use my information security team as an example. They tend to use more of the open weight models running on prem because they really don't want the highly sensitive security data going to the cloud. There are other things where, in some cases, faculty may just be experimenting with different models. In other cases, are particular models that are better for particular disciplines. And I think that's something we'll increasingly see in the future, where, you know, if you think of the frontier models, many of them are... The corpus of information is very broad. But what's really useful in, let's say, biomedical sciences and the legal field and some of these other fields are models that have been... Where the corpus of information has been more carefully curated. The legal field is a great example of this. You don't necessarily want a model providing answers to an attorney that has, I don't know, you know, some random information from the internet that... Because the corpus of knowledge came from Wikipedia or something like that. So I think that's true. Now, you had a question about students and homework. You know, that's a challenging question. And we, in so many ways, we don't view that as a technology problem. We view that as a people challenge.
00:47:48 · Kevin B. Boyd, Grace Gong
You know, faculty are being encouraged to be really clear with students about what the expectation is for how they can use AI in the class. What's appropriate for them to use and what's not. The other thing I'll say is that universities are being really forced to rethink how they do evaluation. So, you know, prior to AI, many courses used papers. You'd go home, you'd write a paper, you know, you'd turn in a 25 page paper. In the world of AI, it doesn't work so well. Because as much as you might say, please don't use AI to write the paper, you know that some number of people are going to do that. And so what we're seeing is a shift to more in class evaluation, whether that's tests or quizzes, and doing it in a way that AI isn't available to individuals. So you actually determine, have you mastered the subject? [00:48:51] S0: What's your thought on open weight model? Since nowadays people are kind of fear of having their data to be put into some of these like large labs and like everybody's like doing like open weight model these days. And do you see the same trends apply to the university or it's like not necessarily the biggest concern? [00:49:15] S1: I think it's not necessarily the biggest concern. I think there is a level of concern about data security about what the source of the open weight models are. And, you know, I think that is a consideration, but there's no question that the open weight models have place to play in research.
00:49:15 · Kevin B. Boyd, Grace Gong
[00:49:15] S1: I think it's not necessarily the biggest concern. I think there is a level of concern about data security about what the source of the open weight models are. And, you know, I think that is a consideration, but there's no question that the open weight models have place to play in research. [00:49:38] S0: Joly Betty is asking like, one of our audience is asking like, what's the biggest security challenge you face as more research teams started using AI?
00:49:48 · Kevin B. Boyd
[00:49:48] S1: Yeah. That's an interesting question. So, you know, as I said, I think that AI can be a tremendous benefit to research. But we also want to be mindful of the fact that research often happens with sensitive data. So, you know, as I said, we're a large biological sciences research university, and there's often a lot of HIPAA data in there. And so we have some pretty important guidance to faculty about what they can and can't do with that type of sensitive data. We have a separate instance of Claude that is okay for HIPAA sensitive data, where you don't have, you know, it doesn't preserve the history and some other things like that. So I think, you know, with the data security concerns, it's having clear guidance for sensitive data. And I think we still want to have the tools available for researchers that are working with the sensitive data. But you know, the other thing that we talk a lot about is we can have really good contracts with the vendors about what is okay and what is not okay. But we also know that this, the AI space is moving really rapidly. And sometimes the models do things that they weren't intended to do. You know, we've all seen what's happened in the last few months with Anthropic, with OpenAI and others where the models have, essentially escaped the guardrails and done things they really weren't intended to do. You can imagine someone using, let's say, a coworker code to do research and the, you know, a particularly capable model decides that the right answer for how to achieve the prompt that it was given was to send the data to someone else who's an expert and get help from them. There's actually been at least a few cases of that happening. That maybe doesn't matter so much if the data that you're working with is astronomy data. It matters a whole lot if it's patient data.
00:52:17 · Grace Gong
[00:52:17] S0: For sure. One of the audience, Christopher, is asking, so you guys are protecting learn to think first step in class, which I love. But out in the real world, there's a lot of entry level jobs, the ones where people used to learn the work by doing the ground stuff is getting automated. Are you seeing that hits your grads when they go looking for a first job? And is this about AIP starting to change how you think about teach to fill the gap?
00:52:48 · Kevin B. Boyd
[00:52:48] S1: I think that's a great question. It's one that we've talked a lot about with many different groups. I think that, you know, there's been a lot of fear in the last couple of years about AI taking jobs away. And, you know, I tend to think that like any, you know, major transition, there are jobs that will change and jobs that will evolve. And as that happens, there probably will be some jobs that are no longer needed in the future. What my own belief is, is that the entry level jobs are going to be the ones that are affected the first. You know, it's things that you used to have a, you know, a junior associate in the law firm or, you know, an entry level communicator drafting things. And in the future, think that's going be done by AI. The really difficult thing is that when you're in one of those entry level roles, that's how you learn how to progress to the mid level role. Know, you're you're off... You're doing that grunt work for the first few years. It might not be the most fun, but but you really learn that particular profession, whether that's a legal profession or being a communicator. I do think, you know, that there's going to be a challenge of, okay, how do you get from being a very successful student, you've gotten your degree and graduated, if there are fewer of those entry level jobs in the future, how do you get the experience you need to get to the mid level jobs? I don't know that I have an answer for that. I think, you know, the individual who asked the question is like, okay, does that change education? I think it's maybe too early to know that, but I think it may have to.