So Founder's Guide to Building with AI.

Introduction

You might have seen this graph, so why for every founder, every innovator, AI is so important to look at, right? We've seen a few months back GPT beat TikTok as the fastest adoption rate, fastest app to reach 100 million. 1And the difference between the previous apps that reached that level of reach was that with GPT, you can actually embed it. It's an API that you can embed it into different products. So it will permeate a lot of other applications. And we're seeing this.

1And so this slide is just to show why it's so important that we look at AI if we're looking at innovation, because it can permeate really all different industries.

The Speaker's Background

A little bit about myself. So I've been looking for the last seven years at AI and how to embed AI into products, either as a product manager or a founder to startups that were products based on AI.

I've worked with the founder of Ubuntu managing the strategy for AI and machine learning. I've also led documentation at Kubeflow, which is a product, which is a machine learning operations project that came out of Google and is now part of Vertex AI. So I've been thinking deeply at how to leverage AI to create products that are differentiated and that bring value to the users.

AI Building Blocks

And the first part about that is, what are the building blocks? With all this range, all the things that are happening, how can I think about building blocks to embed into my products?

I just last week launched an article on this, which I'll share the link in the end, so you can look at what's there and deep dive, but basically we can separate the solutions, the building blocks into four types, right?

The output being text. It could be code, LLMs, agents talking in a text format. It could be audio, image, and video.

And you've probably been inundated with all different forms, right? Like videos generated or images with Dali and Midjourney and all of that.

Differentiation and Strategy

1So if you look at these different building blocks, then the next question is, what am I going to build that is going to be differentiated? How am I going to look at so much fast innovation? Things are changing. How am I going to be different from everybody else in the market?

And to look at this, I'd like to bring this kind of layers of AI, right? And we have a data layer, infrastructure layer, model layer, and then application layer. And this, I think, is a useful framework to think about where am I going to compete?

Am I going to compete at the infrastructure layer, at a model layer? Am I going to have better models than everybody else? Am I going to build better apps? Do I have better data?

And so as a founder, whether you're in an early stage defining a product or at a later stage and you have a company with 200 people and now AI shows up, what do you do? And so the main goal of this talk for me is to kind of go through this exercise of thinking strategically of how do I compete? Where am I going to put my money, my efforts, my resources, the time of my team?

Competing with Models

And if we look at models, it's If it was like five years ago, maybe it was good to invest in models. Now it's becoming very, very hard to compete. It's a deep tech race.

And you might have seen this, that in Hugging Face, there's already, it crossed a million models. There's an abundance of models, which are very hard to compete. They've been trained with a lot of investment. And so if you are starting today or if you're just getting on this journey, it's maybe competing on the model and starting the model from scratch, probably not a good idea. Maybe go to Hugging Face and see what's already out there and build from that.

Practical Use Cases and Approaches

And here just to bring back to practical use case. So in my last startup, I was trying to leverage AI text-to-speech for the publishing industry. And the thesis was that audio books were very expensive to produce and that I could use text-to-speech to make it a thousand times faster and cheaper.

But I didn't start with any special text-to-speech model than anybody else. So what I started with was looking at the existing models. And I went to Google, Microsoft, and Amazon, and NVIDIA. And OK, what are the best models out there that I can leverage?

I built some automations to leverage those models and increase quality. And I used the same infrastructure as everybody else, AWS. And the thesis was that by doing that and by working with traditional publishers, I could start building a data set and I could start building some internal models. And so you can start...

Maybe starting with differentiation on models and data is not a great way to start, but you can eventually get there. Infrastructure is at this moment undifferentiated, and it's a money game. How much money do you have to invest in NVIDIA chips? And we've seen Zuckerberg say that he's going to invest, I think, around $9 billion in 350,000 H100s. And Elon Musk kind of putting $10 billion to rent Oracle's hardware.

So you can't compete in this race and have much better infrastructure than anybody else. So if you're a founder, if you're a startup, and you are in these type of games that are about money, okay, how do I invest in infrastructure? How do I get the technical talent? It's interesting to look at where the money is going from a venture capital perspective.

Investment and Industry Trends

We see that VCs have invested a lot more in generative AI from 23 than before. Of course, it's kind of a nascent field. But the interesting thing is 71% of the capital is going to model makers. but it's a few tickets to the players who can compete at that level.

And so probably, If you can't compete at this level, it's best to go into the application layer. And we see that 7% of the money is going to applications with proprietary models, but actually 17% is going to applications without proprietary models. So people who have found a good use case who found a good use case, and they leverage existing models, and they know very well their industry, and they're just going for it without wasting so much resources in trying to compete at the model layer.

The Role of Data in AI Strategy

And then we come to data. So traditionally, you wanted large volumes of data to train your models. Now with things like RAG, retrieved augmented generation, you can use not so large pieces of data to actually inform existing models and come up with specialized output.

And you can, perhaps if you're, for example, an healthcare provider and you have patient data that is proprietary and that you can leverage to come up with something really unique, then maybe you can come up with a nice solution. One of our clients at Altar is actually helping clients that have this data.

So it's building a solution for sales teams. where a sales manager can input all the information about the company. And through a RAG, it creates a model that a sales rep can chat with. Or you can actually generate decks for training of an internal sales rep.

And so that you can leverage data or build solutions that leverage data. And with all of this, probably you'll be competing at the apps, right? So finding product market fit, you know, finding a great use case that you uniquely understand and building an MVP, testing it with users, same old, same old.

AI in Startups and Evolution

Now, and there was a statistic about like 40% of YC companies being AI ML, You can start with a thin layer that is on top of the existing APIs, prove the concept, prove that you can make money out of it, and then you have a business case to fundraise and then eventually thicken your tech stack.

AI Agents and Efficiency

Yeah, a little bit about agents. So agents, you might have seen multi-agent systems and how it's a bit crazy what's coming from an operations efficiency. And I wanted to share here a use case.

I don't know if you've seen Klarna. It's a FinTech solution and they've implemented internally a customer support agent system. And in one month they had this bot have 2.3 million conversations, handle two thirds of their customer service interactions. So basically doing the job of 700 full time employees in customer service across 30 plus countries.

And they've reduced average response time from 11 minutes to two minutes. So we see already solutions in the market that are having a tremendous benefit for the company implementing these solutions. And it was kind of one month after implementation, Imagine the amount of resources they've saved and how, as a large organization, they're looking differently moving forward at operations.

Conclusion

Anyways, to wrap up, because maybe I'm on time. What's a winning formula in this time?

1If you have proprietary data, or if you can come up with specialized ways to fine tune existing models, then your data plus models can be a differentiator. Then user experience, how are your users going to interact with your solution? Distribution, which has operations in it, so you can use agents to improve distribution and perceived value to customers over the customer lifetime.

Multimodal Applications and Future Directions

And just to wrap up, I wanted to also touch on multimodal apps. You might have seen this week the GPTO launch, and it was pretty cool how... You know, we started with building blocks. OK, there's text-to-speech, which is kind of separated from speech-to-text and from LLMs kind of understanding the world and replying. But now it's kind of all together, all merged.

And GPT-4.0 will be available via API. It's going to be interesting how the API works and how you can. So it's O for Omni. So it's kind of multimodal. And it's going to be interesting how you can embed that into your solutions and come up with really creative products and differentiated products.

If you're interested to dive deeper into that article, you can scan it now or come to me later.