Episode Description

“Last week was a great year in GenAI,” jokes Mark Ramsey—and it’s a great philosophy to have as LLM tools especially continue to evolve at such a rapid rate. This week, you’ll get to hear my fun and insightful chat with Mark from Ramsey International about the world of large language models (LLMs) and how we make useful UXs out of them in the enterprise.

Mark shared some fascinating insights about using a company’s website information (data) as a place to pilot a LLM project, avoiding privacy landmines, and how re-ranking of models leads to better LLM response accuracy. We also talked about the importance of real human testing to ensure LLM chatbots and AI tools truly delight users. From amusing anecdotes about the spinning beach ball on macOS to envisioning a future where AI-driven chat interfaces outshine traditional BI tools, this episode is packed with forward-looking ideas and a touch of humor.

Highlights/ Skip to:

• (0:50) Why is the world of GenAI evolving so fast?
• (4:20) How Mark thinks about UX in an LLM application
• (8:11) How Mark defines “Specialized GenAI?”
• (12:42) Mark’s consulting work with GenAI / LLMs these days
• (17:29) How GenAI can help the healthcare industry
• (30:23) Uncovering users’ true feelings about LLM applications
• (35:02) Are UIs moving backwards as models progress forward?
• (40:53) How will GenAI impact data and analytics teams?
• (44:51) Will LLMs be able to consistently leverage RAG and produce proper SQL?
• (51:04) Where can find more from Mark and Ramsey International

Quotes from Today’s Episode

• “With [GenAI], we have a solution that we’ve built to try to help organizations, and build workflows. We have a workflow that we can run and ask the same question [to a variety of GenAI models] and see how similar the answers are. Depending on the complexity of the question, you can see a lot of variability between the models… [and] we can also run the same question against the different versions of the model and see how it’s improved. Folks want a human-like experience interacting with these models.. [and] if the model can start responding in just a few seconds, that gives you much more of a conversational type of experience.” - Mark Ramsey (2:38)
• “[People] don’t understand when you interact [with GenAI tools] and it brings tokens back in that streaming fashion, you’re actually seeing inside the brain of the model. Every token it produces is then displayed on the screen, and it gives you that typewriter experience back in the day. If someone has to wait, and all you’re seeing is a logo spinning, from a UX experience standpoint… people feel like the model is much faster if it just starts to produce those results in that streaming fashion. I think in a design, it’s extremely important to take advantage of that [...] as opposed to waiting to the end and delivering the results some models support that, and other models don’t.”- Mark Ramsey (4:35)
• "All of the data that’s on the website is public information. We’ve done work with several organizations on quickly taking the data that’s on their website, packaging it up into a vector database, and making that be the source for questions that their customers can ask. [Organizations] publish a lot of information on their websites, but people really struggle to get to it. We’ve seen a lot of interest in vectorizing website data, making it available, and having a chat interface for the customer. The customer can ask questions, and it will take them directly to the answer, and then they can use the website as the source information.” - Mark Ramsey (14:04)
• “I’m not skeptical at all. I’ve changed much of my [AI chatbot searches] to Perplexity, and I think it’s doing a pretty fantastic job overall in terms of quality. It’s returning an answer with citations, so you have a sense of where it’s sourcing the information from. I think it’s important from a user experience perspective. This is a replacement for broken search, as I really don’t want to read all the web pages and PDFs you have that *might* be about my chiropractic care query to answer my actual [healthcare] question.” - Brian O’Neill (19:22)

• “We’ve all had great experience with customer service, and we’ve all had situations where the customer service was quite poor, and we’re going to have that same thing as we begin to [release more] chatbots. We need to make sure we try to alleviate having those bad experiences, and have an exit. If someone is running into a situation where they’d rather talk to a live person, have that ability to route them to someone else. That’s why the robustness of the model is extremely important in the implementation… and right now, organizations like OpenAI and Anthropic are significantly better at that [human-like] experience.” - Mark Ramsey (23:46)
• "There’s two aspects of these models: the training aspect and then using the model to answer questions. I recommend to organizations to always augment their content and don’t just use the training data. You’ll still get that human-like experience that’s built into the model, but you’ll eliminate the hallucinations. If you have a model that has been set up correctly, you shouldn’t have to ask questions in a funky way to get answers.” - Mark Ramsey (39:11)
• “People need to understand GenAI is not a predictive algorithm. It is not able to run predictions, it struggles with some math, so that is not the focus for these models. What’s interesting is that you can use the model as a step to get you [the answers]. A lot of the models now support functions… when you ask a question about something that is in a database, it actually uses its knowledge about the schema of the database. It can build the query, run the query to get the data back, and then once it has the data, it can reformat the data into something that is a good response back." - Mark Ramsey (42:02)

Links

• Mark on LinkedIn
• Ramsey International
• Email: mark [at] ramsey.international
• Ramsey International's YouTube Channel

Transcript

Brian: Welcome back to Experiencing Data. This is Brian T. O’Neill. Today I have Mark Ramsey on the line from Ramsey International. How are you, Mark?

Mark: I’m doing great. How are you?

Brian: I’m doing really awesome. So, we met at CDOIQ recently. So, you gave a really interesting talk on specialized, large language model deployments in the enterprise. I think at some point you said your favorite saying was, “Last week was a great week in GenAI.” [laugh].

Mark: Yeah. I mean, I love the phrase, “Last week was a great year in generative AI,” because it’s moving so quickly that every day seems like a month, and every week seems like a year.

Brian: Yeah, yeah, yeah. I screwed it up. “Last week was a great year in GenAI.” So, we met three weeks ago. I mean, when was CDO? It was like three weeks ago. So—

Mark: Yeah, it was—yeah, a lot—

Brian: So, what’s changed in three weeks? Are we two solar systems away [laugh]?

Mark: Yeah, it’s like we’ve gone, like, three years. But if you think about it, that’s actually happened because, you know, Google has announced an upgraded version of their Gemini model that is significantly better than it was just a few weeks before.

Brian: Yeah, yeah. I’m curious, for you—so you’ve been a chief data officer at different places, now you’re in consulting, and it sounds like you’re doing a fair amount of work in these large language models inside enterprise, internal tooling, things like this. Is that correct?

Mark: Oh, yes.

Brian: Some customer-facing as well?

Mark: I mean, I’ve been in the data and analytics space for over 35 years now and a lot around the foundational type of work, like, the platform work that needs to be done. But, you know, in the last few years, organizations are really focused on how to leverage generative AI and drive, you know, return on investment.

Brian: Sure, sure. I’m curious, like, so you just said, you know, so Google just released yet another version of a foundation model that’s better. Do you have favorite personal testing criteria you use to go pound at these things on your own, just to get that initial feel, like: oh, this is better; oh, it’s about the same; it’s a little worse. Like, how do you go about that, just in your personal curiosity when these things drop? Or do you have, like, a rigorous thing you go through?

Mark: Well, I mean, it’s interesting because, I mean, rigorous, I don’t know about that, but, I mean, with generative, what we’ve done is we have a solution that we’ve built to try to help organizations, and one of the things we can do is build workflows. And so, we have a workflow that we can run and ask the same question of a whole variety of models. And so, we might ask the Claude model, the Gemini model, the OpenAI model, and so forth, and ask the same question and see how similar the answers are. And it is quite interesting that, depending on the complexity of the question, you can see a lot of variability between the models. And so, I wouldn’t call it an extensive testing, but just from experience in running those workflows, we can also run the same question against the different versions of the model and see how it’s improved.

A lot of times the performance of the model is really where we see a lot of improvements because folks want, sort of, a human-like experience interacting with these models, and so if you ask a question, and it takes 10 or 15 seconds, or 20 seconds, for the model to start to respond, that’s okay, but if the model can start responding in just a few seconds, that gives you much more of a conversation type of experience.

Brian: Yeah. I’m a fan of, and I assume this was intentional on the part of, you know, at least ChatGPT and OpenAI, that they seem to be streaming the response, which, from an interaction standpoint, is great. You get a few tokens back, or a few words at a time. It’s much better than waiting for a 2000 word response for 20 seconds. Same amount of content, different experience, you know?

Mark: Yeah.

Brian: And that’s what I wanted to talk to you today about, was, kind of, user experience considerations for these large language models. Because I don’t hear this talked about too much, and companies are throwing a lot of resources at these, so that’s kind of where I wanted to go today.

Mark: Yeah. No, I mean, I think one comment on the streaming, I mean, what folks don’t understand is that when you interact with, like, a ChatGPT, or you interact with the Claude AI interface, when it starts to bring those tokens back in that streaming fashion, I always tell folks that you’re actually seeing inside the brain of the model because every token that it produces is then displayed on the screen, and it gives you, kind of like that typewriter experience back in the day, when we used to have those. But it does make a huge difference, like you’re saying, because if someone has to wait, and all you’re seeing is a logo spinning, from an experience standpoint, what we’ve seen is people feel like the model is much faster if it just starts to produce those results in that streaming fashion. So, I think in a design, it’s extremely important to take advantage of that. But it is actually, as the model is releasing a token, it then gets displayed on the screen, as opposed to waiting to the end and delivering the results. And some models support that, and other models don’t.

Brian: Yeah, yeah. And you make a good—I mean, this is a fairly well known thing in the design world, but maybe not in the data world, that perceptions of speed and simplicity matter much more than actual simplicity. Because we used to hear like, oh, it should never be more than three clicks to get something done. This is completely false. The only thing that matters is someone’s perceived progress at doing a task.

So, it actually might be longer, but if they feel like they’re making advancements there, that perception is all that matters. And so, this is a difference between measuring something quantitatively, which is like, you know, you could throw a workload at it and then count the numbers, and it will tell you how many seconds it took. Very different than, like, putting this in front of a human being and asking them how they perceived it to be. So, I like that they’re streaming the token. I think that’s a really smart interaction.

Mark: Yeah, it’s very cool. And I think the other thing is, for an enterprise delivery of these kinds of solutions, there’s a number of things—I mean, the question is interesting, right? The pers—the consumer, the knowledge worker, types in a question, presses enter, there’s a number of things that really happen behind the scenes before the model even starts to answer the question. And I agree with you, it’s how that information is delivered back to the end-user so that they understand that progress is being made. So, if it takes five seconds before the actual question gets submitted, or if it takes even a second, helping them understand that there’s a few things happening before that information is provided to the model, and then as quickly as the model can reply, all of those things make a difference in how people really feel like this is a real time conversation, or it’s something that’s just really slow.

Brian: Yeah, we call that—I mean, the design lingo would be feedback, right? Is the system providing me feedback so that I know it’s—

Mark: Doing something [laugh].

Brian: —doing something, as opposed to the beach ball in macOS [laugh], you know?

Mark: Right. Right. For sure, for sure.

Brian: My favorite beach ball moment being, I don’t know if you ever watched Family Guy—

Mark: Yes.

Brian: —you know this cartoon? And there’s this—there’s an episode where the spinner just kept coming on, and so Peter’s like, “Oh, wait, wait, it’s doing something.” And they just killed minutes of air time, which I thought was awesome, just waiting for the loading [laugh] spinner in the cartoon.

Mark: Right, exactly. It’s like, oh, you think it’s doing—I mean, the one I like is the progress bar, where it—

Brian: It gets stuck [laugh]?

Mark: —goes across the screen, it gets all the way, like, 99%, and it does that in, like, 30 seconds, and then five minutes later, it finishes the last 1%. It’s like, well, yeah, this was really well-designed [laugh]. It’s crazy.

Brian: [laugh]. No, I understand. Talk to me about the specialized—you talk about your title—I think the title of your talk had the word ‘Specialized GenAI.’ And so, what are you trying to distinguish there, especially in, you know, the enterprise context?

Mark: Yeah, there’s really two approaches that organizations can use in taking advantage of large language models. One is referred to as retrieval augmented generation. Or a lot of times people say, oh, I’m using RAG. It’s basically when a question is asked, the information around that question is packaged up and provided to the model so that the model is actually generating the response based on that information or context that’s actually provided with the question. And so that, you can use, you know, the generalized models, and it reduces things like hallucinations where the models start making up answers.

A specialized model is kind of a different word for fine-tuning the model. So, you’re taking the base model, and then you’re training it on content so that it’s very knowledgeable about a very specific topic that you’re interested in. So, we chose to use the term ‘specialized model’ because we’re training the model to be very well versed in a very narrow domain, and it can answer questions in that domain extremely well. And what that does is it reduces hallucinations, but it also reduces the complexity of the model, and you get a lot more reliable responses when questions are being asked. So, those are, kind of, the two approaches. One is providing context when you ask the question, and that takes a little bit of time to package that up and submit it, and the other is actually fine-tuning the model into a specialized model so that it can answer the questions about that particular area very succinctly.

Brian: Got it. A little bit of a side thing here—and I don’t want to go too far into implementation because that’s not so much what this show is about—but during CDOIQ, I think following your talk, my—I don’t know if you know Juan Sequeda; he runs the Catalogs and Cocktails podcast, but he gave a very convincing talk. And again, I’m not as technical as you two are on knowledge graphs as being a way to highly improve the quality of responses that come back. Are you a fan of using those as well to improve RAG and make sure the system performs accurately?

Mark: I think there’s a number of ways that folks can, kind of, go at this. Our approach tends to be, you have to do a good job when you’re creating the content, and I think that’s one area where folks can focus. So, you know, the way to think about it is you have a large amount of unstructured data, like, you know, documents, and PowerPoints, and all of those kinds of things that contain the knowledge that you’re wanting to make available. That all has to be packaged up and stored in a vector database so that when you ask the question, you can retrieve the pieces that are relevant to the question. So, I think that is an area where the focus needs to happen.

But then also, once you’ve retrieved that information, there are models that do a better job at ranking the content to the question than even just the vector store. So, what we typically see is, you know, let’s say you have a thousand PDF files that you vectorize, and you put in a vector database. You ask a question that’s related to that content. It goes, and it pulls a hundred pieces of information about your question, then you can actually use a re-ranking model that looks at your question, looks at all the content, reprioritizes all of that. That’s where we’ve seen the biggest improvement.

Knowledge graphs can help you because they begin to understand the relationships between the data, and so there’s some ways that you can use the knowledge graph to improve what information is selected. But again, I think there’s no perfect answer. We specifically have not engaged the knowledge graph approach because we’ve found that it’s much faster to use vector searches and then re-ranking, versus heading down the path of a knowledge graph. Because, with a knowledge graph, it requires you to then go through the process of ingesting that information and organizing it into the relationships. Because, again, a knowledge graph, kind of, is a network diagram that breaks down the relationships of all of this. We’ve seen as good, or if not better, performance by doing the vector search and then re-rank.

Brian: Mm-hm. So, you’re in consulting. You worked in-house for several different—Nokia, I believe, and pharma as well. So, now you’re in consulting, and I know you’re doing work in this LLM space. Are you allowed to talk about—you don’t have to name the names of the companies, but what are some of the use cases for these LLMs that you’ve been a part of? And then maybe we’ll go into some of those a little bit.

Mark: Sure. One of the things right now is that—and you’ve probably heard this at CDOIQ—but a lot of organizations are in the prototype phase with large language models, right? So, they’re dabbling, some of the research that was done by the team, I think it was Randy Bean and Tom Davenport did research, and found that less than 5% of the projects around large language models are actually in production. So, 95% of the effort with organizations right now is still in prototype learning phase, and they’re really trying to understand how to drive value from using the large language models. I think the other thing that really has been a little bit of a challenge on deploying is that there are concerns about privacy, there’s concerns about how the models perform, and so forth.

So, one of the things we recommend to organizations is, start with something that’s already public. And so most, if not all, organizations have a website. All of the data that’s on the website is public information. So, we’ve done work with several organizations on quickly taking the data that’s on their website, packaging it up into a vector database, making that be the source for questions that their customers can ask, so instead of having—you mentioned three clicks to get to the information most—websites that I at least work with, it’s many more than three clicks to get to data within the organization. And what folks have also done is they have layered—so they have web pages connected to PDF documents, and all of this, so that it’s difficult to really find the information that’s out there.

So, they publish a lot of information on their website, but people really struggle to get to it. So, that is where we’ve seen a lot of interest is vectorizing the website data, making it available, having a chat interface for the customer, so the customer can ask questions, and it will take them directly to the answer, and then they can use the website as the source information. And this really, it gets around the challenge around privacy because all the data that’s on the website is already public, and it also gets people started and really helps customers and folks that are interacting with the website get to the data that they want to know about the particular organization. And it reduces—there are some organizations that have chat interfaces that actually go to live operators. That’s a pretty expensive way to service and answer questions from the customer, and so by actually using generative AI, they can reduce that expense and immediately drive value, as well as have a chance to actually check the models out.

So, that’s where we see a lot of organizations starting. I know there’s a lot of media around—like, you mentioned, pharma; I spent four or five years working within large pharma. There’s always a focus on automated drug discovery, and things like that. And there are benefits of using generative AI in that area. But really to get started, we recommend that organizations start with something that is quick to implement, they can get their feet wet, they can start getting value, and then begin to move it into other parts of the organization.

Brian: So, effectively though, if we’re deploying this against a website, we’re talking about fixing broken search. Are we not talking about fixing the local search for the company, like, effectively, the website search is not working.

Mark: Again, you know, this is a big thing now. You know, in the media, with OpenAI getting ready to release a search to compete against Google. People are comfortable with search, where they do a search, it gives them 2 million results, they look at the first page, they pick something. That isn’t answering a question, right? I mean, what we’re talking about is, use all of the data that is in the website to actually create a conversation with the customer, answer questions, don’t just give them a search, right?

Because if you pick any website, and if you do a search in Google, or you do a search on their site, it’s just a hit of a lot of word associations and things like that; you’re not really answering questions. Whereas with a large language model, the customer can actually ask questions and get answers. So, for example, we’re working with a healthcare company. We’ve ingested their entire website. A customer can go into the chat and say, “Is chiropractic care covered?” “Where are the hospitals located for this healthcare organization?” You know, they can actually ask questions. They don’t just do a search and get a list of website hits.

Brian: I can see the risk people already cringing when it says, “Yes, chiropractic care is covered by your plan,” and someone in the back has a hand up, like, “Yeah, but it also depends on which plan, and are you in Massachusetts, and which hospital network is it? And, by the way, what’s your deductible? And do you have secondary insurance? Are you on Medicare?” [laugh].

Mark: Like, but the thing you have to keep in mind, Brian, is that a lot of that stuff is on the website, right? So, that’s where it’s organized. So, for example, the way the model would answer that question is it would say, “Hey, it appears that it’s covered, but you should check with your plan and whatever,” because that information is also captured when they talk about the different characteristics and different services and so forth. So yeah, there is that risk, but that risk exists whether you’re using a call center operator that’s either onshore or offshore because they’re reading a script. And if you can feed all of that in, then the model can do that in a much more efficient job.

And I think also people because so many folks are interested in generative AI, we actually recommend the clients, on the button, it actually put chat with us via GenAI, or actually highlight the fact that they’re chatting with a large language model, so that the people can say, hey, this is cool. Now I—it’s, kind of, my ChatGPT for my healthcare company, or ChatGPT for my retailer, or ChatGPT for whatever the organization is, but it’s highly specialized to the content for that organization.

Brian: Sure, sure. I’m not skeptical at all. I’ve changed much of my search default to Perplexity, which I’m sure you’ve probably heard of before, but it’s competing with Google as well, and I think it’s doing a pretty fantastic job overall in terms of quality. Because it’s returning an answer, it’s not returning—it’s returning an answer with citations, which I really like is that you get an answer, and you get citations with it, so you have a sense of where it’s sourcing the information from, which is fantastic. And it’s not returning a list of web pages that might have the answer on it that I have to go do all the work for.

So, it’s interesting, but I think it’s important, like, from a user experience perspective, this is a replacement for broken search because no one really goes to an insurance, to Blue Cross or pick whatever insurance company, because I really want to see all the web pages and PDFs you have about chiropractic care.

Mark: Right.

Brian: Nobody says—nobody wakes up and says that. They have your question, which is this procedure covered or not? Like, I just need to know if it’s covered and if it’s not covered, then what do I have to do to get it covered? And this is great [laugh], it’s, I think, a great example of how to use that and answer real questions.

Mark: Right. And I think it’s also going to be a progression, right? So, like I said, if you start with the basic website information that can’t answer the question specifically for you because it’s general information, and then as the organization gets more comfortable with leveraging generative AI, the next step is to interface it with their—you know, so have the person sign into the website so now they’re a verified user in the system. Now, you have access to a lot of the information that is specific to them. You’ve got the entire policy document, which, again, nobody wants to dig through all of that to find out.

So, once you get that verification step, then you can begin to provide customized responses based on that particular person. So, there’s a lot of benefits, but I think right now organizations are still, you know, at the very early stages. And so, I think what I really advise organizations to do is get started, you know, start working with the technology, but do it in a protected way. And I think some organizations take that to mean that they’re just going to have an internal kind of a mechanism, so they use their internal information, and it’s only available to their employees. But again, that’s where you’re—you do run into some concerns about what information you’re actually including, and do you have the risk of disclosing sensitive information, and those kinds of things. So, it all has to be done within an architecture that supports all of that.

Brian: Sure. I’m curious, like, when you are working on these deployments, do you do any actual human testing of these to understand someone’s perceived experience using this kind of interface? Because I would venture to guess that most people do not have good experiences with chatbots, particularly for customer service. And while it may be more expensive—if you measure it in the short term, a finance person will say it’s really expensive to field these requests with human beings answering questions over the phone, but the quality, the brand impact can be significant. It’s just harder to measure the brand impact of a bad customer service experience. So, I’m wondering, do you ever sit people down and actually do testing with them to see if they feel like, hey, I’d rather talk to the chatbot at Blue Cross than talk to an agent on the phone because I’m getting better information, I’m getting it faster. Do you ever do that as part of your work, setting real users down?

Mark: Oh, yeah. I mean, I think it’s critical to do that, right? I mean, you have to do that in order—and we do it in a couple of ways. We do, like I mentioned, the workflow where we could ask the same question to a lot of different models. It’s important to create a portfolio of questions based on either live experience, right—so there’s a lot of call centers and a lot of customer service that they log the questions that come in and the interactions, so take those real experiences and boil it down to questions that you expect the consumer is actually going to ask, feed those into the models and see the response, and you know, look at those and see, is this something that would represent the kind of answer that you’re looking for, and then make adjustments so that you’re getting that type of an experience.

And I would say, you know, we’ve all had great experience with customer service, and we’ve all had situations where the experience with customer service was quite poor, and we’re going to have that same thing as we begin to ramp out the chatbots, is that we just need to make sure that we try to alleviate having those bad experiences, and also have an exit where, if someone is running into a situation where they’d rather talk to a live person, have that ability to route them to someone else. We’ve all, kind of, gone through these technologies. If you think back 15 or 20 years ago, when voice response units, and we were all pressing numbers and saying account numbers and converting it, and not working, and, “Say your phone number.” “Is that correct?” [unintelligible 00:24:32].

So, we’re trying to avoid those kinds of experiences with the chat and really make it a human-like experience. And that’s why the robustness of the model is extremely important in the implementation. And so, people—I think there’s a lot of discussions around local models versus the more cloud models, like an OpenAI or an Anthropic, compared to a Llama or a Mistral model. And over time, there will be parity amongst a number of these models, but it’s going to take quite some time. And right now, organizations like OpenAI and Anthropic, they are significantly better at that experience, and that human-like experience back and forth.

Some of the other models are local models, and they’re getting better, but their level of human interaction is not at the same level as, like, an OpenAI, or a Gemini, or an Anthropic. But I think it’s critically important that when you design something like this, that you actually have—and it would be the same as if you’re delivering any, kind of, solution for a customer—is do acceptance testing, do checking, see if it actually delivers the experience that you’re wanting. You mentioned citations or sources; that’s something that we’ve found that if we can provide the person with a source link that they can click on, and it immediately takes them to that page on a website or that document so that they can confirm that this is actually what they’re interested in, and it’s in alignment with their questions. But it’s also good for testing where we have a UI where you can ask a question and the answer comes back, and then you can look at all of the sources that were provided to the model to make up that answer. And it’s a way to say, “Now, wait a minute. It’s drawing upon this part of our website, or it’s drawing on this document, and that’s not really in alignment with this question.” Maybe you have bad information that’s on the website, or you need to refine some of the documents and things. So, it’s an iterative process.

Brian: Yeah, yeah. I wasn’t sure if my question came across [laugh] the right way or not. So, when I was asking about, like, testing this with users, what I meant—I guess what I’m trying to separate is, there’s knowing whether or not the facts are coming back accurately, and that’s one way to measure the quality of an LLM experience is, how often is it giving factual responses.

Mark: Right.

Brian: But, to me, there’s a difference between just the facts and the experience of going through getting the facts back. And so, the experience of that, we hope it’s going to be better with these solutions, but that’s something that can really only be measured through a qualitative study by observing a user and giving them, you know, a sample question to ask: “I’d like you to figure out whether or not this item is still under warranty, and if it’s not, what is your recourse to get it fixed?” You know, a task like this, and then watching them step through it, and then evaluating at the end how they feel about that experience and all of that. So, that would not get captured with classical acceptance testing because all that does is say, like, did it give the right warranty response, but it didn’t tell you what the experience was. So, I’m curious if you separate those out at all.

Like, do you do any of that one-on-one user testing with an actual, you know, Blue Cross customer. And we’re picking on Blue Cross; no offense [laugh] to anyone listening from there. I’m just… that’s a big insurer—for our international audience—in the United States. But picking on them for hypothetically, right? Like, do you ever bring in someone that has their insurance with Blue Cross and actually run them through this application and some real use cases like that?

Mark: Yeah, like I said, I agree a hundred percent that that’s critically important. And maybe the term acceptance testing, you know, to your point, it’s like, did it work, versus, was it acceptable to the customer?

Brian: Right.

Mark: So, I agree with you that, just like if you would roll out a new app, a phone app—we’re picking on Blue Cross—if Blue Cross rolls out a new phone app, they’d better take actual customers, have them try it out, give suggestions, is it working, does it do the things, as opposed to just have the technical team they, “Hey, this is a great app. We’re going to roll it out.” So, just like that, it’s critically important to have actual consumers of this try it. And to your point, there’s the facts—that’s important; you’d like it to be very factual—but it’s also the style of the answer. Here we’re dealing with what should be a human-like experience, so you want the customer of this to feel like they are interacting with a human.

So, it should use terminologies, and they should feel comfortable with the answer. They should be able to work their way through answering some of these questions. So, I agree a hundred percent, it’s not just about does it work, and does it deliver the facts, but that interaction. And that’s something that I think is brand new. We’ve talked a lot about natural language for years and years and years, but we’re at the point now where we can actually build these types of solutions that the customer is going to have either a conversation with, so voice is going to start being, you know, more leveraged than typing, but right now, typing is still an interesting way to capture the input and probably the most common, but you want it to be a human-like interaction.

I think it is important to use real customers, try it out and get their feedback as to what does work and what doesn’t work. Because otherwise people just geek out on the technology and say, “Oh, this is great. We’re using generative AI. Isn’t it awesome?” But the customers don’t adopt it, and I think that’s the thing that’s critically important, is making sure that it does provide an experience that is better than experience that you can provide in some of the other ways.

Brian: Yeah, yeah. And I think just to advocate for this, for people listening, qualitative study of this is really critical to do. Even if you’re going to try to run a survey about their experience afterwards, surveys do not tell you as much rich information in terms of knowing what to change because someone’s feelings are not really captured in a survey. We don’t get body language, we don’t see the way they’re talking about it during the study with us.

Like, are they annoyed or not? The analytics aren’t going to tell you whether they’re annoyed, unless they choose to tell you that they were annoyed. And what about all the people that don’t answer the survey, right? So, this is, like, this is classic usability and user experience testing for any type of solution. And I think this is a space where, like, improvements could be significant when you think about just how bad, even searches, like, web search on most enterprise websites, it’s just horrible. Like, on a general level, I would say it’s arguably horrible. A lot of people even know how to use [laugh] Google to go—

Mark: Right.

Brian: You know, do a search, a localized search, on a website because they know, they expect it to be pretty poor. So, it’s very easy to baseline current level, and then bring some people in and see if you’re actually making an improvement. But I don’t think we should skip that because it’s not—just the facts is not the complete story [laugh].

Mark: Right. No, I agree. Now, I’m going to show my age here, but 30 years ago, I worked for Nationwide Insurance in Columbus, and back then, we actually created a usability lab where we had a room that had cameras and one-way glass, and we would bring folks in and have them work with these applications, and hardware, and different things, and we would record it, and they were given instructions and asked to do things, and then you record all of that. And so, I even think, like a, you know—it doing something like that, most organizations don’t make the investment these days to do those things, but I think it gives you phenomenal information, like you’re saying. It’s actually understanding body language, it’s understanding the interaction that’s happening, the responses, all of those kinds of things.

Now, maybe you don’t need to go all the way to the level of a full usability lab, but at least capturing how people are interacting with the technology, what works, what doesn’t work, those kinds of things. And it also helps you design a better application because what you may find is, you know, most of the models, we try to have them communicate at a specific language level, but maybe if people are struggling and asking a repeat question within the process, you begin to lower the level of the grammar. There’s all kinds of things that you can do with the models. You know, it’s kind of like that show, Are you as Smart as a Fifth Grader? It’s kind of like, you start at a certain level.

And that’s one of the things that generative AI has done a great job of, is taking very complex—like, we load in a lot of the 10-K reports for public corporations, and these are 150-page documents that are extremely technical about financial details, but it’s very difficult to find actual knowledge in them. But you can bring this information in, and then use the model to ask questions, and have it convert it to a non-finance person language to better understand. Or it’s the same way with medical reports, you know? You get something back from a physician, you can use these models to translate that down into English as opposed to the very complicated terminology that’s used in most cases for tests and diseases and things like that. So, it really helps the absorption of the knowledge because you’re bringing it at a level that people can really understand.

Brian: Yeah, my brand-new logic board on my Mac has been kernel panicking every single night, and it’s like, you get this ‘Send to Apple’ report. So, I pop it in GPT, and every night it panics, and I pop it in, “Is this the same error? Is it related to the last one?” And it gives you a nice—it’s beautiful for stuff like that because I don’t read… I don’t even know what that stuff is. I don’t read that language, though. Like [laugh].

Mark: Right. Right. Nor should you really take the time to, you know? Again, and then if you call tech support, they’re going to have you go through all of that, and they’re going to start from scratch every time. So, whether it’s diagnosing, you know, a computer issue, or health, I mean, it’s great to feed the information in and have it translate, oh, this test does this, and all that.

And because you’re providing it the actual report, it’s drawing upon that to answer the question. So, I think that’s really where we’re seeing phenomenal benefit is that changing the level of complexity of the language that’s used in the creation of the source material.

Brian: Yeah, yeah. I kind of feel like with all this advancement here, we went backwards with the interface: we’re back to a command line. And my challenge with the command line is knowing what to type requires you to know something about computers if you want to get better results. So, I think a lot of people are going to use these, even localized models, like, you know, a website-specific model to do searching.

And so, my question for you is, how do we expose the capability of this information in a way that’s not out-of-band with someone’s task? Like, “Is my chiropractic care covered?” I don’t want to have to learn some vocabulary to correctly prompt the machine to get back information about my level of care around chiropractic care. Nobody wants to go learn a language to use this. So, do you think we’re at, kind of, stage one of interfaces, which is just give everyone a text area box to type stuff into, and good luck.

If you know the magic keywords to type in, you get a better response, and if you don’t know what prompting is, well, then you don’t get as good of a response. This is a pretty bad us—from a design standpoint, this is very immature. It’s basically throwing all the effort on the user. Are you seeing any advancement here? Do you see us going in a better place than just leaving this, kind of, blinking cursor as the interaction model?

Mark: Yeah. I mean, there’s a couple of statements and comments and questions in there, but I think fundamentally, folks that have jumped onto ChatGPT, and I think ChatGPT is great, however, it’s only using its training data set unless you go in and extend it and do a lot of things yourselves, right? So, if you just go into ChatGPT and you ask questions, what it’s using is its training data, which is a huge collection of information, but it’s, kind of, ten miles wide and one each inch deep. And that’s where I think the models where you’re using augmentation give you a better result to what you’re talking about. And I think also the enterprise types of solutions, it’s going to be critically important that the prompts that get included with every question really help address the things that you’re talking about.

The person shouldn’t need to type in that when they’re asking the question. So, to your point, if you’re asking a particular question behind the scenes, the enterprise has to bundle in prompts that help envelope that question, and then the context that surrounds that so that you get an answer back that’s acceptable. It still keeps with your command line typing a question, but the answers come back that are very relevant, but that is all being done behind the scenes in a properly designed chat solution for an organization, right? So, that’s the big difference between deploying an enterprise solution for a healthcare company or a retailer, and going into ChatGPT. So, because people think ChatGPT is awesome, but it’s not focused on answering questions around certain domains.

So, it will answer a lot of different questions—and the thing I always help people understand is if you go into ChatGPT and you say, “Who is Mark Ramsey?” It’s going to come back and say, “Well, there’s a lot of Mark Ramseys. There’s a jazz music consultant or analyst that is named Mark Ramsey, and there’s a bunch of other Mark Ramseys, so I really can’t help you.” Using that as an example, if you then go into our website, and you ask that question, it’s going to come back and give you the exact answer of who I am, my history, my education, all of that because you’re asking the model that is using augmentation to answer the question about a particular focus area. So, I think that’s extremely important that people don’t get concerned about using something like ChatGPT and really having to put in a lot of effort to get the answer they want.

We need to build enterprise solutions so that a lot of that complexity—and even ChatGPT will continue to evolve and will continue to get better, but I think it’s just there’s two aspects of these models: one is the training aspect, which provides a lot of information, and it gives it that human-like experience; and then there’s the using the model to answer questions. And I recommend organizations to always augment with their content and don’t just use the training data because you’ll still get that human-like experience that’s built into the model, but you’ll eliminate the hallucinations, you’ll eliminate getting bad information because of the fact that you’re focused in. And I think you know, even today, if you have a model that has been set up correctly, you shouldn’t have to ask questions in a funky way to get answers. And that’s really this layering that needs to happen when the question gets asked is, that the organization will surround it with prompting. That’s not something that we should ever expect the user to, and then the context.

And I think the UI will change to voice. We’ll be talking to these models just like we talk to Alexa or whatever. Hopefully the experience will be better, that you know, we’re not in that mode of having to ask the same question six times, but, you know, I think that will improve. But the even with the text, when you ask the question, you should be able to ask the question in whatever is most comfortable for you, and then it should give you the answer that’s in alignment. Now, maybe if we can hit that at a high level, I mean, there’s always going to be cases where the model misunderstands and answers a different question, and those kinds of things, but that same thing happens when we’re talking to a live person as well. So, we just have to, kind of, continue to improve in that particular area.

Brian: This has been a great Mark. Thanks for taking the time here. The, kind of, last question here is, I’m less focused on the technical side in my work, but I’m seeing a lot of strong opinions about how this is never going to work reliably for analytics-type questions, especially when we need numeric responses back because there’s just simply too many different possible queries that need to be run, and there’s all this work to get the RAG to basically translate everything into a structured query against a structured database. Are we going to replace business intelligence and analytics with these interfa—I mean, is it just a matter of time, or what’s your, kind of, take on these very numeric-type answers, where calculations are required to deliver a response? Is that… are analytics teams at risk? What’s your, kind of, outlook on that?

Mark: I mean, we’re still early days, obviously, but I do think that those will be replaced. So, BI tools will be replaced with a natural language. So, instead of building a dashboard, you know, you’ll ask a question and have the technology develop that dashboard and display, or develop a graph and those kinds of things. Where people need to understand is that a generative AI model is not a predictive algorithm, right? So, it is not able to run—in the generative AI it’s not able to run predictions, it struggles with some math, so that is not the focus for these models.

But what’s interesting is that you can use the model as a step to get you to answering those questions. And that’s where a lot of the models now support functions, and the reason that that is so important is that when you ask a question about something that is in a database, it actually uses its knowledge about the schema of the database, it can build the query, run the query to get the data back, and then once it has the data, it can reformat the data into something that is a good response back to the customer that has entered the question. So, it’s not doing the analytics—and that’s extremely important is that the GenAI models are not executing the analytics—but they now have capabilities that allow us to extend into some of these other realms. And we also have to keep in mind that if you look at something like an OpenAI or an Anthropic, these models are actually made up of a whole collection of models that operate in a lot of different ways. So, if you send a picture, and you ask it to understand the picture, that’s using a different model or a different component of the generative AI environment than if you submit text and ask for it to generate a response to a particular question.

I think over time, we could assume that those types of analytics will also be different aspects of the model. But today, that’s being done through functions, and we’ve actually worked with organizations to be able to ask a question, have the model execute a function, get data back, reformat it, and deliver it back. So, the model is not actually doing the analytics; it’s being done in the SQL query, which is the way that a lot of the BI tools operate anyway. And so, I think that is something that will happen. It’s just—to your point—that the models don’t do everything, and so it’s important to, kind of, break down the different types of questions and the different responses that you’re looking for, and make sure that there’s good connections between those. And I think over time, that’ll get easier.

Brian: Yeah. What I meant was, from a perception standpoint, though, as a business user, the model is doing everything because they have no idea how many—that you’ve chained together five models, and there’s two function calls, there’s a SQL query, there’s a render, you know, blah, blah, blah. That’s all plumbing, and so the perception is, is the model is doing this. And again, this is, you know, the perception versus—

Mark: Reality.

Brian: —reality. And this is what I was curious about, is whether or not you see teams getting to the point where they’re going to be able to cover every possible function, and that every query will be properly interpreted by the LLM, turned into the proper SQL, and that the data that comes back is believable. Because then you get into the question of, how do you know if the number of projected sales is right? How do we know if it’s right unless you read SQL? And I have seen, like, you know, ChatGPT Code Interpreter, which is nice.

You throw it a quer—you know, you give it, like, survey data, and it returns answers, but it also shows you the SQL, which on a simple, like, you know, like, on a survey, I can—you know, I know enough SQL to see, like, okay, it queried from the right table, and it asks—you know, I can interpret enough about it. But I could see, in a large enterprise, where it’s like, all these table joins, and you have no idea what these column names are, how would you know if it’s accurate? It seems like it’s a problem that would just be never-ending to try to cover the gamut of analytics questions that need to come back. But I don’t know. Is it solvable, or is it just, like, a never-ending, like—[laugh].

Mark: Well, I mean, here’s a question. In the organization, if you see a dashboard, how do you know it’s true?

Brian: True.

Mark: How do you know that the person who developed the SQL didn’t make a mistake when they were developing that SQL to generate that table? I mean, it’s the same type of thing. The way we know that is through verification. And also, to your point, I think we have to get a lot better at, kind of, the show your work button, right? So, if you ask it to answer a particular question, it comes back with the answer. We need a little show your work button that then opens a window, and it shows the steps that things have gone through.

And not that every business user is going to be able to understand that, but again, it is something that—it’s kind of like self-driving cars, right? I mean, we’ve talked about self-driving cars for years and years and years, people are still concerned, is it going to work? Once in a while, it doesn’t work and there’s an accident, and makes the front page of the news, but we forget that every single day, thousands of accidents happen, hundreds or thousands of people die in the US every day just by trying to drive themselves. But if we get one accident on, you know, automated driving, that makes the front page news. And so, I think it’s going to be the same kind of thing here where we’re just going to have to work through some of these areas, we’re going to have to provide the context, and do a lot of verification. But if implemented correctly, I think the models will be on par, if not better, but we just need to make sure we are able to, you know, show the chain of thought that has actually led up to the answer.

Brian: Well, and you make an interesting point. For the analytical minds out there, dying from a driverless car is probably not the same thing as a loved one dying from a driver in a car, a human driver. They’re both death. The experience is not the same. It does not feel the same way.

And understanding this, I think, is really important to figuring out, are we creating better outcomes for customers and end-users of these solutions? You’re not going to be able to interpret that just from analytics. Analytics are just going to tell you how many calls did we offload from the call center. It will give you answers like that, but it won’t tell you about the benefits and the outcomes to the people, unless you’re doing the proper user research with actual human beings, and talking to them, and observing them using these systems. We have to do that.

Because you’re not even to know what to look for. And in this ca—you know it’s, we’re not talking about death by a vehicle here with an LLM, but we don’t even know what we don’t know about these because we’ve opened up this interface that theoretically seems like it could answer just about anything. It’s a huge thing to open up. You’re giving a lot of power over to users, which in one ways, is great. But I want to give you the last word.

Any closing thoughts—before I, you know, get your contact information and tell people how to find you—any closing thoughts on user experience and UX for LLMs or GenAI in general? Or any advice that you might give to data product leaders listening to the show?

Mark: Yeah, no. I mean, thanks, Brian. I mean, I think it’s critically important that we view the user experience on an LLM like we would do with other user experiences. And I think in the case here, our objective is to make it as human-like in the interchange as possible, and make it easy for folks to ask questions. I mean, we’ve talked about natural language interfaces to computers for years.

Brian: Yeah.

Mark: It’s always been very clunky. To your point, you had to know exactly how to format the question to get the answer. What we’re really targeting here is, have a conversation with the computer, do it in a way that is natural, and you get an answer back that’s natural. And that’s really the objective. And I think that’s the target that organizations should set is that, in whatever kind of business use case they’re going after, is to have a very natural interaction to the point where—I don’t think we need to really measure, hey, does this pass the Turing Test, does it actually feel like we’re talking to a human, but it just needs to be a significantly improved experience from where we’ve started, with things like pressing buttons on the phone and trying to shout our account number six times into the phone and get it picked up.

I mean, it needs to be a very human-like and natural way to interact. And I think if we can do that, it’s going to be very well accepted because there are a large number of folks out there that they don’t like understanding the technology, right? They don’t like reading manuals, they don’t like reading documents, they’d like to just ask some questions, get some answers. We’ve really missed the mark in search on that because we ask a question, and we get 2 million links. We don’t want to go through all those links and try to boil it all down.

That’s the real potential here, is boiling down all of that information into an actual answer. But it’s been great today. I mean, we’ve gone from AI to BI, and from analytics—

Brian: [laugh].

Mark: —to death, so, I mean, I don’t know if we could have covered any broader perspective here, but it’s been a lot of fun. I enjoyed it.

Brian: “From Analytics to Death.” Maybe I’ll call the episode that. So [laugh]—

Mark: [laugh].

Brian: Mark, this is great to chat with you. Do you hang out on LinkedIn? Where does someone go suck up some Ramsey knowledge if they want to learn more about you and your work?

Mark: Yeah, no, definitely. I’m on LinkedIn. We’ve got a YouTube channel which shows a lot of the cool stuff that we’re doing. Or they can shoot me a note at mark@ramsey.international. But the best way is just—the easiest way is just hook up on LinkedIn.

Brian: Cool. Well, I really appreciated your talk at CDOIQ. I was able to glean a lot from it. It was very straight shooter, not a lot of fluff. I didn’t feel any vendor jargon and all that baloney, so it was just nice to get a straightforward, business-oriented presentation on that, so great job on the delivery that with your partner. I forget his name—

Mark: David Paget-Brown from Modak.

Brian: David.

Mark: Yes, yeah.

Brian: That’s right, yeah. David. It was great. So, thank you. I hope you stay in touch with you, and thanks for coming on Experiencing Data.

Mark: Yeah, thanks for having me.