The AI Iceberg

We see a lot of news about AI everywhere, and much of it can bring anxiety and uncertainty about the future. While nobody can predict exactly how far AI will go, think back to when the internet was first created—who could have imagined it would one day replace our home phones and completely change how we watch TV?

But what is just hype, and what should we actually take seriously? That is exactly what I will explore in this article.

First, let's look at the "AI iceberg" represented in the picture below, where I will try to break down each layer in a friendly, easy-to-understand way.

On the surface, it represents all the most well-known, user-friendly tools available on the market to use the power of LLMs ( Large Language Models ), Agents, and image generation through prompts— but wait: What are prompts?

Prompts are instructions that you can give to the AI to do a task, create an Agent (I will explain later what it is), create a skill, generate an image, summarize a text, or just correct spelling and grammar errors.

For example, this is the prompt that I often use when I am writing emails or other texts, like this article:

From now on, for any text I paste here in English, French, Dutch, or Brazilian Portuguese (just for this session), please fix any grammar and spelling mistakes. Keep in mind that my native language is Brazilian Portuguese. With as few changes as possible, adapt the text so it sounds natural in those specific cultures: US/Canadian for English, Metropolitan French for French, and Netherlandic Dutch for Dutch. Keep the original meaning and wording as close to mine as possible.

To keep this text from getting too long, I'll stop diving into this ocean of an iceberg at the Agent level.

Let's now dive a little bit below the surface, where all those popular Generative AI tools use agents (or: are actually agents). With the following image, I will explain in a nutshell what is what, without too much technical jargon or detail.

Why are they called Large Language Models? The word "large" refers specifically to the sheer number of parameters inside the AI's network—typically ranging from billions to trillions—which makes them massive compared to older language models.

When we talk about "parameters," think of them like the billions of connections in a human brain. They allow the AI to learn how language works, using statistics to predict the next most likely word.

Model Context Protocol (MCP) is a little technical, but to keep things simple and understandable to everyone, it provides a standardized way for AI applications to discover and interact with predefined functions (from services on the internet or locally), external tools, and data sources, making them accessible to LLMs.

Memory can be classified as short-term or long-term and can use databases, files, or other storage mechanisms. Memory helps retain information that can be retrieved later by the runtime orchestration system.

Software that manages the execution loop, state management, tool execution, routing and handoffs, error handling, and safety limits. It could execute the following process, for example, as shown in the image below:

This is a very good question. We see a lot of hype and news aimed at driving engagement, which can cause anxiety for many of us, especially if we don't understand what AI is and what it's capable of.

Think of AI as an iceberg. Deep at the bottom, its foundation is mathematics and statistics. Then, as we move up a level, we have machine learning, followed by deep learning, until we reach the surface with generative AI (ChatGPT, Copilot, Claude, Gemini).

AI can perform tasks based on what it has learned during training. Depending on the AI system, it can do things like generate text and images, drive a car, pilot a drone, or describe images and videos.

A good metaphor I heard from a colleague is: "Using AI is like giving your brain a car."

However, there's an interesting nuance. A car doesn't automatically make someone a better driver, and AI doesn't automatically make someone a better thinker.

AI can improve your productivity if you use it as a tool for many tasks. For example, as a software developer, today I can create new features for an application in minutes or hours, whereas before, it could take days, weeks, or even a month. But AI doesn't replace me. Quite the opposite: it takes my knowledge and expertise to determine whether the generated results are safe, maintainable, and correct before they can be used.

But what could happen as AI evolves? Job transformation may be more likely than complete replacement. A team could become more productive with the same number of people.

Some fields are already feeling the effects of AI, such as translators, voice actors, and visual artists, including illustrators. Other fields are also being affected by the use of AI-powered robots in automation processes.

I cannot say what the future impact of AI will be as it continues to develop, but we can already see changes in our work environments. For some professions, the way we work today is not the same as it was three years ago.

Some companies are also reducing the number of junior professionals they hire in AI-exposed occupations as AI advances.

The only recommendation I can give you is to learn how to use and get the most out of AI tools like ChatGPT, Claude, Copilot, Gemini... you name it. These tools could become as essential to our work as Excel, Word, and PowerPoint became 10 or 20 years ago.

Another thing to consider is how we should use AI without harming our careers and reputations. For example, some people use AI to read documents, generate summaries, and respond to colleagues. I don't think this is always a good idea. If your colleague is asking for your opinion, are you using Copilot to help you understand the document, or are you letting Copilot form your opinion for you?

I'll leave this question open for you to think about: How are you using AI today?