Transformation of the Week
Nicolay Gerold

Transcript

Lusk: Hey, Nicolay, thanks for joining me today. Congrats on winning the Transformation of The Week, tell me a little bit more about your background.

Nicolay: Thanks, Chris. So, yeah, I’m running an agency in Munich, Germany. We are focusing mostly on doing LLM implementations. Also added a venture builder last year, so we basically use the agency to find interesting problems in different domains. Then basically build startups for it, finance it through the agency, and either sell it or keep it as a cash flowing asset, or fail at it. As in most of the cases, it’s going quite interesting. We sold the first one last year. Now building up the second one and hitting some nice MRR at the moment. And I also have my own podcast on how AI is built, so it’s quite interesting to be on one for a change.

Lusk: So tell me a little bit more about the types of businesses that interest you.

Nicolay: I think LLMs in general as a technology, it’s really well suited for more like transforming different pieces or different formats of content into each other. And among this is basically like going from unstructured to structured. And I think with that, since we are focusing on LLMs, it’s a lot in document heavy industries. So a lot in banking, insurances, healthcare, where we really see a lot of values. Because a lot of the like assets are in unstructured data and you suddenly can make it workable and usable by other types of AI, with which you then can build really interesting and more complex systems, which combine basically LLMs for the information extraction, then classifiers so more traditional, traditional models and ML for the analysis.

Lusk: So what are some big challenges that you’re trying to overcome right now in Gen AI?

Nicolay: The biggest issue in a lot of cases is reliability, especially in enterprise context. How do you actually make it so humans can trust it, but also how to get the accuracy high enough, like the task completion rate, that it’s complete automation. And in many cases you probably won’t reach that, but you will still have a human in the loop in some way.

I think like the second issue all of us are facing in terms of AI is like, okay, how do we separate the signal from the noise? Because there’s so much news popping up all the time, and so many new developments and new breakthroughs that you actually have to be like, pick and choose your battle. Like, what do you want to investigate further? What is actually really like worth putting into an implementation? Or like when do you think a technology is mature enough to put in an implementation?

And it’s like an issue you’re facing constantly with Gen AI because the space is evolving so fast, And every tool, basically you have to assess it from new like. Do you trust the development team behind it? Do they show good progress? Also, like how many open source supports do they have or how many maintainers and things like that. And in Gen AI, it’s really challenging because most of the tools have been around for like a year or two.

Lusk: Now you know the AI makerspace motto is build, ship, share and a couple of weeks ago you shared a post that blew up, you got a ton of new followers. What was the post about and why do you think it resonated with so many people?

Nicolay: So I think, like I hit on a really like a top of mind topic at the moment. So I wrote about like basically RAG is dead, long live RAG. And basically went into a paper and a benchmark and added like some things we are seeing in practice or in implementations we have done. I think like why it blew up, because it actually reflected the opinion of a lot of people who were doing implementations, which is probably the group which like, recommended the post.

The hook was really like clickbaity at the same time, like RAG is dead, long live RAG. So it got a lot of people attracted to it and a lot of reshares, a lot of likes, a lot of comments, a lot of of engagement at once, which was like an interesting experience to have, to be honest. Like I had to turn my phone off for most of the workday to actually be able to focus on some work.

Lusk: So what words of wisdom do you have for people that are watching this who are considering a move into Gen AI?

Nicolay: I think it’s never been easier to get started because generative AI really like equalizes a lot of stuff. It’s way easier to build all software applications and to code. But also you don’t need any training data anymore to get like an AI model into production because you have really good base or foundational models to build upon. At the same time, everyone is at the moment just figuring shit out, and there aren’t many best practices or established processes yet, which you would learn through years of experience.

So the guy who is having like ten years of software development experience in Gen AI stands on the same footing as the guy who was just coming out of college. I think like really picking a topic, trying out a lot of stuff will get you up to speed so quickly and at the same time, because the technology is moving so fast, cycles are way shorter.

And what used to apply to like GPT 4.0, GPT 4.5 like to the regular models or like instruction fine tuned models, doesn’t really apply to the thinking type of models anymore. And you it’s like the same prompts, for example, don’t work. The same fine tuning practices don’t work as well, so you have to adopt them and learn new stuff.

And it’s a lot of trial and error. And the thing like trial and error is something everything, everyone can do. So basically you try out stuff, you look at the results, you do an error analysis, like what could you do better, what went wrong? And then you basically iterate, learn and improve.

Lusk: Nicolay, congrats again on winning the Transformation of The Week, where can people connect with you?

Nicolay: So I think like LinkedIn is probably best. I’m also on Twitter and Blue Sky. If you want to follow like my content, what I’m putting now like the podcast, “How AI is Built” on wherever you get your podcast. But if you have any idea, just want to chat AI, just reach out on LinkedIn go into the DMs and I’m happy to connect.