Dhaval:
welcome to the call, Martin. Thank you for joining. Tell us a little bit about your product.
Martin Pichlmair:
Okay, so I’m Martin. I’m the CEO of Write with LAIKA. And our product is a kind of creative writing tool that is using large language models, in our case, quite small, large language models to, support writers when they get stuck or when they need more text, that is influenced by their previous writing.
Dhaval:
Wow. Okay. So when the writers get stuck or when they. Interested in continuing with the style and the tone of their previous work. They can use your product including the contents, the storyline, or anything along those lines.
Martin Pichlmair:
Yes. How LAIKA works is that you you upload existing writing. You have when you get, for example, stuck in a murder mystery because you don’t know who the murder is. Funnily, we had that case twice already with users. And then you upload what you have written before and our, system fine tunes a language model With your text and then you can prompt the model to continue writing in your voice, in your using your characters. You mentioned using scenes you have been writing about in the past and very much sounding like you. Now you can do that with your own text. Or with the text of famous writers, we have, for example Dostoevsky in there and Jane Austen in there. And a lot of, all of them, of course, dead and out of copyright writers that you can also collaborate with in a similar way by asking them how they would continue a sentence, for example.
Dhaval:
Wow. So it has memory and context as well as style and the personalization built into it. So is that. Large language model that’s very different from Chat GPT 3, which would spit out very confident phrases very long phrases. But they’re also having the same style. Is that, how is that different from the large language models? You said that you have used large language models or you have you built on top of them or like, help us a little bit on how have you built this.
Martin Pichlmair:
So we’ve built this on very small, large language models. They’re still in the same architecture and come from the same family, but they’re very small because that gives us the ability to fine tune them very quickly. It takes like five minutes. If you upload , a half done book, for example, takes five minutes and you get your own, we call them brains because that’s a nice metaphor. Your own brain based on your writing to interact with. Now, of course it has an understanding of the context, but it’s not always super, like it doesn’t have an actual understanding. It can just play with probabilities of words, just like all of those language models do.
Dhaval:
Wow. Very cool. Let’s dive a little bit into your product journey, is this your first startup? Is this your first AI product? Tell us a little bit about your background, Martin.
Martin Pichlmair:
So I have a weird background. I did a PhD in computer science originally at the University of Vienna, at the tech university, and then worked in academia for a couple of years. I got a little bit, I don’t know I wouldn’t say bored, but I wanted to do something differently. So I started a video game company and then after a year started another video game company because the first one didn’t work out. It didn’t work out, but it also didn’t not work out. It was fine. It was just not meant to be a longer existing thing. The second one actually is still around. It’s called Broken Rules and makes awesome in the games. But I’m not involved anymore because I decided at some point to go back into academia. So that’s where I spent the last seven years until last year where I just realized with my partner, That we have a huge connection between what I was doing in research, which was using generative AI to create systems for video games and her background, which is writing for video games. So we sat down and, uh, started workshops during the Covid Pandemic when everyone was sitting at home. We started online workshops where we introduced writers. To, the newest possibilities in language models, using very, very clunky tools at that point. And after three or so of those workshops, we realized the workshops are always poked out, but it’s really hard to work with the tools that are there. So we decided we have to make our own tool, and that became a research project that was funded by the Danish state, uh, in the beginning. With the intention of turning it into a product. And since last November, we founded a company and turned it into an actual university spin off that is based on yeah, research that is now working on a product that we will commercialize within the next month
Dhaval:
Very interesting background. You do have like a very traditional computer science background, making you very competent in this area. Right. So, quick question. You mentioned that you launched this in November, but you haven’t commercialized. It doesn’t mean that the product has not launched yet.
Martin Pichlmair:
Yeah. We have a wait list and we have, a data with, nearly 2000 users. So there are a lot of people using it every day, but, it’s not a commercially launched yet. We’re still only free for select users.
Dhaval:
Very cool. Is this. Is this a self-funded or have you bootstrapped this whole thing? Are you intending to, or have you raised capital? And are you intending to raise capital as you move forward?
Martin Pichlmair:
Well, we got some funding from Danish State again. The program that we’re in that funded turning research into a product last year that was still in the context of my university has a follow up program that funds your salary basically. So we are kind of weirdly half bootstrapped. We have no investor, but we have, our salaries covered by the Danish State but a very low salary. But still, it’s good enough to, know that we’ll be, we’ll be around for another year at least while this funding runs. we’re looking for investment in the moment. It’s, we are talking to a lot of VCs. It is. It just takes a while it seems.
Dhaval:
Yeah. how is that playing out in this current market? Like how is that, can you give us have you done this before? And if you have, like how is it compared to the current market, if you can speak to that.
Martin Pichlmair:
So I haven’t done it before, but a good friend of mine has a very similar company, actually a very different company, but also an AI company also in Denmark. And he also has an academic background. It’s otherwise very, very different because it’s B2B and started out much bigger than we are. But it looks like they, the climate they saw two years ago is very different to what we have now. So I’m getting all my tips from him and half of them don’t work anymore. The climate is not good in the moment. Even in the hype space of creative ai, there is a lot of chicken egg problem, happening in the sense that investors wanna see. They actually want pay traction very often .They want to see some pay traction or immense numbers in weightless users or something like they wanna have proof of actual viability very early on. But it’s such a new area that you are actually creating a market. So it’s very hard to say where this whole journey is going because the whole, like AI is not super new. But generative AI is really something that is only a thing since like a year or so. It’s very hard to say where the journey goes in the moment and, like it could all still just be overhyped. Then I would understand the need for having paid traction, but it could also be that we are just opening, creating a completely new market here, and then a little bit of trust would be nicer than having to prove things too early.
Dhaval:
Yeah. Yeah. Where are you in the product stage in terms of product development? Are you close to? I know you mentioned you were ready to launch in a few months. Are you close to finishing your product development? Is is that almost there? Like, if you can share that. Cause my follow on question’s gonna be on, what were your top learning lessons around creating an AI generated product? What was that experience? What was the top learning lesson there?
Martin Pichlmair:
So I think the whole idea of finishing a product is not really how it works with software as a service anyway, but especially in this extremely fast moving area of, AI in general, but especially generative AI where new technologies come out on a sometimes weekly basis. There’s a lot of competition, but there is also just a lot of speed of development. I don’t think we will ever be at the moment where we say, now we are done with this. Instead, what we are, where we are trying to get is to a point where we can say, this is our 1.0 version and we hope to be there in the month actually. And, from then on we of course continue building. and one of the main challenges to my surprise actually, and maybe that was the main learnings, is since what we’re working with is so new, it is really, really very hard and crucial to explain to users what we even doing here, because they are very curious, but they have a very hard time understanding these new ways of interacting with the computer. And, it takes. Nearly as much time to make the product as it takes to actually package whatever you’re building in a way so that an everyday user can understand what is even going on here. So that’s quite challenging and that is something that we also are constantly in the process of figuring out.
Dhaval:
Yeah. One of the biggest challenges for generative AI products is gonna be around user experience and managing that human AI interface. What has been your biggest learning lesson in that area for your product?
Martin Pichlmair:
So we have a nearly no prompt system. So most, in most AI systems, how you interact with what a generative AI systems. If you interact with them, you do some kind of prompt engineering as they call it, so you get better and better at formulating what you wish to get out of the model. More and more precise over time and you just, it’s a learning process. You prompt them with text usually, and you just get better at describing what you want in the way the machine understands. which is an interesting interaction, but it’s actually surprisingly technical. And our approach is a little bit, since we are all about making this more accessible, these systems is our approach hides that. We are, of course also prompting in the background very much from the user and tries to make sure that you get a higher percentage of. Good results automatically instead of giving you this huge space of freedom that you have with more technical systems. But those systems exist so if someone is technically inclined, they can just use a different system. It’s just not what we are making a software for.
Dhaval:
Yeah. You mentioned something about prompt engineering and how does that play out when users are engaging with the product, do they get to like decide exactly what’s gonna be the better prompt to ask, or is that something done behind the scenes?
Martin Pichlmair:
Most of it is done behind the scenes. So the base use case of LAIKA is, you upload your materials to train with your voice, and then you just write half a sentence or a full sentence and a question mark, for example, and then you ask the system how it will continue from there. Now, of course, technically you are just prompting. But in practice it feels like you are writing and now, and then you give a little bit of control to the machine and then you override the result immediately and work in this very tight interactive loop with the system. So it’s a bit of a, yeah, more. There’s a lot of discussion about generative AI in the moment among artists that, uh, somehow do not appreciate the fact that there are so many machine generated images out there and what we are doing is very different because there is no result. That is trust. What comes out of the generative ai, it is always the author with. Their own thoughts remixed by the machine creating something. So it’s very much 80% is human authored and what is even made by the machine vanishes, hopefully. So it’s a slightly different way of interacting. It’s using the same base technology and the same post model base modalities of how you interact with a machine learning model. But it hides them and it makes the loop even tighter than existing systems.
Dhaval:
That’s where the magic is, making that loop tighter. Thank you for sharing all your knowledge. Martin. Anything else that you wanna share about the future of your product? Where do you think it’s going? Do you have a big vision around where it’s gonna be in about 10 years from now?
Martin Pichlmair:
About 10 years. Oh boy. Well, in the end I think that I actually have a big vision how, where it will be in 10 years or where society will be in 10 years. I wrote a blog post about it recently. That’s why I have. The thing is I actually think that a lot of interacting, interaction with generative models is not so much about getting results and much more about the experience. So what we are doing is we are interacting with cultural, the cultural heritage of humanity, in a live and, uh, uh, back and forth ways. So we synthesize something that uses images that are up to thousands of years old, like the photos of those images, but still human expression that is hundreds, thousands, and so on, years old, and we synthesize them into something new. And that activity is actually a very interesting activity of interacting with the, what humanity is with all of our culture. And I think it will become just something very normal for us to do. That we’re not only interacting with the way a specific historian has recorded our history, but in a much closer, again a tighter loop, interact with past images, past writing, past words, past figures, maybe historical figures and so on in a, and of course, synthesized way, but still very much informed by who humanity was in the past. And I think that just will lead to a different understanding of who we are. And that will maybe take 10 years, but it’s gonna be interesting. That is a really big vision. I know.
Dhaval:
Yeah. That’s amazing. I, I look forward to get to there as well Martin Write with LAIKA 2000 beta testers, how many people who have signed up for the wait list.
Martin Pichlmair: Well, not many more.
Dhaval:
Not many more. Okay. And you are ready to launch your product in about a month’s timeframe and, you have exciting, future ahead. I am looking forward to following your journey. Martin any last thoughts on anyone who wants to start on this journey? AI, product creator who wants to build an AI product. In generative AI space using text as their, as their tool. What would the, your advice be for them? any tools that you would recommend for them to use to get started?
Martin Pichlmair:
I think the important thing is to find a niche, because I think in the very near future, like next year, most likely, Microsoft will trust, integrate generic functions for creating text into all of their products. And then that’s just gonna cover 90% of the use cases. So if you want to get started in text, it’s like you have to find a very, very, very specific niche, maybe in the business to business sector, to be able to still be around in a few years when that will just be a standard function of word processors. So I think that is my main tip to just. Yeah, keep it, keep it focused.
Dhaval:
Now in terms of execution to keep it focused, would that be custom models? Would that be built in tune with GPT 3 or GPT 4, whatever is the next version? How would you suggest that?
Martin Pichlmair:
I do like our autonomy. We own nearly all of our stack, all of our core stack is owned by us. And that gives us the ability to define our own policies of how people can interact with our system it gives us safety. It makes us independent of the server costs of other people. Of course we are hosting it in the cloud. It is a little bit cheaper also in the moment. So this autonomy is something that works very well for us. Now, I think it’s of course, expensive to build a product, using your own stack, but I think there is will just be more and more need for diversification of what you’re doing. And we can have more diversification than any product that builds on the vanilla GPT 4 version or 3 or 3.5 or whatever.
Dhaval:
You cannot really have that niche if it’s built on top of a vanilla version. Well, thank you so much, Martin. It’s been a pleasure having you on the show. I’m looking forward to following your journey,Thank you so much, Martin. Have a great rest of your day.
