The Idea

I Need a Book

A month ago, my friends and I were on our way to Paraná for our annual fishing trip, and in the car I tried to explain to them what artificial intelligence (AI) is, because we'd asked ChatGPT a question. I talked nonstop for 20 minutes, thinking I was explaining perfectly how it worked. When I paused for a moment, I looked at their faces and realized that this whole thing wasn't so clear or simple to them. I don't think they understood me very well, because they didn't ask me anything else. Not a single question.

That reaction—or rather, that lack of reaction—confirmed to me that explaining what artificial intelligence is, what an LLM or AI model is, and how it works, was impossible to summarize in a car ride conversation. We all use it, but we don't know exactly how it works or what we can do with it.

If I wanted to explain it properly, I needed a different format. I needed a book.

Once back home, that conversation—with no questions—left me thinking. As I watched the fire in my wood-burning stove, the idea finally took shape. Electronic music playing in the background, a quiet Sunday, and that sense of stillness that only comes when you have nothing urgent to deal with. I felt a deep urge to share my journey over the past few years in the world of artificial intelligence. I've always enjoyed talking about what I do, but the world of software isn't the most common one, and I don't always find someone to share my stories with or tell about my day-to-day experiences. In fact, of all my friends, I'm the only one immersed in this technological world.

Explaining what artificial intelligence is and how it works is quite a challenge—there are many elements at play, many pieces of the puzzle, and the terminology is difficult and not commonly used. All of this pushes me to find the best way to share my experience using this technology over the past few years.

That's why I thought about writing a book and trying to convey the knowledge I've gained and the experience I've had over these nearly four years since it all began for me.

The ideas started coming one after another. My first book. It must be quite a challenge. I don't know where to start, nor do I know how to write it, how to draft it, or what words to use—or what tone to adopt. Nothing. I started by doing a little research and asked the AI, Claude—Anthropic's model—to give me a basic outline of how to structure a book, but that didn't help me much, because even with the outline, the hardest part was deciding what I was going to write about and how I was going to tell the story or put it into words. Finally, I told myself to start with the hardest part, which is just sitting down to write. So I went back to the beginning and started writing.

Not everyone sets out to write a book. In fact, I know very few people in my circle who've done it—barely three at most. It's undoubtedly a very complex task that I may have oversimplified. But even knowing it was very difficult, I decided to move forward anyway.

After a few hours of mulling over the idea, I decided to voice it and told my family about it during that lunch at home.

I don't remember exactly how I started. But I do remember that when I finished speaking, there was a silence different from before. It wasn't indifference—it was surprise. They asked me what the book would be about, and I told them a little about the idea behind it. At first they didn't quite understand why, but after a while they got excited.

"I can read what you write and help you," my wife said.

That's when I realized they were interested in the idea, too. My wife loves books. At home, everyone reads; books are constantly coming and going; they share literary workshops; they chat about what they're reading. The image of everyone reading and drinking mate on a summer morning in the garden is a common one at our house. I, on the other hand, have read few books, and they've been very specific. I read Everest¹, a book with huge, incredible photos showing all the routes and the history of expeditions to Everest, and the last one I read, I think, was by Adrian Newey², the Red Bull car designer, who tells the story of how he got involved in Formula 1 design. When I like a book, I read nonstop until I finish it.

I've often felt the urge to share something I'm doing or thinking, but I don't always work up the courage. I often let some time pass to see if I'm truly convinced by these new ideas. It's amazing what happens when you stop just thinking about something and say it out loud—the idea gains so much more power. It stops being a dream and starts becoming a reality. Something shifts inside your head; you can't just go back on it so easily. It's no longer just yours; others know about it now, and that commits you.

I'm not a programmer

I've always been a computer enthusiast, but I've never been a programmer. As a kid, I loved game consoles—I remember spending hours playing ATARI³ or the TALENT⁴ that my friends had. When I got older, I spent hundreds of hours on the Commodore 64 that my dad bought and used just as much as I did. Then we moved on to PCs like the AT⁵ with 3.5" disks, and over time we kept upgrading the hardware and incorporating computers more and more into our lives at home.

When I founded Wispro, around 2017-2019, I also had my closest encounters with code—perhaps the most advanced and complex ones—because I started developing the company's website and was able to access all the Wispro Cloud code repositories and make some minor changes. But that required understanding a lot about how the development process worked. Those were very happy years of expanding my mind into places unknown to me.

I've known about and used Git since that time. I've always found it very practical because it lets you keep an organized record of changes to the files in a programming project: what was modified, when, and by whom—with the ability to roll back if necessary. In other words, it's software that lets you manage versions.

Even so, my day-to-day work has always been more focused on the strategic direction of companies and business management as a whole. That's who I am—someone who understands technology, applies it, and sells it, but doesn't program it. Although I wanted to study programming or something related to computer science since I was a kid, life took me in a different academic direction, but I always felt I had advanced computer skills.

In fact, as a lawyer, three months after starting my first job, I developed a database in Microsoft Access to manage all the legal cases we had at the firm. I named it "5000" because I imagined that was an unattainable limit for lawsuits.

With that software, we ended up managing 20,000 cases across five branches in different jurisdictions. That doesn't make me a programmer, but I understood the concept and enjoyed getting my hands dirty and building things. I took courses and even wrote some code to build a scientific calculator. I love it and always have, but the time I had available had to be spent on other tasks.

And then ChatGPT arrived

On November 30, 2022—the day it was launched to the world—everything was moving faster. But with every day, month, and year that has passed since then, the pace has multiplied. AI has become increasingly proficient at writing code. For AI, it's just another language, like any other. That's how Cursor, Copilot, and Claude Code won over the hearts and minds of programmers—the first to start using these tools to generate clean, efficient code ready to be implemented.

Companies like N8N⁶ and Lovable⁷ also emerged, making it possible to develop systems and connect them to one another—something that, years ago, could only be done by a programmer—and, suddenly, the line between knowing how to program and not knowing how to program became much thinner.

Another tool that became very popular was one for creating knowledge bases—that is, all the documentation on how to use applications, the user manuals. More and more companies made them available, ready to be used by AI. The models began to respond better and better. Starting with ChatGPT 3.5, each new version was a leap forward.

No tool has ever made me feel so capable of doing things that previously depended on an entire team, but at first I was completely skeptical. I didn't believe in this whole ChatGPT thing at all. I kept working on my own projects, ignoring everything new; I always found something else to do and put off looking into what the increasingly famous ChatGPT was all about for just one more day.

It must have been about four months after the launch—around April 2023—when I started reading about the tool's potential. That's when I said to myself: Hey, this is serious business. What's going on? Why aren't I using it?

So I entered my first prompt. I remember asking something like, "What can you do?" and its response half-surprised me—something along the lines of, "I'm an AI that can answer any question you have." It didn't strike me as magical. But it was enough to keep asking it lots of questions and evaluating how it responded. What lies behind that response is much bigger than it seems.

It was months of continuous learning. Every day, I spent hours reading Twitter posts (now X), the official documentation on the websites of the tools we were using—for example, figuring out the best way to upload content to a RAG⁸ and how the responses changed in different scenarios. Throughout that process, the DeepLearning.AI courses—particularly those by Andrew Ng, the site's founder and a world-renowned authority on artificial intelligence—were essential for me to better understand how it all worked.

Studying, experimenting, testing prompts. Going back and forth, failing, trying again.

Many times I felt like I was wasting my time. That it was too complex for someone who didn't come from a programming background. Many times I thought that maybe this was for others and not for me. But something pushed me to keep going—a mix of curiosity and fear of being left behind. And every so often, a small step forward: a prompt that worked, a concept that clicked, a piece that fit together with another.

It was a slow and sometimes frustrating process, but the ideas eventually began to take shape. There came a point where the pieces connected and everything started to make sense. Like when you're putting together a puzzle and suddenly you see the whole picture.

Two people

That was my entire AI world in 2023.

My first mentor—and the person I could always talk to—is my cousin Harpo, a university professor, expert researcher, and a leading figure in data science, machine learning, and distributed computing. He was born the same year I was, and back then we shared everything we could about computers. He went much further and studied computer science; I, on the other hand, chose a path in law. But we've both always enjoyed that side of things. Today, he's a leader, a pioneer, and continues to work on very important projects around the world. Harpo always truly understood this phenomenon. At first, he helped me with so many things. He was a great support—always.

By 2023, as AI was dominating the tech news, I began to wonder if it would be possible to replace the manual work done by Wispro's sales team—responding to prospective clients via WhatsApp. This idea seemed a bit far-fetched, especially since I didn't fully grasp the potential of AI at the time either.

Still, there was something else turning over in my head, and I kept telling myself, you need to be there, you need to do something that takes you into that new world, you need to build some project that at least keeps you in contact with the technology every day and helps you understand where it comes from and where it's going. I also remember a phone call I made to Harpo, telling him a bit about what I was thinking, and without hesitating he said, yes cousin, it's a good idea to start doing something related to AI today.

That's how Klari was born—as a software project integrated with ChatGPT's AI. With a small team of a few people working part-time, development began. Slowly, of course, but it got started. It was a new project, and there were many other things to do at the time, but we got started anyway. After a lot of work, the first versions of the screen where all incoming chats through WhatsApp, handled by AI, would be managed began to take shape—we called it a CRM.

From then on, the person I did talk to about AI—all day, every day, all the time—was my partner and co-founder of Klari. Although, as a programmer, he focused more on product design, he never stopped keeping up with new releases, news, and all the technical documentation he needed to study to implement solutions in Klari. After weeks of intensive learning, reading articles, researching, and hundreds of brainstorming sessions, we managed to bring the idea down to earth and verify that, in theory, it was something that could be done.

Our process was anything but fast. It took us about 6 to 7 months to have the first use case with my company, Wispro, using the initial version.

But its use didn't meet expectations. We weren't able to replace the sales team's work. What we did achieve was providing a new support channel for customers—a WhatsApp that operated 24 hours a day and responded based on the information we'd provided. We fed it Wispro's user guides and official documentation. With this, Klari's new bot could answer questions such as: "Where can I create a new invoice?"

But we weren't the only ones. A Canada-based company had already been doing the same thing for months and basically had a year-long head start on us. In less than three or four months after the launch of ChatGPT, they managed to bring Chatbase⁹ to market—a company I followed very closely. They offered the same thing we wanted to provide with Klari, only they grew very quickly. In the first version, they had a knowledge base—or RAG—connected to a chat embedded on your website to answer basic questions about products or services. This was exactly what Klari did by uploading all of Wispro's official documentation. It was the same concept.

The founder of Chatbase worked alone, from home, accomplishing in a short time what took us nearly a year.

One person, on their own, from a laptop was able to do it. I wish we had seen the opportunity in time, the way he did. Today Chatbase is a worldwide example of how a single founder grew an application so fast — it's a really compelling case study that I had the chance to research.

But seeing this was also a very positive sign that we were in a growing vertical and that we had a great chance of success. With Chatbase as an example, we could sense that it was possible to create new companies using artificial intelligence.

Here in Mendoza, AI was just starting to take off. Not everyone knew what a RAG was; few understood embeddings or vector databases, LangChain, or LangGraph. Of course, there were some specific cases where they were already working with models and conducting tests. The scientific community is always at the forefront in that regard.

But what I really want people to understand is that in 2023, in Mendoza, using concrete AI applications wasn't very common—at least in the circles I moved in.

1. Gillman, P. (Ed.). (1993). Everest: The Best Writing and Pictures from Seventy Years of Human Endeavour. Little Brown. ↩
2. Newey, A. (2017). How to Build a Car: The Autobiography of the World's Greatest Formula 1 Designer. HarperCollins. ↩
3. Atari 2600. Video game console launched in 1977 by Atari, Inc. It was the first successful console to use interchangeable cartridges. ↩
4. Talent DPC-200. Personal computer based on the MSX standard, manufactured in Argentina by Telemática S.A. in the mid-1980s. ↩
5. IBM PC AT (Advanced Technology). Personal computer launched in 1984 by IBM, with an Intel 80286 processor. ↩
6. n8n. Workflow automation platform founded in 2019 by Jan Oberhauser in Berlin. It allows connecting applications, services and AI models in a visual editor. n8n.io ↩
7. Lovable. AI-powered web application development platform, founded in 2023 by Anton Osika and Fabian Hedin. It enables creating fully functional applications using natural language. lovable.dev ↩
8. RAG (Retrieval-Augmented Generation). A technique that allows an AI model to query external documents before generating a response, improving its accuracy and reducing hallucinations. ↩
9. Chatbase. AI chatbot platform founded in 2023 by Yasser Elsaid in Toronto, Canada. It allows creating virtual assistants trained on custom documentation. Bootstrapped, with no external investors. chatbase.co ↩