Scaling and Nurturing AI Startups: Essentials for Success

Transcript Disclaimer: This transcript has been generated using automated tools and reviewed by a human. However, some errors may still be present. For complete accuracy, please refer to the original audio.

00:00:04 – Sangeeta Gupta - Welcome to the Nasscom podcast on #AIspeaks with Matthias Zwingli, CEO and Founder of Connect AI, a company that provides AI solutions for businesses of any size and interest. Matthias has over a decade of experience in the field of innovation and technology, start-ups and business development. He also wears many other hats as a startup coach, active Angel investor, and a podcast host, and today, interestingly, he is on the other side of the podcast. 

So, Matthias, welcome to this Nasscom podcast and we look forward to an engaging conversation over the next few minutes with you. You know, I'm going to start. There's so much conversation around generative AI, I think every day you can just see a new report, a new article, a new model being launched. And I think it's important we kind of declutter this space in some ways. And I think that's really the objective of the podcast that Nasscom has initiated. And I think, when I was looking through your website, you really talked about how AI doesn't replace companies. Companies that use AI will replace those that don't. And I think the same is said for individuals, but I wanted to understand a bit about your journey into AI. What's inspired you to start your own AI company? And you know, what's it been like. And for especially because we want to take this message out to a lot of other startups that are very, very actively looking at building out their own AI journey.
 

00:01:30 - Matthias Zwingli - All right. I’ll try to put it together. For me, there were a couple of key moments. I was always more inspired by the business side, which I sometimes regret a bit. I think computer science would have been a good choice to study as well, but I was always looking to adapt to it.

The big change for me was that I initially viewed AI from a business perspective. At first, you scratch the surface and recognize its importance and growing influence. But then, I did a three-month Python programming boot camp, which really helped me understand AI better. In the end, I learned two things: I could do it, but there are people who are much faster, younger, and better than me. That was a challenge, but also a valuable learning experience.

During that time, generative AI really took off. I think this changed a lot within the AI ecosystem, especially in development. Before generative AI, machine learning was more predictable—you input something, and a specific output followed. But generative AI produces less predictable results, making it even more exciting.

This shift is particularly interesting for people like me and many others who don’t come from a strict computer science background. It allows for a mix of different skills. Of course, there’s a technical aspect that needs to be understood, but it’s still an emerging field. There aren’t many true experts yet, so it’s a great time to get involved. If you dedicate six months to a year, you may not be a leading expert, but you’ll definitely develop deep expertise in your focus area.

00:03:34 - Sangeeta Gupta - t’s interesting to see your website, especially because you talked about building for companies of any size. When we look at the Indian context, even in generative AI, the only ones making money or investing are the really big companies, mainly big tech. However, the biggest impact is possibly on smaller companies.

How do you see this dynamic changing? How are you advising smaller companies on their generative AI strategy? Or should we just be content with the current state?

I mean, it’s not really clear to me at the moment.

00:04:13 - Matthias Zwingli - Yeah, that’s a great point. I see it the same way. I think the biggest impact we see is that you no longer need a team of five data scientists or engineers to build a product. You can do it with a small team and a relatively small budget. I think it now costs about a tenth of what it used to, and you get results much faster.

AI, especially generative AI, has opened up many new possibilities for AI projects. SMEs would definitely benefit the most from it. However, it's sometimes hard to reach them, and we’ve learned that as well.

I understand why you mentioned it. From a coaching perspective, I would tell startups that they are too broad. Narrowing down the niche is important, and we are still going through this process ourselves.

What we find effective is building solutions in the customer support space, specifically for telecom companies. The idea came from the hypothesis that SMEs would greatly benefit from it. However, it's sometimes hard to sell because it depends on the readiness of the customer.

Is the customer ready for an AI solution? We see that the market is still in its early stages.

What we see happening now is that companies view AI as an interesting and important trend. However, their first step is usually integrating Microsoft Copilot or GPT Enterprise. Where we come in is when we build a generative AI model tailored to their business tasks. However, this is still in the early stages.

00:06:08 - Sangeeta Gupta - Yeah, I think the sentiment is that the excitement, interest, and conversations around AI are very high. But enterprise adoption is still slow. Most people expect it to pick up in the next 6–12 months, and it has to. I don’t think that’s a choice. But yeah, it’s still progressing slowly.

Let’s shift gears a little, Mathias, and get into the technology side of things. You recently posted on LinkedIn about the launch of ChatGPT-4 Omni and its features. Then there were Google’s recent launches, Llama 3 was released, and many other developments in AI.

As an advisor or strategist in technology, how do you guide people in choosing the right AI model for their organization? Most conversations revolve around whether to use Copilot, build on open-source models, or something else. With so many choices, it’s difficult to decide where to start.

00:07:20 - Matthias Zwingli - Yeah, we see the same. People spend months debating which model to choose instead of just starting. My advice is to begin with the easiest solution available.

When choosing a model, focus on three key areas: technical feasibility, ease of development, business value, and data privacy. These factors determine whether a task has potential, is easy to build, or is more advanced.

If you're starting with generative AI, it makes the most sense because you’ll learn a lot as you go. Choose the simplest solution first. We can always make it more complex later, but it's better to start with an easy task and build from there.

I recommend using the dominant model initially. OpenAI models still deliver the best performance. Test with the most reliable option first to establish a baseline, then optimize based on cost, privacy, or other requirements. If cost is a concern, explore open-source alternatives. If privacy is critical, consider on-premise solutions. The key is to start, gather insights, and then refine your approach based on what’s most relevant.

00:09:22 - Sangeeta Gupta - What’s your personal favorite AI model?

00:09:26 - Matthias Zwingli - So far, our favorite has been GPT-4 Omni, and we have swapped almost everything to it. Some of the issues I had were a bit crazy, right? We built Jenny and other support assistants using GPT-4.

Using GPT-4 for simple tasks is like driving a Ferrari to go grocery shopping—it’s overpowered for the job, but the performance is impressive.

And I think we always try to build on cutting-edge technology. The only issue we had was latency. Sometimes, it was not fast enough for human interaction, especially in chat and voice-based applications.

GPT-4 Omni effectively solves that problem. Another issue we faced was cost—building on GPT-4 became too expensive, so GPT-4 Omni helps reduce that burden.

00:10:15 - Sangeeta Gupta - That's a good, it leads into my next question because you know, you're seeing this whole growth of so many AI start-ups, right India, I think last year when we tracked India, I had about 60-70 generative AI start-ups, today 3X of that. And I think there are a lot more out there. So probably, you know, as a start up coach particularly, what are you seeing in the AI start up world? I mean, there are enormous opportunities, but how are they? What would your advice be to them to say, how do you navigate this cost return equation? How do you get to market? What are your spaces to build for? And, you know, don't just get caught in the hype curve, but figure out what you're really doing.

00:11:01 - Matthias Zwingli - That's a very good question. I think it's fascinating to look at it from an investor and founder perspective. From an investor perspective, that's almost the easier place to start to understand the cost of the industry, where the money is going, and what the expectations are.

If you look from an investor perspective, you have the hardware or infrastructure layer, and there you go from NVIDIA chips to Croc for insurance chips. It's probably the safest option if you want to enter a broad multimodel LLM infrastructure like Frontier model—you need a lot of capital. If you can't provide that, it's not the market for you because you'll compete against OpenAI, Google, Tesla, and others who have significantly larger financial resources.

OpenAI mentioned that as soon as you release a new model, the old model becomes financially irrelevant. They don’t get many returns on GPT-4, GPT-3.5, or Turbo models, which is a challenge.

The last layer that remains is applications. I think this is the big bet. Some Swiss investors, like Andrea Scalia, put it well—these S-curves show that we haven't yet seen massive industries emerge where AI has incredible potential. One area of opportunity is the healthcare industry. If there’s a lot of data but the current status quo isn't great, what could generative AI do?

There are many ways to approach this as a startup. What I love about building AI companies is their efficiency. You don’t need 100 employees—just a few very good ones. With that, you can go far. The rest can be scaled by leveraging cloud computing and AI tokens, which allow for near-infinite scalability.

00:13:27 - Sangeeta Gupta - So you don't see AI today as a winner-take-all world, right? Because a lot of the money is going to big tech. And then, of course, there is value in the middle. I think your point is that there is enough value in the middle, and how you monetize that value is how you should build it out, right? So don't just get obsessed with what's happening with the billions and trillions of dollars big tech is spending, but truly focus on what can solve it and what are the best resources to solve that, right?

00:14:05 - Matthias Zwingli - Yeah, choose your market a bit. If you want to compete on the Frontier LLM infrastructure models, I think there is not a winner-take-all, but there will be a few players that run it. I don't think there is too much room—you don't need 100 models. You need two or three good ones, and that's what you mostly base it on.

00:14:30 - Sangeeta Gupta - What we've seen in India is a bit of Indic language models because we are a diverse country with many languages. You're seeing a few companies building out local language models, and I think that will be an interesting space to watch. It may not be as big as what some of the big tech companies are doing, but it's an interesting area.

00:14:57 - Matthias Zwingli - Totally, choose your niche. I wouldn't go for the massive broad everything market that the big ones already do, but that's an interesting niche, right? Country-wise, many languages are difficult. So I think the USA is not the first priority to capture the Indian market. And if you are already strong, perfect. Then you have other industry niches and sectors like law, which is quite country-specific. So you can play a lot with it.

The other day I heard of a company I thought was quite interesting in the US. If you build on an OpenAI GPT-4O model, your intelligence layer is basically the same, right? You can't really differentiate with it anymore, but you have to think, what's my differentiation? I read about a company that does government contracts, so they focused on that. I think that's a smart focus because it's very niche. You use AI, and not many people can easily copy it because it's such a niche sector where you combine your specific knowledge with AI, and that can really help. I think that's the application layer I see functioning.

00:16:19 - Sangeeta Gupta - I think so. We've covered models and technology, or the computer needs. What about data, right? Because AI is what the data we feed in, right? And if you are a startup or a smaller company, you may not have access to all that data. So what would you advise from a data lens, right? Is it about synthetic data generation? Is it about other data sources they should focus on? How do you solve the data conundrum?

00:16:51 - Matthias Zwingli - Depends a bit on the area. If you're more in model training, you should solve it somehow. But what we see is even open-source models. I'm curious about this year, right? With the release of Llama 3, which is a very competitive model, how good open-source models will get. So I think that will improve quite a lot.

And then what we see as well is you don't need massive amounts of data. Mostly, it's better to have high quality, especially later on if you go into fine-tuning or the rack libraries for the models. We mostly don't train a lot of models. We see it not yet that useful for most use cases, especially for SMEs. You get a very good result if you do a very specific rack retrieval augmentation generation library specific to the task, and then you just perform and increase that one, you get quite a good performance already.

And then you can always see, does it need a lot of training? I think most applications don’t have fine-tuned LLMs. They basically just go with what's on the market. In terms of data, it depends, right? But I think if you want to go very niche, data is probably key. Find specific datasets. It doesn't have to be huge, but this can already be a game-changer.

00:18:22 - Sangeeta Gupta - The reverse question to the AI talent shortage is really the fear of workforce impact and what happens if AI becomes so prevalent and productivity improvements of 30-40% become real. The whole concern about AI taking away jobs, even if it augments them. And I think, just to build on the customer service example, Matthias, that you were seeing, and India does a lot of work in the customer service area. Obviously, AI will get implemented, and we are very clear that we should be the ones to help implement it for our global customers. But how do you think through what it does to jobs?

00:19:07 - Matthias Zwingli - It's a tough question. I think a couple of metaphors. What I thought was interesting is I read this about radiology. Ten years ago, a very famous convolutional network scientist in AI emergency recognition said that in five years, there wouldn't be any need for radiology personnel or experts. We still have them.

What we've seen is AI reached the same level as a human in deciphering tumors or malignant diseases on X-rays, but it hasn't replaced the human in the whole job. I see it very similarly. It's still a bit stupid. Even Gen AI is very high, but it's still a bit stupid. It's very narrow.

If you look at customer support, what works very well for us is that we reduced everything that is written and can be retrieved quickly from a knowledge base. In the end, we reduced ticket volume by 15–20%. But it's not that we replace people. I think it will be similar to e-commerce—it's not one or the other. There will be a shift.

AI is here, and AI will stay. It will empower and automate many tasks. Klarna, the Swedish buy-now-pay-later company, was quite interesting when they released their solution—same basic build, but they had 2.2 million compensations in the first month. I think the claim was quite IPO-inflated. I don't see that level right now, but there will be a shift.

It's a big question, and it's always hard to say because we don't really know yet. But if you look back at search, Google, or the Internet, so many jobs emerged that we had no clue about. My dad, for example, has no idea what SEO optimization is or who does it.

I think there's a lot of possibility that new jobs will emerge. It's definitely a productivity game. I see a lot of potential right now in optimizing repetitive tasks. That’s fascinating. But we are not at the AGI level yet, where you put in a Gen AI model and it solves everything.

00:21:59 - Sangeeta Gupta - I'll come back to the AGI question in the end, but I think that's a similar view we really hold, Matthias, which is that AI will automate tasks and not necessarily jobs. I think it'll become your assistant or whatever else may be there to help you as an individual do what you were doing better. Yes, jobs will get reshaped as they've been reshaped over the years. Every technology revolution has happened, and this one is much faster, hence the anxiety levels are higher. I think we need to prepare for it. The whole focus, at least as Nasscom, is that it's important to start skilling yourself. Learn to work with these tools because you cannot say it's not going to happen to me. The more efficient you are in using these tools, the more likely you are to be successful in whatever that new job role would be.

I know we're going to run out of time, but I had a couple of last questions, Matthias. Given the work you've been doing, how do you see startups and corporates working together? I know we've had a lot of good examples in India, but there's still not a huge comfort factor. The process is slow, and sometimes the startup loses patience or things don't always work out. So, you have good examples, but not a lot of them. How are you seeing startup-corporate relationships?

00:23:35 - Matthias Zwingli - First of all, very nicely put. I couldn't have said it better. I think it's exactly how I see it as well from a shop perspective. And I'm always quite optimistic. You can always have two ways, but very optimistic. I think life is more beautiful if you're optimistic.

And then to the question of startups and corporates, yes, it's a challenge sometimes, and I see why. We tried to match during the time of Digital Switzerland and did a lot of corporate-startup matchmaking with a couple of learnings from it. What we realized is if you're very early on, it's maybe too early because if you work with a corporate, the sales cycles take a lot of time. Especially if it gets more technical or data-sensitive, you're in the first loop, then there's compliance, procurement, and so on. It just takes massive time, and you have to be crucial—Is that the right job to spend your time on, or can you actually finance it? Because it's a long process and can take years. Maybe it's not like months, right? So 6 to 12–18 months is the norm, and you have to be able to finance that because no one pays you during this time.

So that's the challenge. I think they work a lot in what we've seen as well. In the beginning, we came in and said, 'Hey, we can solve your AI problems.' Now we come in and say, 'We built this for these customers, it works, you need it.' And then it's a lot easier. It's a lot of trust, right?

But yeah, I still struggle. I'm not an expert on that, but I still struggle with it. It's a tough one, and it's a lot of investment. I think it will be more interesting now because there aren't that many other solutions. Nevertheless, what we see is that most SMEs we talk to go straight to the model. And this is kind of crazy—so many are on Microsoft that their first step is trying out Microsoft's Copilot. It's a medium, and then they're a bit disappointed because it's not the Holy Grail of AI that solves all their issues.

But yeah, just push it and be very upfront. You're an expert as a Gen AI startup, an expert in the field. Corporations normally aren't, so don't undersell yourself. That's what you know, that's what you can do. If you already have first cases, work with them because that builds trust in the end as well.

00:26:48 - Sangeeta Gupta - I think that's a related question. Matthias, it's around the risks of AI, right? While there's never been a technology that has created so much excitement, there's already so much concern around what it can do and what you need to prevent. We've seen use cases coming up—harms, copyright issues, and other very genuine concerns, right?

Of course, Europe has implemented or announced the EU AI Act, and other countries are figuring out what needs to be done from a regulatory perspective. If I'm a startup, a smaller company, or an enterprise, especially a younger company, how should they navigate these regulatory risk frameworks? It sounds very complex, right?

00:27:44 - Matthias Zwingli - It totally is. So for me, on a regulatory side, I look at it—of course, there are risks, but why do big tech companies push for it? And what we see, especially on frontier large language models, is that they're not as protected as we first thought. We see similar performance on closed models and slowly on open-source models. So I think the logical way then is to reduce the open-source space. How can you protect your models? I think that is one of the main reasons we push so fast for regulations because, technically speaking, it's such a fast-moving market. How do you regulate that? You will always be too slow, and sometimes you don’t understand what is happening or what isn’t.

There were concerns about AI being used for robbery, but you can already do that with Google, and AI models don’t perform that great on it. The earlier risks we discussed were bioweapons, then robberies, and so on. They don’t talk about that anymore. So it has come down a bit, but yeah, it's still there.

I wouldn’t regulate heavily or would try to minimize regulations. As a startup, it's hard because, unlike big companies that have enough lawyers to solve these issues, for startups, it's very tricky. If you have to apply all the regulations that come into place, it's going to be very tough. I hope we're not pushing too hard on regulations. If they are right, they should allow the building of easy products. The more complicated your product gets, the more data-sensitive it becomes, and the higher it will be regulated.

00:29:45 - Sangeeta Gupta - Yeah, that's where we are. I know we're almost out of time, but last question, right? Where do you think AGI is headed? Do you see that on the horizon? Is it maybe these super intelligent agents, and humans will be the secondary guys, right?

00:30:05 - Matthias Zwingli - So yeah, you have both saying—Elon Musk says two years, and then you have the other view that is 10 years plus. It could be another, how do you say, overestimation on the autonomous vehicle self-driving cabs moment? From my experience, and that's just my personal view, I think it's 10 years. It's rather 10 years than two. I don't see that level yet. We're still very narrow. It's still kind of stupid and still very task-orientated.

I mean, it's impressive how fast it got to human levels in certain tasks, but combining it—we're still struggling. And I think it depends a bit. The big question on this topic is how do you define AGI? Is it sentient AI technology that rules them all, like Skynet, or is it a smart senior manager advisor that can solve multiple tasks?

I think it's a big question of what level, and depending on the level you choose, the time horizon looks very different. I think we will get better in multi-task, multi-modal AI. That will definitely come, but I don't see sentient AI happening yet. Hopefully, I don't know—that's a bit out of the picture for me for now.

00:31:44 - Sangeeta Gupta - And as Nasscom, what we're really seeing is a focus on adoption, right? There's so much talk, but we need adoption and scale. That is when you realize whether it works, doesn't work, its productivity, and what happens to jobs. Otherwise, it's all theoretical, right? Unless we focus on adoption, we're going to keep talking about all of this, right?

00:32:13 - Matthias Zwingli - I do agree, right? I mean, the ideology or idea of AGI and Skynet was around forever. Now, I think it's a bit sad that we have this amazing technology, and rather than testing it and seeing what we can do with it, we talk about the next being 10/20—I don't know how many—years in the future. So I totally agree. I think it should be about adoption. How can we use it? And it's always with technology. It's like in the past, right? It can be good and bad. We rate it that way, but it's out there, and you can't get rid of it.

So my approach is, let's move toward a sustainable future in a sustainable way. That is not done by prohibiting it because I don't think it's going away. So let's move forward and see how we can use it for good, how we can limit threats and dangers.

00:33:17 - Sangeeta Gupta - So, I just want to pick up the sustainable future point. Do you see AI and the ESG footprint it is creating as a possible deterrent to some of the net zero goals? Or do you see AI being used more effectively to meet those goals or a bit of both?

00:33:38 - Matthias Zwingli - A bit of both. I think definitely it's massive amounts of data and written text. So yeah, it can definitely make it more efficient. It's an interesting field. I didn't focus too much on the adaption in green tech yet, but I think if you look at sustainability in ESG, I think it can definitely help because it's a proficiency tool that helps you summarize, find clusters, and so on. It's such a massive value chain, so optimizing this process helps spot irregularities.

I think we've seen that with the latest trends around South Poland early stage. It didn't promise what it should. If you get the whole process—from selling to creating, monitoring, and analyzing—more autonomous, there's definitely a lot of potential.

Let's see. We have another technology that is almost much better, blockchain, but it's still not happening at scale. So yeah, I don't know. This one is a bit open to me. I think it has potential. What I love about Gen AI at the moment is that it delivers value on day one. We see it in the first month—we reduce tickets, create value, and cut costs. I have hardly seen any other technology grow that fast.

00:35:32 - Sangeeta Gupta - I don't think any other technology has scaled or caught the imagination of so many people across the world. Thank you so much, Matthias, for your time. I hope this was helpful for you. It was very informative for us, and we really look forward to engaging with you outside this broadcast.

00:35:54 - Matthias Zwingli - That would be amazing. Thank you so much. It was an absolute pleasure.

00:35:57 - Sangeeta Gupta - Thank you.