EP 002 - Large Language Models for Business: Strategies and Applications

Svetlana (00:00)

Thank you for watching!

Edgar Dyck (00:02)

Hello and welcome to our second episode of the AI Boardroom. I do this again.

Svetlana (00:09)

Hello, hello, excited about this one.

Edgar Dyck (00:13)

Let me repeat this.

Hello and welcome to the second episode of the AI Boardroom. Svetlana, lucky, I'm happy that you're here. So yeah, I also lucky you're here, honestly.

Svetlana (00:27)

Lucky that I'm here. No, it's awesome. Looking forward to this episode, I think, where we're gonna be diving into the large language models and what they can do for businesses and how businesses are leveraging them right now to kind of capture some of the business value.

Edgar Dyck (00:48)

Basically, do not ask what you can do for AI. Ask what AI can do for you, right? Ha ha ha.

Svetlana (00:55)

Well put.

Edgar Dyck (00:57)

Okay, so yeah, today's title, we have to come up with better titles, but it's basically, the goal of the episode is that after this hour you will be a lot more able and capable to think about what you actually can do with the new AI stuff that's flowing around everywhere. And yeah, that's what we're aiming for.

Svetlana? Yeah.

Svetlana (01:23)

And something just to set this. I'm sorry, go ahead.

Edgar Dyck (01:26)

No, I just I wanted you to do exactly what you wanted to do

Svetlana (01:30)

Yeah, now, yeah, so I think one of the things that I wanted to kind of maybe start off with is just kind of lay it out there. A lot of hype has been in the market in the media about A.I. And any time they open up an article or anything like that, you know, ads, when they say A.I., typically, I think in the reality of today's day and age is that they usually talk about large language models, chatbots. And so one thing that I want you to make clear that

AI is not only large language models, but because of the attention it's getting, there's an assumption that all AI can do is these chatbots. But it's only a small portion of what value you could capture with AI, which we'll dive in over the course of different episodes. But large language models are definitely, I would consider them low-hanging fruits in the business. You could take off the shelf. You could build.

some versions of your own kind of a chat bot and bring it into your business. But to me, it's a no-brainer. And I think that's why I think a lot of people are kind of grabbing onto that title that all LLMs are, or basically all AI is large language models because of how broadly and how widely adopted they're becoming. What are your thoughts on that, Anagar?

Edgar Dyck (02:51)

I think you said something really important, like low hanging fruit is basically how I started this whole journey thinking about business application. I was thinking about what is the like quickest win to get from this. And another thing you said, which is really important is general purpose. I don't have to train the model necessarily.

be a good option but it's an option and not a must as it was before like and I think it's all coming down to the huge datasets the stuff is trained with but the general usability is off the charts and nothing we've ever seen before that's said I always encounter this when I demo or talk to people

You always have this chatbot interface and it basically looks as it always looked. But I didn't, and I tried, but I didn't yet come up with a better way of interfacing with the AI.

Svetlana (03:56)

Yeah.

You mean for to be able to interact with AI?

Edgar Dyck (04:10)

Yeah, because you need an interface of some sort. And you have voice, you have video, you have text, like a lot of interfaces are thinkable of in terms of how to deliver information. But I always find myself coming back to chat because it's the most natural at the moment, at least for the generation that's working with computers right now.

Svetlana (04:14)

Yes.

Edgar Dyck (04:35)

Because for me, even like I'm really a geek and I really like all things technical and still to this day for me It's not natural to speak to my computer Like I would speak to a person so and that makes it basically makes me basically always fall back to the chat because Yeah, we don't talk yet like confidently to computers

Svetlana (05:02)

Well, I think it's just an evolution. There's always going to be something that is in the middle of that interim period where things are still uncomfortable because they're new. But I do think that we're, you know, chat bots, if you step back and look at the situation, the way that we used to converse with these Googles of the world and everything is really unnatural. You're asking.

a one-sided, you're starting a one-sided conversation and you don't really get feedback other than clicking some resources. Now you're really kind of with these chatbot functionalities, you're really are trying to mimic that natural conversation that happens between humans. And why I think this is so game changing because you can pose a question and then you can build on it.

and you can get clarifying topics. This is unlike you getting education from a teacher, right? You raise your hand and you ask the question. You and I are talking here, where it's not me speaking at you, it's us kind of speaking together and building on certain topics. So I think that's really important. So I do think that that's gonna become more and more evolved as we kind of speak, but I think that's really why people are leaning towards this experience because of this interactive kind of.

feedback that you get by chatting with these chatbots.

Edgar Dyck (06:26)

Yeah, definitely. I think multimodality is something that really will make huge leaves this year, I have a feeling. I can install open source models on my computer, they're not even that big, that can read pictures or look at pictures and say what's on it. That's pretty...

Svetlana (06:50)

Can you define what maybe for our audience, what multimodal is?

Edgar Dyck (06:54)

So multimodality, to be honest, I had to Google it myself back then. So a modality is basically a type or form of input. So if you have a modality like text in a chat board and you type in text and get text back, text is your only modality. And it's oftentimes was like back in the days, it's the only modality the AI is trained on. So it only is trained to read text

trained to give tags back.

truly multimodal models are able to basically take in raw audio or raw pictures and then like look at it basically so it's not like we did with object recognition stuff like that pattern recognition like it is pattern recognition but in the end it's basically a true understanding of the content you're

presented and the magic is if you have a true multimodal AI you can give in a picture and it gives you back text which is obvious but you could also give in a picture and get back audio and that's where it really starts to get interesting because like AI is kind of

only text at the moment, even if they say it's multimodal. Because if you talk into chat GPT in the app, you can use your voice and it can answer your voice. But it takes your voice, translates it to text and then processes it and gives you text back and reprocess it to give you voice back.

And that's important because you lose so much information. What's the tone? What's the emotion in the voice? Was it loud? Was it like, was he barely speaking? Stuff like that. And that's where it gets interesting. So as soon as we are truly multimodal, we can take the audio and not only the text information, the audio, but also everything around it, which are factors that are like really important

and take them and analyze them and also give like audio in that form back.

Svetlana (09:24)

So the current state of AI, you're basically saying it's not truly by definition multimodal. They're converting input, whatever it is, whatever object, whether it's voice, whether it's an image that you, it converts it into some text format, maybe a description of what you're looking at. And it's still the large language model, which is text-based, that is processing that information and then basically outputting it in whatever preferred format you have.

The way you're describing it is that it's not truly multimodal.

Edgar Dyck (10:01)

Where we already have reached true multimodality is with images. So if you upload an image to Chetjiputti and he analyzes it, it's really going through the model. At least with GPT-4. But for example, if you have video, it's not. If you have video, it takes a video, makes a sequence of screenshots, puts them all together, and then processes it. Which also works pretty well, so it's still able to interpret the movement in the series of pictures.

use the pictures.

But it's not really reading the video, including the audio and stuff like that. So you still have to transcribe the audio. I don't want to get too technical in this, but just that multimodal with real modality for waveforms, like what's the basis of audio, that's the trickiest thing to do. There is a lot of stuff. Because you can say like 50 times the same word.

and you have 50 different waveforms. And it's really interesting. It's another topic. And this one, I think, will still take a bit of time.

Svetlana (11:15)

Well, I think you haven't said this before, but you've spent a lot of time in video production and directing. So you understand voice and video really, really well. So it comes from an area of expertise. And this is why I think it's such a maybe even a passionate topic. And I think where voice is also becoming important is chat bots to support customer service. So if you are a voice, you're a voice.

Edgar Dyck (11:40)

Yeah, definitely. Yeah, yeah.

Svetlana (11:41)

speaking to an agent and let's say even behind the scenes it's powered by a generative AI and then as you said like you could speak to it and then it's going to return voice back. If you're not really picking in those nuances in the voice then you're kind of really missing the frustrations people are having right or the urgency of the requests maybe they're speaking really quick and they need to get things done and you have to kind of as humans pick up on those you know unsaid kind of traits in our voice but it's really hard to do that if you're

Edgar Dyck (12:04)

Yep.

Svetlana (12:12)

all you're doing is converting voice to tech.

Edgar Dyck (12:15)

So, like today's podcast should also like...

deal with how can I use it for my business. And for example, if you have customer support, you can like two-fold it because there are AIs, and they're pretty good, that can take basically any modality and read the emotions from it. Like they are trained to interpret the emotions in the waveform, like in the real audio. And you can combine that with the transcription. So you then can build solution

which uses the transcription functionalities to get the text itself and can then also give the LLM, because we are processing language again, that you build up a context and you build up a prompt basically. And in this prompt you can then put the transcription and you can put the emotions. And then the prompt can give you an adequate response to that.

I'm not sure if it's all... I'm not sure, but I'm confident that you could do this like the other way around too, so that you can generate voice that's angry or something. But... Like, just to feel natural, you know?

Svetlana (13:31)

The sales representative or whatever the customer service agent screaming back at the complaining customer.

Oh, that's funny.

Edgar Dyck (13:42)

No, but honestly, that's how you have to think about the stuff right now is if you build solutions, how can you extract information from the context you have, like from the reality you have basically, transform it into the right text information and then let the LLM give the right text answer and then see how you can go from there. And that's basically what every AI solution out there is doing.

That's basically what they do. They take the information, make text from it, get text back and see what they can do from there. That's why I find it kind of impressive that Microsoft is able to take an Excel file and do a PowerPoint from that and then they co-pilot it. But at least that's what they sell. Didn't try it yet. Because everything you get is text and from that you have to work with it.

Svetlana (14:38)

Yeah, and so I think that's interesting. I think we jumped in to the content a little bit deeper without kind of, I wanna say, I would love for us to maybe take a minute to address like, why should businesses, business leaders really care about this technology and what can it truly do for your business? So I think understanding kind of the use cases, I think are great, but what can, what does this technology really do? Like, why can't I just continue working at the same

kind of under the same business and operating models that I currently have, what's changing in your perspective with like, what's the huge benefit that these AI systems really bring and why should these business leaders care?

Edgar Dyck (15:25)

As far as I'm concerned, I'm always look at it in two different categories. The one category is I have a running business, which for example has a lot of support, which binds a lot of resources. And there I can go in, step in and improve on my existing processes.

That's like the one category, like take your existing business model and see where you can be more efficient, more cost efficient, save costs, or maybe generate more revenue through better content, stuff like that. That's on the one side. The other side is companies that are really taking that opportunity to improve on the product or to even ship a new product based upon their existing field they're operating in. And that's where it gets interesting.

for everyone to really dive deep because if you're going that route, the requirements and the effort you have to put in are a lot higher, of course, because, yeah. One example would be BI, business intelligence companies, that they produce graphs from data basically and give you reports.

Svetlana (16:35)

Mm-hmm.

Edgar Dyck (16:39)

and being able to maybe not only give you the report, but also give you suggestions by using AI.

to how to interpret this. I even say, or like my feeling is, we will get to a point where business intelligence solutions won't show graphs anymore, because why do you look at graphs? It's to get a visualization of the data. Why do you want a visualization of the data, sorry? You want this because you want to be able to extract information quickly at a glance.

Svetlana (17:04)

Oh yeah. Hahaha.

Edgar Dyck (17:16)

And if, but if I have a language based solution, which just tells me what the outcome is, and then I still optionally can look at the graph, but I don't have to because I just get the answer that I'm searching for, why the graph? So I think that's the direction where this might go. But this would be an example where you like transform your existing business model and take it a step further using AI.

Svetlana (17:17)

Yeah.

Why do I need to look?

That's interesting. You said something that I didn't really think about much, but yeah, I think it's that interim step, right, that we've kind of talked about is what's, you know, you're looking at graphs and the way that we kind of do business or operate our businesses right now, we just kind of have come to. Again, there's an evolution of ourselves before, if you think about like those graphs, we're doing them by hand. I remember even, you know, when we were in school, drawing all of these like dots, and then all you're trying to do is you're trying to make sense of the data.

Edgar Dyck (18:04)

Yeah.

Svetlana (18:10)

So you have all of these points that you kind of have to represent on the graph to understand trends. Well, we were able to do that with rule-based systems. Now AI can take it a step further. Why do you even have to look at the graphs? Because again, the premise behind them and the goal we were hoping to achieve from those graphs is just basically get the insight so I can make a decision based on this data. Well, what if you could you bypass that graphing step? To get from the problem that you have, you have a bunch of data.

to some insights about that data that you can, it's enough for you to make a decision. And I think that's really key. So, and as supportive, maybe in support of those kinds of insights, maybe looking at some graphs, but that becomes optional. And I think that's truly impactful. But you said something else that I just wanna highlight because I think it comes down to, you know, why businesses should care really is, you know.

The goal of any organization is to drive kind of their top line performance. And there's two ways that organizations really are successful. And I think you kind of touched on both as how do we continue to drive more value for the organization by increasing profits? So what could we do more of? What kind of offers can we provide or services to our customers to drive more business?

And then how do we reduce and do that at a lower operational cost? Right? So it's profits minus costs equals, you know, better value delivered and top-line performance. And I think AI can do both. It can help with, you know, better product recommendations, better user experience that drive more customers to do more business with you. But then, as you mentioned, it could do tasks that we've just kind of gotten so accustomed to that it could streamline.

a lot of documentation reviews and other things. So, and then if you free up people from reading documents to relying on AI to help you kind of do that for you and derive those insights, you've saved like hours. And what if you multiply that by a workforce of let's say a hundred? Just check how much, I mean, just look at the numbers and look how much efficiency that's gaining. Now those people are freed up to do more of stuff, let's say to...

Edgar Dyck (20:19)

Yeah.

Svetlana (20:30)

drive more customers to your business. So it's a virtuous cycle.

Edgar Dyck (20:37)

And one thing I also wanted to point out is, it was like, it was in every, I mean, every transformation, if you have broken processes, I won't fix it. And also you have to sometimes rethink a process because like, of course, like your whole organization was putting out graphs in your BI solution, but that was only necessary, the necessary tool for representation of data

anything else. So now you have to rethink that and see is it still, do we still need that tooling we needed before to achieve what our value proposition is? And have to really ask yourself, because if you just throw AI at an existing process and like it might not even not deliver value, it might even clash and that's a big risk. So yeah, like always think about what you're doing.

Basically.

Svetlana (21:38)

Well, it's going back to the core. Like, why the heck are you using this tool, right? All solutions are built to solve specific problems. So what problem is this software or this process solving for you? So maybe it's to, I don't know, to get insights to make better decisions, right? So right now, again, as we just kind of talked about, using some of these analytical tools or maybe data representation tools to be able to...

display a lot of this busy data so that I as a user can Get those insights Basically quickly, but in my experience, I mean even building dashboards, right? There's usually a person behind the scenes interpreting all of these graphics and then making Connections and all of that. So there there's a human behind that So what if you could streamline that graphing and insights tasks and then you're basically Are freeing up that person to do something else?

Edgar Dyck (22:19)

Yep.

Svetlana (22:35)

And maybe they're the ones who are preparing these reports, those thorough reports and coming to you with those insights to say, based on our analysis, these are the options that we derived from the data. What do you think?

Edgar Dyck (22:36)

Yep.

And I think that's...

a lot more straightforward. And then you can also have a conversation on the data, right? So you can just like, you not only get an interpretation of the data without even looking at it, you also get the option to talk to the AI to see why it comes to that conclusion, which might lead to even better results because you can have an active discussion. And yeah, that's, I think we are also,

honest it's really hard to get this reliably working through a whole bunch of different use cases. So there has to be work put in. If you go that deep, you have to really think about the cutting edge the technology has to offer. But in the end...

like always, we have a new tooling and we have to become better using it.

Svetlana (23:53)

I think you said something else that I want to touch on, because I think it's treat AI. So when you evaluate your processes, it's not about adding another tool. It's how do you either augment existing tools or looking to replace, because it's not about adding more point solutions to support your workforce, but how can you maybe even eliminate some? What can this tool now offer? What problem can it solve now that maybe previous tools are now redundant? And so I think that's.

part of maybe that evaluation process that you have to do, because it's not about accessing five different tools to solve that problem, but could you do all of those five tasks with one?

Edgar Dyck (24:31)

Right.

Yeah, and from like I've done a lot of customer software like business software to optimize business processes through the last 30 years because that was my job basically. And we had this over and over and over again that people were just trying to put their old processes on the new software on the new managed like on the new software tooling which should like improve their daily business.

copy what already worked. And I get this, we humans are like this, and we're like really trying to stick to our known bads rather than taking the risk for a better good. An unknown good, sorry. Yeah. Think that's something every executive has to keep in mind for that.

Svetlana (25:23)

And I think...

I'm sorry to interrupt, but I think what you're kind of saying is really gold. And I think it's resonating with me also is that, you know, we like being comfortable. We like to strive a balance. And I think when organizations are established, I think I think step number one is recognizing that you have a problem so you can address it. But ultimately, it takes, you know, all of the organizations are born from somewhere, right? They started with small.

Edgar Dyck (25:46)

Yeah.

Svetlana (25:54)

you know, five people teams, and then they've grown out to be thousands, if not hundreds of thousands of employees. Well, in order to operate at that level, you kind of have to put frameworks and things in mind to keep things stable on the ground, right? So you just kind of have established ways of working. You have tools that you standardize across. And so change is hard. And I feel like that's also something that organizations consider. So sometimes yes, change is hard, but starting may be small and really kind of testing to see.

Is there an ROI? Is there a benefit to my business? Maybe at small scale, because if you multiply that by the number of employees you could expand this to, I mean, the benefits are explosive. So.

Edgar Dyck (26:35)

Yeah, and also like, I like now started hiring people for my company, young people, unskilled people. It just, because I love to like work with people and see them develop. And I'm the one helping them do that, like it just satisfies me. But I also really, really look forward to give them the tooling, the AI-based tooling,

and I'm really curious to see in that. I will definitely on that podcast will give some reports on that how it helps how it helps them to be more efficient and like be able to do things they weren't able to do the day before.

I don't know if I should tell this here, but every year there is a challenge called Advent of Code. And Advent of Code is basically, you have a calendar and every day you get a new challenge, a coding challenge. And I did it in 2022, I did it the first time, and I was like, oh, that's nice, I learned a lot. Let's try this again next year. And this year I just said, okay, I'll try a new language, and I'm going to use AI for that.

Svetlana (27:44)

I'm sorry.

Edgar Dyck (27:53)

didn't use AI like you would think of it, like hand in the problem and try to get a solution. I was giving it the solution I wanted, but in a new language. I wanted to learn the new language. And it wasn't even a big language with a lot of resources. It was a niche language. And I was able to learn the language in a day because at first I got the code and the solution was what I suggested the solution should be. So I thought about the solution myself.

And then the next step was I tried to understand the code because that's necessarily what you have to do if you wanna learn a new coding language is to be able to understand the syntax and the concepts behind it. And the AI was just able to explain everything in the detail I wanted it to. So I was able to learn in a day in completely new programming language which was not even concept wise near what I'm used to use. And it was awesome. It was like for someone like-

Svetlana (28:46)

Wow.

Edgar Dyck (28:53)

me who's always thought about educating people, I was in awe.

Svetlana (29:00)

But it truly resembles an active learning type of style. It's like the question answer. And again, I think this is why I think AI is game changing, is because it keeps the conversation relevant to you. So instead of you reading, and the problem that I have with reading books and courses, I just want to know a specific thing. Just tell me which page it is on so I can get my answer. So I don't have to read the entire book to get to that. But.

Edgar Dyck (29:25)

Yeah.

Svetlana (29:30)

AI is able to keep that relevance and not waste your time going through the background of knowledge. You have a specific question and you ask it, explain to me this. And it stays on task and it stays on topic instead of taking you down the rabbit holes of the history of this language is this. And things you need to understand before it gives you what you want to know.

Edgar Dyck (29:54)

Yeah, and that's also how you could, I think, leverage the stuff in your business, like create documents and upload them to the AI. Like we talked last time, and we will also have another episode how you do this like safe and privacy within your company. There are definitely ways to do that, but just like have the option to set up

just a lot of data, which the language model could have a look at, to then train new people, train staff, train, maybe just reskill or upskill your existing workers with processes they just didn't know yet and you haven't had time to assign someone to teach them. So there's definitely some room for that.

Svetlana (30:47)

So you're basically referring to like upskilling workers in, or even training, or like maybe onboarding. You could take it as far as onboarding, yeah.

Edgar Dyck (30:52)

Yeah, if you see what I've experienced, I had my personal teacher. He was explaining exactly what I needed to know and giving me all the explanations that I wanted to know to understand the concept. So if I take it a step further and give the AI a bit more autonomy to have some goal in how to teach people.

you will, you are already there, that you can give the AI content, which just can read, and then try to repurpose it to teach someone that same content. I think it's a huge opportunity here for every business.

Svetlana (31:42)

And I think that you said something that it's maybe important to emphasize a little bit. So it's not about just you or the employee driving the conversation and then figuring out, oh, well, I'm expected to know what I need to be asking the system, but you could actually set goals. There's a way to say based on this conversation of what this employee or...

whoever needs to know here are the goals for this conversation. So be able to make sure that you touch on all of these things, but guide the user at their own pace through those conversations. So if they have specific questions about a specific topic, they can linger on that topic for as long as they need to get those topics right. But the model understands that it's one of 10 objectives that you've set for it, so it knows to move on to other topics. So there's some instructions that you can give it

Edgar Dyck (32:17)

Yeah.

You could even think about how to check if the goals are reached, like give the model tools to use to check stuff. And also what I wanted to say, I think you said, you can make it achieve goals at their own pace, like learn at their own pace. And it's not only at their own pace, but also in the language they understand. So if I...

If I say, okay, I don't get the explanation the way you are giving it to me, maybe just switch the explanation up a bit. And the LM is totally capable to do it today. So, because different people learn differently. But we have standard curricula for like ever.

And now you can go ahead, have goals and topics which you wanna teach, like contents you wanna teach, and everyone can get its own curriculum. It's a huge opportunity.

Svetlana (33:40)

So education is, within organizations, key. And I think, again, bringing it back to operational costs and everything like that. So think about time it takes to onboard these individuals and scheduling and then allocating people to be able to be available to answer those questions. You could streamline a lot of that. And guess what? Once they complete the onboarding training, you can still give them access to the same tools. So if they have questions weeks from now.

have access so they're not bothering HR or trainers for that. That's their proprietary or basically their own personal assistant for onboarding. They can go back and ask questions anytime.

Edgar Dyck (34:22)

And that's basically also what the big win is. You don't have to then like, for example, you've learned something, something else came in between. The person couldn't work on the topic and then he basically forgot everything. It doesn't even have to be a year later, like a month later.

And then you can go ahead and say, no problem, the teacher is always there. He can always teach you again. So, and that's also where like to just repeat the voice example from the beginning of the episode, that understanding voice and the emotions in the voice in a teaching situation would be really valuable.

Svetlana (34:46)

Hehehe

So I know, and I know we've been touching on a lot of productivity topics and graphics and just training, but something that I think a lot of organizations do is also creative in nature. So marketing, websites, search engine marketing and search engine optimization. So what are your thoughts on how does LLM start to deliver some of that value?

workers in succeeding or doing more or optimizing their processes? What are your, I mean, kind of high-level question, but what are your thoughts on the creative aspects of these large language models?

Edgar Dyck (35:49)

So my take on this always was something that's really deeply rooted inside me and that's I am furious how much human brain we waste on mundane repetitive tasks. And we all do, like think about your workday, think about it. How much of your workday is really creating value and how much is managing stuff?

necessary to be able to deliver the value, but it's not really adding anything but work. And I think we all have still the higher, like I think the higher you go, the more you can debate stuff. But I think we have a lot of people doing a lot of so-called like bullshit jobs, which could be automated to a really high degree if not completely.

to then use that brain power, which oftentimes also really educated people, to use that brain power in some fields or topic where it's basically creating value. And yeah, that's my high level approach to all the AI in business stuff is how can we get people to do less mundane tasks and spend more of their time creating real value.

Svetlana (37:18)

And when they are dil...

Edgar Dyck (37:18)

And we need that.

Svetlana (37:21)

And I think you've said it really nicely, and it just kind of sparked the thought in my mind that you're freeing up them employees to do more work that they deliver value. And that's why they came to your organization in the first place, because they wanted to create an impact, they wanted to make a statement. And so you're not just basically improving your business, but you're kind of...

creating happier employees, because they're there to create an impact. Then if they're able to do that by doing their best work and you're freeing them up off of these like mundane reporting tasks, I mean, isn't that a win-win?

Edgar Dyck (38:01)

Yeah, definitely. And it goes even further. Like for example, I have one customer, I have to program some stuff. And I have the whole solution in my head. Basically just need to program it and test it. Yeah, just. Like I've had the solution, I've had time to think about it. And I have a solution, now I have to test it, and then I have to think.

Svetlana (38:16)

Hahaha.

Edgar Dyck (38:27)

And that's like, people that like, oftentimes you see videos of people programming and hacking stuff in and it's like, code flying around and stuff like that. It's not like that. It's mostly like, you think about a problem, you have a solution, a possible solution, you have to try it out. You type it in, you have to program the whole fucking solution. And then you go back and say, huh.

Svetlana (38:46)

Hehehehe

Edgar Dyck (38:53)

this didn't work like I expected it to work. And then you think again, which costs you some minutes, and then you go back to programming. And that's basically why programming takes so long, because the more complex the solution gets, the more developers you need, because every small thing in and of itself needs to go through this process of thinking, trying, thinking, trying. And of course, the more expert and the more experience you have,

through this process and the more often like the first idea might already be the right one but still you have to program the stuff. I have to look for my language a bit. I am, I am. I'm a passionate person like in general but yeah. So but what I wanted to say is like

Svetlana (39:34)

You're really passionate about this topic, I can tell. Ha ha.

Edgar Dyck (39:47)

You have to imagine how much more you would be able to do if what you're thinking about could be processed quicker on the PC or like could be even thought like alongside you with an AI. But yeah, on the high level, how can we create more value in the same amount of time? It's always like that.

Svetlana (40:13)

But it's also creative, and I think we maybe touch on this as well, because I think folks may not realize that what you're talking about is already a function of these large language models. You could prompt the system to write a lot of the code for you. Quality, I mean, there's still, again, human in the loop that's very much required to evaluate the output. But the other day, I was kind of brainstorming some ideas with my team. And again, very non-technical in nature, so I don't write code.

Edgar Dyck (40:29)

Yeah, definitely.

Svetlana (40:42)

So I needed that technical person on the other side to evaluate, but I was like, hey, let's do a quick proof of concept. Like, could we do this? And we're taking like some complex diagram. And I was like, I wonder if it's gonna convert it into like Python code, like that I could just give to one of the developers and we could like develop a solution, you know, or, you know, use it for another use case. And I was like, I wonder if this is gonna work, but I need you to like evaluate the code behind. He's like, yeah, I can do that.

But you could, so you could take, and this is kind of going back to that multimodality, you could upload the diagram, like back of a napkin type of sketch and say, convert this into logic and then Python code that I could use for my application. But you could also use it to QA your code or you could even write code based on a specific prompt. You're like, here's what I'm trying to do. I'm trying to achieve this following result, right? The code and everything like that. So...

Edgar Dyck (41:27)

Yeah.

Svetlana (41:36)

what took you multiple iterations of trial and error to figure out, you know, it could save you quite a few hours. And then in writing code, and I think we spoke about it before that you rely on AI for some of your work as well.

Edgar Dyck (41:53)

Yeah, definitely. I couldn't live without. I had it... One afternoon I was sitting in a hospital cafe waiting and didn't have internet access and I didn't have AI and I was basically lost. Not lost, but I had to type out everything myself. Like, who does that? Like a caveman.

Svetlana (42:17)

Oh, it's so funny. No, but if you think again, going back to the productivity gains, right? So like if you have access to some of these tools, so, and then let's say you've streamlined or helped save one hour of development, and then let's say you have a team of five. So on a weekly basis, you're saving five hours for these developers to do more work, or you could put them to develop other projects, or, you know, create different types of business value.

Edgar Dyck (42:41)

Mm.

Svetlana (42:45)

Again, think of AI and not just like an add-on tool in addition to existing, but it could truly create so much value and so much efficiency. And just by the nature of the work that they do. So code, and we spoke about decision making, data analysis, and all these other things. But I want to even tap into maybe a little bit more on, you know, going back to the creative aspects. But, you know, what are people doing? Or do you know of any use cases for using Dali? Or?

kind of the visual aspects. So, ChadGPT again.

Edgar Dyck (43:15)

Mm-hmm.

our logo, for example. I create the AI boardroom logo, like at least the first version, because oftentimes if I have to think about a logo, sometimes I have an idea, but oftentimes I don't. And just typing in like, hey, this is what we have, can you come up with some ideas? It's already really useful. And also I was trying to come up with an elevator pitch for a new company.

And I did make three elevator pictures and thought, hmm, let's see what Chechiepedes has. And then I gave him my pictures and said, please, that's what we do. That are my pictures, please create new versions. Like do not like work with what I have, but just come up with three new versions. And all three of them were better than mine.

Svetlana (44:14)

Hehehe

Edgar Dyck (44:17)

And that's how I use it often. I gave it like my general idea and what the context is and just try to get some ideas from it. And you can do this for your postings, you can do this for your...

for your briefings, you can do this for a lot of stuff. So it's not a ready to go solution. It basically never is. But you have to think like, if you would have had a human doing it, you would have given him the same context and he would have come up with an idea and then you would iterate on that. And that's the same how you would use the AI.

Svetlana (44:56)

Yeah, and maybe part of the process too, there's probably some user research or asking opinions and things like that, but because it's trained on that knowledge, it kind of streamlines, I apologize, even that aspect of it for you. And then you go back and you iterate on the, you know, whatever it is, whether it's designs or texts or versions of texts, and you're going to ask more people. So again, it helps to streamline a lot of these tasks that we take sometimes for granted and just kind of gives you an input, but it's based on data.

based on data that has been trained on. So it doesn't just come up with a specific direction just because it feels like it or it has an opinion about something personally. Again, everything that AI can do is based on data. So that's something that I want to emphasize. So that's why I think you're finding the quality of its recommendation pretty good.

Edgar Dyck (45:51)

Yeah, and then that's also like, I think we have to touch risks in that too. So, and data and bad data is a risk in that case. So it's like always, I say bad in, bad out. It's not as strong as the original saying, but yeah, if you have bad data, you will get bad results. And that's.

often times how it goes, like it was like this before and it's still like this. And yeah, that's something you have to be aware of.

just having more intelligent systems does not make, it can make use of more data which you would have gotten at all back, I think in your case, it's like that, you used data which was basically useless for 100 years and now you can capitalize on that. And that's...

Svetlana (46:43)

Mm-hmm.

Edgar Dyck (46:57)

That's something that is possible, but still you have to be careful with that.

Svetlana (47:06)

Yeah, I think there's a lot of tasks that are related to data cleanup and everything like that. So you shouldn't just package your data into some unstructured way and then just assume you're shake your hands, wipe your hands and then be like, it's ready to go. I think that there's a lot of against the practices that experts know data engineers have a good.

understanding it for how to clean up that data, how to look for biases. There are even tools that are built to automate kind of evaluation of your data to make sure you have representative data so you don't have again biased data that's overly representative of specific groups or maybe interests and quite subjective kind of to your organization. So there are things again in the process that you could do to help eliminate some of it, but

Edgar Dyck (47:43)

Mm-hmm.

Svetlana (48:02)

And does it have to be perfect? I don't think so. I think there are going to be things, again, edge cases that you're going to uncover, which I think you just have to get comfortable with considering some of these AS systems and the data it's trained on is like good enough. And it's ready to go because users will give you input and they will highlight, um, you know, things that are, again, those edge cases. And the reason why I'm emphasizing the word edge case is that you don't want to

Edgar Dyck (48:17)

Yeah.

Svetlana (48:31)

do what I think Google Tai or Tay did when they released their chatbots preemptively without evaluating the data it was trained on. So it started to make sexist and racist comments back to people. So I think that there's, again, some of these automated tools that look for those types of things in data, the output. So you do want to get your solution at least 80% curated against some of these, you know,

big scenarios that you know, so potential issues that you can uncover. But the 20%, leave it up to the users to help solve or help highlight it back to you so you could solve them at a talk basis.

So.

Edgar Dyck (49:17)

Yeah, and also what I want to emphasize on that, of course, it's definitely a risk that these systems aren't 100% right all the time. But I want to emphasize that they already bring value even at like 80% or 85% of times when they are correct. So you just have to, oftentimes it's with a human in a loop, I think at some point we will get systems who will evaluate other systems.

But yeah, that's where we have to be clear. Being not 100% correct all the time doesn't mean that it doesn't have value. It has a lot of it. So imagine if you have like 95% or 90% of your support requests solved by AI, what would this free up in your company? So that's how you have to think about it.

Svetlana (50:19)

Yeah, and I think...

Edgar Dyck (50:19)

Yeah. So sorry, just for the bias stuff, there are toolings which you can set in front of a request that find malicious requests in the first place and also filter out malicious answers, stuff like that. So, yeah.

Svetlana (50:40)

Yeah, there's definitely a lot of value, I think, added to the organizations. And I think there's ways to quantify that. But you evaluate the accuracy and I think that what you mentioned, it's not going to be a hundred percent, but something is better than nothing. Right. So even if it's able to help streamline, you know, 90% of the tasks, but you hear a lot, there's human in the loop that's required in a lot of what AI currently does. And so it's.

It's a value adding technology to the organization, but it's also going to help with change management because AI systems are not perfect. Hey, we still need you. We still need humans in the loop to help guide these solutions. So, and that's how those AI systems get better is by human in the loop feedback. But if they're able to streamline or save five hours a week and do that work really well, and...

the remaining 10% is something that the humans need to evaluate, you're still saving a significant amount of time of your employees. And I think that's value added.

Edgar Dyck (51:50)

And something really interesting from that is also, and that was something even some altman from OpenAI and the OpenAI team was surprised with is, they thought even at least in coding that it would, it might cost people jobs. But the result is they just do more code and they even hire more people because now they can do more stuff they couldn't done before because it was not like

ROI.

there was no ROI on that. So now you can do things at half a cost and suddenly they are profitable. And that's why, and I wasn't thinking about that possibility either. So for me it was also like, can I do my job in five years from now? Or will it be necessary for me to code anymore to hire someone to code? And the result is, yeah, you still need the human in a loop

of like firing 90% you just do 90% like a lot more than you have done before and you do stuff that just wasn't cost effective but now is and then that's...

That's something that got me personally really exciting because now I don't have to tell my staff we install AI in our company and our processes and we don't need you anymore. But we can install that and say, hey, you can do more with less. How about that?

Svetlana (53:28)

And I think, you know, I'd love to maybe spend an episode even talking about use cases, but what you said there and something to think about, right? So if you're able to free up development tasks, let's say by 50%, guess what? You could resurrect some of the old use cases that may have not been, again, as profitable before, or you thought, oh, it's way too much cost to actually do it with real rule-based or traditional ways of doing it.

Edgar Dyck (53:46)

Yeah.

Svetlana (53:57)

Well, why don't you evaluate it with AI now? So does it make it more lucrative and is there a better ROI now? So those same people who you've freed up from development can help build proof of concepts for you and deliver more value that you thought previously was not possible with AI before. So, you know, something to also think about this instead of letting go and decreasing those operational costs by letting go employees, could they do more business value adding work? So.

again, driving that top line business growth by building new solutions, delivering value to your customers in new ways. And so just something to think about. And I thought we need to touch on that before.

Edgar Dyck (54:29)

Yeah.

Yeah, maybe we also, after that general business stuff, we might even do an episode for startups, I think. Could be also interesting to how you, as a startup, can now do business models, or like start at least business models, which you were just like to, yeah, which were just not cost effective, and now with AI might be. So it's definitely a really interesting topic.

Svetlana (55:04)

Absolutely.

Edgar Dyck (55:06)

I would love to, after like, we have a lot of theory, but from your perspective, what would you actually need? What do you actually need to start an AI project in your business?

Svetlana (55:25)

So I would choose a business function. So if you're asking about like, how would I prove the value behind this? So again, with the perspective of, can this deliver value? What process in your business can be automated? So for example, if you have someone that drafts a lot of requirements or does a lot of research, could you use AI to help with that? The drafting piece. And again, you could upload examples of these documents.

and feed it specific information and then do that in half the time. So I would say, you know, first of all choose a business function that tends to be quite repetitive in nature, you know, so again, like these are these tend to be You know research-based marketing maybe brainstorming ideas drafting requirements Reviewing documents like for example legal they have to look at different agreements

relationship management before they invoice or they speak to their customers or review, you know, marketing material or anything like that, they look at their kind of partnership agreements. So if they're looking for specific questions, could you, if they do that a lot, could you streamline a lot of it, maybe referencing brand guidelines for design, and they want to make sure that they're kind of abiding by certain principles. So what things in your business and again, some of it could be core,

some, you know, let's say marketing agency, you're going to look at the core of your business. What do you do a lot of that is that tends to be repetitive? And that would start there. So maybe it's drafting emails. All you do is like in sales is you draft a lot of emails, copy and you personalize it to make sure it's curated. Well, could you connect that to a CRM? I think we talked about that before. It's like, you probably collect notes, post every conversation that you have with your clients. So could you personalize and automate that email drafting?

on, you know, for sales and then really free up their time to reach more clients and have more live conversations with people. So I would say again, start with a business function, look at a repetitive process and then see what either existing systems that you use like CRM or email that you could plug this kind of large language model, this generative AI functionality in to help streamline tasks.

And then I think an important consideration is not to automate it end to end, but still have a human in the loop, because those AI systems are not yet... I don't want to say 100% again, as you mentioned, accurate. So before you press send or before you press submit, there's definitely something that some QA that needs to be done, which is why you need to have people who are really knowledgeable to evaluate that output. Yeah, and I would

start to start there.

Edgar Dyck (58:15)

Yeah, but if you, I think if you, thank you first of all. I think if you have something like emails, it always depends, right? So all the, like the amount of control always depends on the possible damage. So in sales, if you have like high ticket sales, how damaging would it be to have like a mis, like a misinterpreted email?

Svetlana (58:29)

Mm-hmm.

Edgar Dyck (58:44)

put the contents in an email and then sent on customized offer. What would the damage be if this goes wrong? And like 5% of the times, 1% of the time, whatever. Benchmark this and see and then decide, is it worth to sit there with a human in a loop? Or do I just swallow the pill and have some bad emails?

Svetlana (59:10)

Yeah, no, I think the risk aspect of it too. And I think it depends again on your industry too. So something to think about, I have experience working in healthcare where that's not even like a non-negotiable. You can't even have errors like that. So definitely before you put anything, you can't automate processes end to end. There must be a human in the loop. Where again, where stakes are high, maybe banking, financial industry, highly regulated

Edgar Dyck (59:22)

Yeah, definitely.

Svetlana (59:37)

industries.

You definitely want to have some governance structure to evaluate not only the overall model, but maybe have some checks and balances in place to make sure that whatever is done across the organization meets certain standards. But again, not everyone's like that. And so I think you've mentioned startups or small organizations, like small businesses may have a lot more flexibility. So I don't want to scare everyone that this is like you must have all of these things in place. But.

If you are part of a highly regulated... Yeah, exactly. Know where you're playing and what the stakes are if you have an oops moment in your business. So, yeah, make sure you have kind of those checks and balances in place.

Edgar Dyck (01:00:06)

You know your game basically.

Yes. Yeah, and if, just to maybe plug it here, if you have anything you wanna discuss, you don't know really how to solve it. Yeah, Svetlana is already doing active sessions as far as I know. And like for everyone who needs a German input,

It's always free to reach out to me or to one of us, and we're happy to help you. And yeah, see how we can build a value in your company using AI. And yeah.

Svetlana (01:01:04)

Yeah. I always want to hear from you. So if there's anything that you feel we didn't really dive into, or maybe there wasn't sufficient amount of use cases, or you'd like to see more of something, I think people crave examples. So happy to do it. So I think, as you mentioned, maybe just kind of hearing from you, what do you think you want to hear more of? What can we?

touch on or what specific topics, because that's going to give us also some thoughts on how to make sure we provide the most value for you guys listening. So your time is valuable, and we want to make sure that we maximize the impact we create with that time. So.

Edgar Dyck (01:01:43)

Yeah, totally. So yeah, that's it for today's episode, I would say. We're already an hour in. Like time flies by, right? So, yeah. Yeah, like for me, it's like kind of a youth dream come true with all the possibilities that are emerging right now. So I keep talking as long as you keep listening, so.

Svetlana (01:01:54)

Oh, it does. Yeah, no. It's a passion that, again, both of us are really excited about this topic. So I think it feels. It feels.

No, this is great stuff. Well, I hope everyone got something out of it. I hope anyone listening had their aha moment today. And as always, we're looking forward to seeing you guys tuning in next week.

Edgar Dyck (01:02:15)

Okay.

Yes, thank you also very much for all your time and listening to us to the end. And yeah, put down in the comments, write us an email if you want something else, if you have topics to discuss. We are always looking out for that. And

With that said, I wish you all a really good week, or a really good day, wherever you hear us or see us. And yeah, last words as always.

Svetlana (01:03:00)

Yeah, all right. Thanks so much.

Edgar Dyck (01:03:06)

Bye bye.