Shop Talk: Accountability and Opportunity for AI

June 23, 2023

In this bonus episode, Eric and Kostas chat about all things AI.


The Data Stack Show is a weekly podcast powered by RudderStack, the CDP for developers. Each week we’ll talk to data engineers, analysts, and data scientists about their experience around building and maintaining data infrastructure, delivering data and data products, and driving better outcomes across their businesses with data.

RudderStack helps businesses make the most out of their customer data while ensuring data privacy and security. To learn more about RudderStack visit


Eric Dodds 00:03
Welcome to The Data Stack Show. Each week we explore the world of data by talking to the people shaping its future. You’ll learn about new data technology and trends and how data teams and processes are run at top companies. The Data Stack Show is brought to you by RudderStack, the CDP for developers. You can learn more at Welcome to The Data Stack Show Shop Talk where Costas and I talk shop cost. This has been a while since we’ve done one of these lots. I have a new baby in my household. So that’s kept me away from the microphone for a little bit. But this is exciting.

Kostas Pardalis 00:43
Well, yeah, yeah, I miss restyling with you. So let’s do it again.

Eric Dodds 00:48
One of the interesting things about the primary interface and ergonomics of chat GPT is that it’s one, I think it’s amazing, because it’s just so familiar, right. And it’s just such a, it’s actually almost, you know, 10 years ago, it’s like, you know, the most popular, like, the most widely adopted, like AI, LLM is going to be like a text exchange. You know, I mean, it almost sounds crazy when you step back and think about it. But brilliant. From an interface standpoint, one of the things that’s really interesting about presenting a bunch of options as search results that have some sort of ranking that’s at least nominally transparent to the end user who executed the search, is that you’re presenting some level of choice. One thing that’s very interesting about the chat interface is that when you app when you put in a specific query, the results are largely singular. Right? And so you’re actually not presented with a lot of choice. And that really, in some ways to return to the accountability question. What’s really interesting about presenting choice is that there’s more accountability for the user, because they have to decide, right? Like, I’m offering you a variety of things. And I’m applying some weight that may or may not be transparent to you. But you ultimately have to decide what you believe to be sort of the best match, but it’s very linear. And you’re getting a pretty singular response. What do you think that causes? I mean, I think it ‘s one of the roots of the issues of accountability, right, is that it’s almost a binary choice, like, do I trust this? Or do I not? And I don’t have any surrounding context, that gives me a toolset with which to make that decision.

Kostas Pardalis 02:58
Yeah. 100%. And also, I think it’s even worse than that, because even if you ask the same question, if you ask, like, CPT, it might generate a different response, right? Like, there is like this, it’s not as deterministic, let’s say, as it is, like when you said like on Google, right, like in Google, if you’re liking a keyword like Moz are going to get like the same results all the time. Right? Yeah. With the sub CPT, like, that’s not the case. And I think it’s one thing like when and that’s like, where things like I think are getting, like, quite interesting when it comes to the grading age with products and trying to productize it, right? Because how do you account for that, especially in your skills, these were critical decisions made, right? Like, if let’s say, I’m, I don’t know, let’s say I’m a doctor, and I’m using like, subjectivity, like to summarize stuff. Yeah. Right. How can we build a user experience product that will help the doctor to do the summarization, but at the same time, we’ll say it guards, both the model and Docker. From there either breathing like the wrong information or like making mistakes that it’s like collusion like interpreting like hallucinations like for example, it’s just like, what the like pretty much like everyone I guess, like hears about like, a large language models. So I think that’s the interesting part. And like, that’s why I’m very excited and I think like the people who should be extremely excited about AI and especially MLMs, our product managers, I think there’s like, a very new modality that is given to them. And it’s a very new tool, which is almost like completely uncharted territory right now. I think there’s a lot of space for creativity in product design and management to make something very interesting then create new types of experience there.

Eric Dodds 05:27
Yep. Yeah, I totally agree. One. So, almost this is distilling it down to like, two levels, which is not actually reality. But reality is a construct, so it doesn’t matter. And I’m gonna roll with it.

Kostas Pardalis 05:40
So even the matrix, right, like, yes.

Eric Dodds 05:45
I think that there will be a day actually, I rewatched. That was in the hospital when the baby was born. And I forgot how good of a movie that was. I remember watching it for the first time. Yeah, the first one. Yeah. Like, yeah, unbelievable. Yeah. But one thing that is a little bit disappointing to me about the hype cycle is that you mentioned earlier that this is going to dramatically accelerate the amount of human creativity that can be put into jobs by doing things like pulling information out of PDFs, and summarizing it, right? I mean, if you mentioned the medical industry, right? I mean, there are entire, like, sections of publicly traded companies that have armies of people who go through PDFs, and manually, like copy information from a PDF into like, some other system, right? I mean, that is wild to think about, right? And it’s like, Okay, those are creative human beings, if we can remove that from them and allow them to, like, you know, attack different problems like that, to me, the first wave of the types of things that this will do will actually be like, pretty utilitarian like that, and they will just unlock a bunch of bandwidth. Which, to me, is extremely exciting. I know, there’s concern about job loss, but I’m with you. I’m actually not concerned about that, because I think that it will free up human creativity, which will be a net benefit to society, which is pretty exciting to me. But then I think then, you know, once a number of those problems are solved, you and even now, what people like to talk about more is the tricky thing, should this LM be able to make a diagnosis for a doctor? And where is the jurisdiction, you know, from the doctor and from the patient? And how do you regulate that? And, you know, that’s where things get really tricky. And that’s where people tend to focus the conversation, because they’re like, what’s going to happen, you know, is this continuing to like, wrongly diagnose your disease, but to your point way earlier, there’s going to be monumental changes for things that are not that exciting, like from a news cycle standpoint, right?

Kostas Pardalis 08:10
Yeah, 100% Actually, I think there is a, I keep, like, sharing these insights, because I don’t know if it was like, very surprising for me when I experienced it. But I think like, it’s, it’s a glimpse to the future of like, how AI is going to be like, in my opinion, like implemented in like, how we are going to embed it in our everyday life. So I don’t know, I guess everyone is familiar with Invisalign, right? Even if they haven’t used it, they probably know what it is. If you go through the process of doing a treatment with Invisalign, you’re going to notice something if you pay attention, okay? The whole process can happen without actually having a doctor there. Like you don’t need the doctor. Like you don’t even know what the doctor is doing when you go there, like at the beginning. It’s something that you can pretty much do on your own, like you scan your teeth. They take the scans with a phone, by the way, you can do it. Feed it to the algorithm, and the algorithm comes up with a treatment. And the treatment is then fed into like a 3d printer, they print Invisaligns molds and they send them to you. Right but no one does that. Right? Like we always go either to a dentist or like an orthodontist. And we go through this process. And the reason we do that is because the modality is so different, that like humans would never like, throw us that Right, like, it will take time to get to the point where you’re like, I’m going to put something in my mouth for like the next six months. But it’s a computer generated by a machine, and it is a completely generic machine. The doctor there is there to take care of like, okay, some cases that the outbreak the edge cases or the algorithm, or like complications, right. And most importantly, to make you feel safe, and trust in the process, okay? Because you are interacting with a human right. Now, the benefit did like the opportunities, like lost their jobs, no doubt that like more and more people at the end got access to these kinds of treatments in lower prices, and they scaled their job because now they can in their like, offices, they can have, I don’t know, like five eggs like the patients will they had? Right. I think that’s what is going to happen like with other professions that have lawyers, like they’re acting with the public sector, like all these insurance, you know, insurance claims, right? Like, all that stuff. We will just be able to do more. And that’s going to be like, in my opinion, like a positive sum game or the end compared to the negative sum game that is like crypto wars.

Eric Dodds 11:29
Yep. Yep. Yeah, that’s super interesting. Okay, one last. One last thing, what have you built? And have you built anything with an LM yet?

Kostas Pardalis 11:43
Yeah, I’m using it. I’m using them like, a lot. I mean, I’m using them. Yeah.

Eric Dodds 11:52
I’m beyond just like a personal because, like, I think a lot of people who are interested in it, use it for different things on a daily basis by just sort of inputting a bunch of prompts, but have you like, I think you’ve built a couple end to end things on it, right?

Kostas Pardalis 12:08
Yeah, no, I’m building some stuff. Yeah. Like, for example, like, one of the ways up, like, and it’s funny, because you mentioned something similar. Like, for example, I wanted to build, like, a synthetic data generator. And, you know, especially when you have low cardinality data that you need there, it can be super helpful because it can create much more realistic data. So for example, let’s say I want to even like with let’s take, like the RudderStack event, right? Like, if you want to create, like synthetic data for, like the agent field, right? Instead of like, trying to figure out like, inputting, like, just like a junk of likes, characters there. Yeah. You can get like, I don’t know. Yeah, like, like, variations of like, hundreds of them. And just sample them and, like, use them like to generate data, right? But if you don’t have to continuously do that, like, just to save your data without because it’s sort of slow to do it, like every time and expensive. Yeah, but um, he’ll do these kinds of things. And you didn’t like to code. Super, like useful, especially like, I was always like, I’m gonna like, I was never like the person that enjoyed writing tests. So now it’s pure joy to write tests, using like, copilot. I’m doing a lot of, like, old like, I got to use like, whisper a lot to analyze like, the episodes here that we are. Yeah, we are recording. I’m using my journey to generate some assets for presentations and stuff like that, like things that like, okay, like, Hey, by the way, keep in mind, I can’t even like to draw like a straight line, right? Like, for me, it’s like a huge enabler, that’s, you know, I can go there and create something decent. So like, it’s going to be perfect. It will never be perfect. And yeah, like, needs a designer. I will never like, substituted a designer. But at the end, the designer can iterate much more on what I’m asking because of these tools, right? Yeah. So we can end up liking what I need, like Foster. What else have I done? I think these are like the things so far. I’m using the love like, why like gold, to be honest, like, he is really interesting. It cannot substitute coding. But it changes the experience a lot like what has changed is that I’m not using Stack Overflow anymore. Yeah. Motto Yeah,

Eric Dodds 15:00
I mean, I’m not writing code in the same way that you are, but querying data a lot. I mean, I’m not, you know, I never, I’ve never had a job where one of my, like, stated responsibilities is writing SQL and so, or like regex stuff, you know, and so, if you’ve never done that professional, and I just haven’t spent the time to actually sort of get over the hump of like, you know, figuring out how to do, how to structure a query in a way that is efficient, and, you know, whatever. It is amazing that, where it’s like, I’m trying to answer a question. And I know, like, the basic structure of how it should look, and it’s like, just correct this. No, it’s amazing.

Kostas Pardalis 15:43
Yeah, the other thing that’s also like Golang, mixed with a conversation about, like Google, for example. They’re like, Okay, I have, like, I’ve never learned or that or like specific terminal commands that I keep forgetting exactly like what the parameters, right? Yeah. One of the biggest problems that I have is like I always forget, like how to do an SCP. And I used to just go and google it, because you pretty much get the exact answer when you want. But what happens is like people have like gamed the system show mobs where you are, like, ending, like in a landing page where there is a command, and there’s so much advertisement on meat. Like a shopman. Like, it’s so annoying. Like, I can’t see anything.

Eric Dodds 16:29
I know exactly what you’re talking about.

Kostas Pardalis 16:33
Yeah. Like it now instead of doing that, like, I’m just asking, like, the layman like it gives me you know, like, the commands. And even if it doesn’t work, it’s close to what I need. Right? Right. Right. But I didn’t have to go through all that crap of like all these like, dodge of like, ah,

Eric Dodds 16:50
Yeah, I totally agree.

Kostas Pardalis 16:54
That’s my goal. So you know, like, like googling their way, also, like hearing themselves by allowing this to happen. Because at the end, it’s all about the experience. It’s just like, it’s the same information. There’s nothing creative. You like your body like on Google?

Eric Dodds 17:12
Yeah. It’s an absolutely fascinating study in the user interface, I totally agree and accessing information. It’s also going to be really interesting to see Mike, you know, to see how this impacts Microsoft. And kind of what that does in the market with the big players, which will be super fun. All right. Well, Brooks is telling us we are at time here. And maybe we’ll try to. I don’t know if we should, should we try to ask an LM to generate an entire episode around a topic and see if they get our tone and voice.

Kostas Pardalis 17:58
That would be interesting. We can do something we can create. Let’s call it a sub GPT.

Eric Dodds 18:07
Yes. That would be fun. All right. Well try it. All right. Well, thanks, everyone for joining us for another shop talk. Kostas, thank you for sharing all of your thoughts, even if they’re, you know, just a construct and we will catch you on the next one.