AI and User Confidence with Igor Ranc ( Product Leader, Curator Handpicked Berlin)

Andrew Maeer Welcome to byProduct Live, where we unbox everything SaaS and product. We talk with experts and leaders in SaaS and product to understand their approach to leadership and ways of working to share their knowledge and understanding. Andrew Today on the byProduct Podcast, I'm delighted to be joined by Igor Ranc. Igor is a Product Professional, and some of our listeners will definitely know Igor from his newsletter, Handpicked Berlin.

Igor's early career included roles at Daimler Financial Services and Mercedes -Benz Bank. At his most recent position, Igor was Senior Technical Product Manager in the AI products team at InsureTech platform WeFox. I'll be talking to Igor about his thoughts on AI and its application in products and considering if we're about to see a dip in confidence in the use of AI at all. What started as a side project and a hobby, Igor publishes Handpicked Berlin, a five-minute Monday morning inspiration for everything Berlin.

It's a great resource. They've got around 8,500 subscribers and counting on Substack. And we'll talk a little bit about this too at the end of our conversation today. But for now, Let's dive into AI and products.

Igor, welcome to the show. How are you today? Igor Thanks for having me Andrew and thanks for this beautiful intro. Andrew No problem, no problem.

So I think what would be really useful for our listeners, if we just put a little bit of context of your experience and your last role, you're a technical PM in the AI team at WeFox. You just want to give a quick overview of what you're responsible for there. Igor Yeah, basically I did, I just started this role at the start of this AI hype. So it was really great to be having experience working on this.

So mainly we did a lot of machine learning projects, but what I mainly focused on was sort of a chat bot where you upload certain documents and then you can basically discuss, discuss questions, sort of a Q and A on a document. And this was very classical LLM, large language model application. And yeah, that's why we are also today talking about this. Andrew Yeah, cool.

And I think a good place to start with AI, and we mentioned the hype, is actually the challenge around implementing AI models. So what kind of challenges have you seen in that role? Igor So I mean the technology, at least when you remember when we first had the experience of this chat GPT, everyone was extremely amazed, right? So this of course also then translated to all the companies, everyone wanted a piece of this.

It's like one of those, it was like sort of a new blockchain technology, right? Everyone thinking, okay, what are we gonna do now? there's, I think it's still the case, right? You've seen a lot of, there's a lot of experiments.

Andrew Yeah. Igor There's been a lot of different use cases, a lot of hype. We've seen the investments coming in. And I think the challenges that is similar to blockchain in this sense was that use cases are very hard to really pinpoint because technology is amazing.

And then a lot of companies are now trying, still trying intensively to find use cases for this technology. But I feel that the main problems are basically this inherent property of LLM models that they hallucinate, right? But they are very convincing in it. you write something, I mean, you tried it probably with ChatGPT and it will be very confidently telling you completely wrong information.

And this is extremely hard to prevent. It's basically literally impossible to prevent. So you can't be never, you will never be 100 % sure. Andrew Yeah.

Hahaha. Yeah. Igor that what the model is telling you is the truth, right? So you can be 99 % sure.

And this, especially in the applications where you're looking at B2C. So for example, we've seen this with, I think was a Canadian airline where a chatbot promised a discount to a customer. And then basically the company was liable for this. So everyone is extremely afraid.

Everyone is extremely afraid of this. So Andrew Yep. Igor That's basically the hallucinations. then obviously everyone knows about this.

So everyone is also interested in AI safety, right? So how will I ensure that my product is not gonna lie to the customer? But it turns out this is also extremely hard to achieve, right? Even with the best, I mean, we've seen what happened to Google.

It happens to the best companies with the best resources. And I think that's been like sort of Andrew Yeah. Of course. Igor a common occurrence in every single application besides the ones that are basically for the expert to check, right?

And when you say, okay, who is the expert? Who is the expert in this case? It's a very good example is software engineers. They are experts.

They know the code and for them, I think this has been a massive win. Andrew Yeah, absolutely. And I guess it's all about control on the models. And that then comes down to the prompts that people are using.

Because if the prompts are lazy, then I'm guessing then that the outcomes that the model is going to deliver are not going to be accurate. Igor yeah I mean this is what I mean you we all see it right when you try when you say to ChatGPT or now basically Claude is much much better now when you just say something quick two sentences write this do this you will get a very general response so whenever you work with LLMs it usually makes sense that you really think about the prompt so this like usually if the longer it is, the more precise it is, the more you experiment on the prompting itself, the better the final result will be.

And this is often what we, mean, in general, we want results fast, but because the model of input is basically text, right? It's very hard to think of it as a PM. You always need to think about user stories requirements. So if you would really like a model to produce something extremely good, you would need to think in very precise matters, you would need to really define things really, really in detail and that's what we usually don't do.

So usually if you don't put effort in the prompts, also results will be terrible. So, and that's exactly what makes a difference. So if you want to have a very good assistant, you need to spend a lot of time with teaching or training or prompting in this way that you're getting the results that you want. Andrew Yeah, and I guess there's just so many variations and connotations in different circumstances and the way that different people would ask different questions.

I guess it almost becomes a vicious circle that if the data is not going inaccurate enough or the model is learning from that data that's not accurate, then it's going to keep creating those inaccuracies. And I guess it would be difficult to get to a point where everything's where it is 100 % accurate. Igor Yeah, I mean, I don't know if you notice, but for example, ChatGPT they always changing the model, right? So you can have like, you can have identical prompt, but then the result will be slightly different.

For example, in some formatting tasks, it would just stop putting a dot at the end of the sentence, even if for three months, everything worked perfectly. So that's one of these main risks where you rely on third party models, which maintained and run by them. For example, this was very common in this ChatGPT wrappers, where you see that certain things can work really well, but then suddenly they stop working really well. That's in the cases where you don't control the model.

And so if you don't run your own models, then you're going to always have this question, okay, what is my supplier now doing with the model itself? So I think the most successful companies will anyway... anyway be the ones that will run their own versions of the models. Andrew Yeah, yeah, yeah.

And I guess if we take that back to your work at WeFox and we put that into a chat bot example, it can obviously be an incredibly powerful tool if you upload a data and then the bot can understand what you need from that document. But equally on the flip side, if it gets it wrong, the impact on a business could be super negative, really bad customer experience. But ultimately, as you mentioned about the airline, it then ends up costing the business more money than the bot is supposed to be saving in terms of the way it works.

Igor Yes, so that's a very good pro... I mean that's one of the biggest problems exactly, it's this confidence in the result and it's not programmatic, right? So it's not one plus one equals two, it's usually two, that's the LLMs, that's the main problem here and even with all this mitigation nowadays a lot of talk about rec, so that's retrieval augmented generation, it also has its own problem, Andrew Hahaha. Igor It is an amazing tool, but you can always, you need to always take it with a grain of salt.

So that's been, that's been clear for everyone that seriously worked with it. But of course, this doesn't mean that the technology technology itself has much more utility than for example, blockchain, because people that are really, I think the most beneficial industry of this is software development because they are able to, I mean, this is just software. development will never be as easy and as cheap as it's going to be from now on, right? So each year is just going to get better and better and I think also you will have all these very very targeted models because one thing that is also a problem you mentioned, right?

If I upload massive amounts of data to a model this also costs me money especially if I don't run it by myself Andrew Yeah, of Igor But even if you run it by yourself, it costs you GPUs, you need to render the computing power. So it's also expensive to run this technology. And that's why you also see that OpenAI, I mean, they are by far the market leader, but they are burning so much money because it's also very expensive, right? So you said like, okay, but if I don't have the use case, mean, Klarna did a lot of, at least publicly, they said we did a lot of this with customer service.

So yeah, I think everyone is trying, but there's not been a killer. Andrew Yeah, and I think in terms of the cost, think I saw recently the latest Nvidia chip, is it $20 ,000 per chip or something crazy now that it's costing? Igor Yeah, I mean, it's bizarre. Now everyone is investing, but this is more for the training part, right?

When they train this foundational models with all these billion parameters. So that's one part. So you need to be big. But the second step, think what's going to really matter is that you're going to have a lot of very small specialized models.

For example, if I have now a task for your podcast that I just want a particular way that this model will This will be, think, in a couple of months, maybe a year, it will be extremely accessible so that everyone will be able to have a little army of very smart interns, which will be very cheap to run. Especially this, if you imagine if you have somebody very smart, like extremely smart intern that comes to your business. if you think about models in this way and you tell them okay for this particular task you need to do this particular things and you repeat this this is where we are talking about training data right you expose you expose a lot of information the way you want it the model will be able to understand this and then it will be able to output what you asked it right and this will be this will be definitely changing many many jobs or many many companies in a good way especially in the efficiencies but of course there will be never a magical solution where you're gonna put your shit where you're gonna put your shit prompt in and then something magical will happen Andrew No.

Yeah, definitely. So I think when we're talking about confidence, then I think that there's two sides to this, right? So you're saying that software engineers and professionals are going to continue with the hype because they understand the benefit it's going to deliver. But then you've got the consumer and the user side where are they going to believe what's happening?

And there's going to be a situation perhaps where there's a confidence dip on that side of things, would you think? Igor Yeah, there's one thing is definitely this. But I think also the consumers will finally not even get it exposed, right? I think companies will be very reluctant to push some chatbots there because I don't know what your personal experience is, but it's also very painful to chat, right?

With your problem, because even for the user, you need to explain, okay, my ticket, I couldn't book it. Like, you know, let's say some booking .com or whatever. Andrew Yes.

Igor It's painful, right? It's painful to type all this. So you just rather talk with someone. So maybe if it's going to go in this direction, that you're going to have agents that are really good in explaining things through the voice, right?

Which we're also seeing. I think that's maybe even more interesting. And of course, people will be always asking themselves, am I talking to AI now, a real person? Andrew Yeah, a real person or not.

Igor But I think this will come eventually. We will have this definitely. But the question is also how brave the companies will be to release this into the wild. Because even the best AI safety I briefly mentioned, there is this Twitter accounts which are just basically called jailbreak, whatever.

And they always find the prompt that will break the model. it will just do the model with just... So it is always this problem. that is going to be very hard to mitigate.

Everyone obviously is trying. So I think software engineers, they're going to love it, continue to love Everyone that are also in the companies basically thinks that somebody will check over, right? Imagine there is a result and there is still a person a bit checking this. All of this will be massively adopted, but the things that nobody is checking, but the end consumer is getting it.

I think a lot of especially older companies, they will be a bit reluctant to push this out. Andrew Yeah, I guess the danger for software engineers comes though is when the AI model can create its own software language and then build the models and applications without the software engineers being involved. Igor Yeah, I mean, obviously, obviously, but I don't I mean, yeah, there's a lot of discussion in this area as well. I think we are still far from that, that there will be some self aware.

I mean, this is this this AGI everyone speaks about. And then we can just close the shop if this happens. Andrew I was going to say that's when the software engineers need to pull the plug on it and not switch it back on. It'll still be working, yeah.

Cool. look, we've had a really good chat through the application of AI. And obviously, there's some of the risks in there and the safety and actually where the confidence pieces are going to be. So we'll wrap up our conversation there.

Igor Yeah, but this won't work. When they will unplug, this stuff will still work. Andrew Why don't we just have a quick two minutes on your newsletter, Handpicked Berlin. How long have you been producing the newsletter for now, Igor?

Igor I think it's been now two years and a couple of months. And yeah, I mean, as you said, I enjoy it. It is my side hobby. And basically that's also the reason why we speak, I believe.

Andrew Yep, yep. So I'm signed up to the newsletter. I get it every Monday and it's really informative and interesting and got varied pieces of news in there. What kind of things do you include in there just for the benefit of people who may not have heard about Igor Yeah, so Yeah.

So it's basically a weekly digest of what's going on in Berlin and Germany from the perspective of tech startups who got funded things that I got interesting. What is interesting in Berlin? What's interesting in Germany? Carreered stuff.

It's a bit of a mix here. What I find personally interesting and luckily so many other people, so many other people also find it interesting. And yeah, I'm really enjoying doing it. So of course I would be happy to get.

Andrew So do the people. Igor new people checking it out and also giving some feedback. Andrew Yeah, cool. Well, let's wrap up the conversation for today.

So we always like to do some key takeaways. So I think number one is it's obviously important not to get caught up in the hype around AI and ensure that AI is adding value to your product and ultimately whoever your customers or your users are. I think as we said, it's ultimately all about the data quality. Rubbish in is going to equal rubbish out.

I think everybody would agree on. And then that dip in confidence is likely to come on the consumer side. So do people know that they are actually interacting with AI? And then how easy it is to interact and how frustrated people get.

And it's likely we're probably going to see a little bit of a dip and then the hype will come back again. But obviously, as we've discussed for software engineers, they're going to be loving the tool because it's going to make their life easier and help them be more productive. And we're have a little extra takeaway today and that is to make sure all our listeners check out Handpicked Berlin. So sub stacks the place to find that right to sign up and subscribe and then that'll hit your inbox every every Monday morning.

Igor Yes, exactly. And yeah, don't forget not to do lazy prompts. Andrew Yes, no lazy prompts. That's the other thing.

Cool. Well, thank you very much for your time today, Igor. It's been really interesting and good to connect. And we'll catch up soon.

Igor Yeah, thanks. Thank you for having me. Andrew Maeer Thank you for listening today. We hope that you found the byProduct Live podcast valuable, useful and informative.

You can subscribe on all major podcast platforms, including Apple, Spotify and also YouTube. If you have the time, we'd love for you to leave us a positive review. As a new podcast, this would really help us out. We'll hope to see you soon on the next episode and thank you again.