Capital Efficient Strategy for Generative AI Startups: RavenPack CEO Armando Gonzales (Part 6)

Sramana Mitra: How big a deal has your infrastructure been as a cost element in the past and now? It does take horsepower to process data.

Armando Gonzales: Because we don’t have access to that type of capital, we had to build a very cost-effective system. Even these days, it’s probably less than 10% of our gross revenues. Because of the type of software and infrastructure we’ve built, we are able to operate very efficiently.

Sramana Mitra: The bottom line is, you’re not trying to boil the ocean. You have a limited set of data, which is valuable data. You’re basically modeling that and delivering value based on that. You’re not trying to mine the entire web.

Armando Gonzales: That’s a good point. We probably have more than a billion documents every day. We’ve come up with a strategy to filter the content that is more relevant to the things that we need. We only focus on the relevant content.

Our strategy is to narrow down the content that we believe will have an impact on the market. I don’t think you need the entire web to do that. Having access to it is important but having specialized computation of content that matters is where you can reduce cost and deliver high quality results.

Sramana Mitra: Whatever problem you’re solving, you need data related to that. You don’t need the whole world’s data on everything. It’s specific.

Armando Gonzales: Yes. If you add content that is not domain-specific, you can confuse the model.

Sramana Mitra: How does the advent of ChatGPT impact your business?

Armando Gonzales: It has mostly impacted our business from a business and marketing standpoint. After a decade or two of nothing but headwind, now we have a little bit of tailwind. We’ve produced a lot of literature and studies. Everything is meant to be academically-researched.

From a technology standpoint, we are mostly in the business of classification of content. We’re not in the generation of content. We do use it to effectively teach our own systems about different ways that people can write content. There’s a limited amount of ways that we would have systematically thought of. Because it’s so inexpensive to produce variations, we can synthetically teach our system different ways of writing about markets. It’s been helpful in that respect.

Sramana Mitra: Can you double-click down on that? When you use generative AI to train your system, how are you operating? Are you operating with the publicly available ChatGPT or a specific language model?

Armando Gonzales: We built our own model internally. It has its advantages and disadvantages. We are testing and contemplating other services. We need to try different things.

Sramana Mitra: My two cents is, this industry is going to evolve the way PaaS has evolved in other parts of the industry. Cloud computing has a lot of PaaS that do a lot of the bottom stacks. Then startups or application vendors comes on top of those and build the applications.

This industry is going to evolve the same way. There will be generative AI platforms and PaaS who will do a lot of the stacks and people like you will come on top and leverage one of those. It doesn’t really make sense to reinvent the wheel.

Armando Gonzales: I totally agree. We had our own rack in a data center. We had our own computers. We had a technician. It used to take us three to six months to generate an archive for training. When we moved to Amazon, we learned how to do that in three days. That was a true transformative moment. Just like that, we are going to see a lot of these platforms.

Sramana Mitra: There will be a lot of very capital-efficient startups that will be able to build on top of those platforms. The operating differentiator is going to be the domain knowledge. That’s how I think the generative AI entrepreneurship is going to become scalable.

Armando Gonzales: That’s exactly what has to happen in every case. We started super broad thinking our AI would be applied in healthcare and finance. We weren’t hitting a single home run.

Sramana Mitra: Part of it is also timing. It was before the mainstream market woke up to AI. Now the market is awake. There is a lot of money going into these fat startups that are doing platform plays. That’s fine. There will be fat startups that will build those platforms. Then there’s going to be lean startups who will do startups on top of these platforms and build significant companies.

Armando Gonzales: Exactly. For us, lean is the right word. We’re proud of it. Now with compute and storage getting more affordable and scalable, we are benefiting from the economies of scale.

Sramana Mitra: It was very nice meeting you. Thank you for your time.

This segment is part 6 in the series : Capital Efficient Strategy for Generative AI Startups: RavenPack CEO Armando Gonzales
1 2 3 4 5 6