00:23:27 · Grace Gong, Kevin B. Boyd
[00:23:27] S0: Speaking of that, so like if you're building like a AI assistant that searches like the university documents, how do you make sure it's only like retrieve the information that's particularly, you know, that the user is like allowed to see? [00:23:41] S1: Yeah, that's a great question. We've been having lots of conversations about that. And, you know, I think... I guess I would give you a couple different answers. So one of the things that is possible with Claude and with other tools like that is through the use of connectors, we can connect to, you know, various other... Or data storage, whether that's Box or OneDrive or Google Drive and things like that. I think generally those connectors are built in a... In such a [00:24:19] S0: way
00:24:19 · Grace Gong, Kevin B. Boyd
[00:24:19] S0: way [00:24:19] S1: that they're pretty good about people only being able to see what is available to them as a user. But I think one of the problems that exists is that often people have been given access to more than maybe what they actually need. And that hasn't been all that apparent in the past when they actually had to search through and scan through and look at things in some of these file storage systems. It becomes harder when all of a sudden the AI can see everything. And you know, if you ask it a question, it's going to pull back an answer based on everything that is available to it. And that's something where with many of these tools, we're trying to make sure that there isn't data oversharing, and that in tools where you have sensitive information, we're being intentional about maybe not making that sensitive information available to the AI. There's there's different ways to do that. So I think that's that's more on the... Some of these file sharing tools that are incredibly useful with AI. Then on the other side, when we're building agents, whether that's in finance or HR or others, so much of that is making sure that we're designing it, testing it, and really limiting access and putting appropriate guardrails in place so that we know exactly what it does have access to, what it doesn't have access to, and and we're and getting some warning if it seems to be exceeding guardrails in some [00:26:02] S0: way. When an AI is answering something wrong, so how do you figure out whether the problem was the motto or the retrieve document or the prompt?
00:26:02 · Grace Gong, Kevin B. Boyd
[00:26:02] S0: way. When an AI is answering something wrong, so how do you figure out whether the problem was the motto or the retrieve document or the prompt? [00:26:16] S1: So you're talking about less agentic in that question and more someone is interacting interactively with the tool? Is that what you mean? [00:26:28] S0: I would say, yeah, like more of like interacting since, you know, I feel like if you are building out, let's say like in the, yeah, I guess like also in the school context or like in whether it's like, you know, yeah, like, maybe, like, a faculty or, like, people are trying to pull diff... Different level of, like, information from the system.
00:26:28 · Grace Gong, Kevin B. Boyd
[00:26:28] S0: I would say, yeah, like more of like interacting since, you know, I feel like if you are building out, let's say like in the, yeah, I guess like also in the school context or like in whether it's like, you know, yeah, like, maybe, like, a faculty or, like, people are trying to pull diff... Different level of, like, information from the system. [00:26:49] S1: I mean, I'll be honest with... Like, we don't spend too much time helping people, like, with individual queries and things like that. We are spending a good amount of time helping people to understand how to use AI effectively. And that's true both with the staff, so the administrators across the university, as well as faculty and students. And so I think there's an element of many people when they start to use these tools, they think, I can just ask a simple question like what I do with Google, you know, or some other search engine and I'll get a response. And unfortunately, when you do that a lot of times, it's garbage in garbage out, you know. You know, what I think is really helpful for people to understand is the more context you provide, the better the answers that you get. And that's I think a key thing that we're really trying to help people understand, is how do you provide enough context, develop the harnesses, to get the results that you're looking for, whether that's something you're doing with an interactive chat or something you're doing with a tool like code.
00:26:49 · Kevin B. Boyd, Grace Gong
[00:26:49] S1: I mean, I'll be honest with... Like, we don't spend too much time helping people, like, with individual queries and things like that. We are spending a good amount of time helping people to understand how to use AI effectively. And that's true both with the staff, so the administrators across the university, as well as faculty and students. And so I think there's an element of many people when they start to use these tools, they think, I can just ask a simple question like what I do with Google, you know, or some other search engine and I'll get a response. And unfortunately, when you do that a lot of times, it's garbage in garbage out, you know. You know, what I think is really helpful for people to understand is the more context you provide, the better the answers that you get. And that's I think a key thing that we're really trying to help people understand, is how do you provide enough context, develop the harnesses, to get the results that you're looking for, whether that's something you're doing with an interactive chat or something you're doing with a tool like code. [00:28:03] S0: So how do you defend... So, like, going back to the security aspect, so how do you defend against prompt injection when something AI reads, try to trick into ignoring it's an instruction or exposing data?
00:28:03 · Grace Gong, Kevin B. Boyd
[00:28:03] S0: So how do you defend... So, like, going back to the security aspect, so how do you defend against prompt injection when something AI reads, try to trick into ignoring it's an instruction or exposing data? [00:28:23] S1: Yeah, that's definitely a problem. I don't know I the answer for that. You know, I think with where we are providing a chatbot, let's say a tool that allows someone to ask questions about benefits. You know, so much of that is us doing enough testing to make sure that prompt injection isn't possible, that we're really clear on what happens if someone is asking something that exceeds the bounds of whatever that chatbot has been built for? [00:28:55] S0: I guess like when we also talk about the research that the UChicago people are doing, What are something that you find most interesting when you are like looking at the researches?