AI Unmasked

Intelligence without the hype

I find it amusing when I describe things as “just math,” and some people get so caught up in the “just” part that they seem to completely forget the “math” one. Galileo Galilei believed the universe was expressed in mathematical terms, with math functioning as a language. Imagine this. So, for instance, if I say “this is just our universe,” getting stuck on the “just” part and forgetting the “universe” one is counter-productive to say the least.

I start this article with this thought. When I confront the idea of “magic” or “miracle”, I am doing so in favor of something much better, much more powerful, and much more interesting. Through systematic observation, the concept of “magic” appears to be related to a measure of distance from some particular aspect of objective reality. Not all reality, but some aspect of it.

The Rails Magic

In 2005, a keynote by David Heinemeier Hansson (aka DHH) took the world by storm! He presented a demo of his creation, Rails, a full-stack web development framework. This was a time when the tech stack that dominated web development (for most cases) was either LAMP (Linux, Apache, MySQL, PHP/Perl) or J2EE (Java 2 Platform, Enterprise Edition). Building software back then was an extremely bureaucratic activity. Lots of moving parts, dense documentation, lengthy configuration files, entity relationship diagrams, UML diagrams, and, depending on each case, it would take days, weeks, and sometimes months to see a not-so-complex feature come to life.

Senior software engineers of that time thought and acted a lot like civil engineers. Building software was similar, in spirit, to building an apartment complex.

But DHH was never too conventional. When it was common for software engineers to be antisocial and avoid any contact with users, DHH was an entrepreneur, agile practitioner, book author (New York Times Bestseller), race car driver, and keynote speaker. He was and continues to be good with marketing. So he came up with this presentation title: “How to build a blog engine in 15 minutes with Ruby on Rails”.

That… just… worked! What would easily take days to be done, he did, live, in just 15 minutes. And it was about a pretty UI. I am talking about database modeling, backend development, frontend integration, input validation, unit tests, etc.

There was no generative AI back then. This was 100% deterministic. Developers could not believe this new crazy thing. Rails grew rapidly in popularity, and it reached its peak between 2006 and 2010. Startup founders used it to build their solutions, as it happened with Twitter, GitHub, Shopify, Airbnb, Kickstarter, Zendesk, Twitch, SoundCloud, Hulu, and many others.

You can watch the full video below:

With Rails’ success, new terms emerged among its advocates, such as “The Rails Way” and “Rails Magic”. Of course, this phenomenon caused great debates. Some people thought the magic was much needed. Building software was way too complicated. It was time for a change. Other people thought that the “magic” was an illusion, and that not really knowing what they were doing would lead developers to irrecoverable problems down the line. Both sides had their merits. History shows that while some Rails apps were incredibly successful, to this day, the market was flooded with Rails apps built by people who only knew the magic and nothing else, and many of them did not survive scaling needs, which led to the “Rails doesn't scale” mockery of that time.

It Was Never Magic

A dear friend of mine, Paulo Fagiani (in memoriam), who would become one of the most prominent advocates of Rails and agile software development in Brazil during the 2000s and 2010s, showed me DHH's video shortly after it was published. I watched the entire video without saying a word. I was absolutely fascinated by this magical thing called Rails. I have been building software since 1998. I could not believe that some of the things that would take me so much time to do were now so quickly and easily achieved.

After watching the video, I was in awe. I couldn't stop thinking about the many projects I would build using Rails. I remember telling Fagiani, “This changes everything!”

Fagiani then walked me through Rails's source code. He explained that Rails aggressively combined elements like convention over configuration, which would replace extensive configuration files with best practices. In other words, if you follow the convention, a series of resources will be automatically available to you. ActiveRecord, Rails's Object-Relational Mapping (ORM), was heavily powered by Ruby's metaprogramming capabilities (code that writes code). Although the model, view, controller pattern (MVC) was around since the 1970s, Rails was the first to make it incredibly and beautifully obvious, built into the Rails framework. Combine all of that with generators and scaffolding engines (with which files and hundreds of lines of code will be generated in just a few seconds), and the perception of “magic” was granted.

Fagiani's goal was clear to me. He wanted me to go beyond the surface and discover that it was never magic. It was something much bigger, much more powerful, and much more exciting: creative, innovative, entrepreneurial, audacious, visionary, and inspiring software engineering.

End users and power users of development tools share a common characteristic: they are similarly distant from the essence of the technology they admire.

This is something some people seem not to have yet understood: when I say “it is not magic,” I mean “Don't undersell it! It is something so much better than that!”

When you refuse to see a technology for what it is, you are surrendering power to those who built it. But when you understand what the technology is, you truly manipulate that power in your favor.

The Problem with the Magic Idea is the Wrong Mental Models

Every technology, especially those related to computers, is made of layers. At the core, we have the physics (transistors, silicon, electrons), then we have the implementation layer (code, algorithms, data structures), then the abstraction layer (compilers, frameworks, APIs), then the user interface (design, interaction, visual language), then the interpretability layer, which is where the user-constructed mental models are made.

You will never hear a chip designer or a compiler engineer talking about “magic.” They will talk with precision about concrete, core technological elements that make impressive, powerful, and profitable results possible.

Rails was supposed to be at the interface of implementation and abstraction, but at least for a new generation of web developers that got stuck in the magic, it was positioned at the interpretability layer, sparking philosophical debates that had little to nothing to do with its undeniable technical merits.

Looking back to 2005 and closely inspecting what is happening now in 2026 with the AI hype, I come to the conclusion that the more distant one is from the implementation layer, the more “magic” the technology looks like.

Although not exclusively, this epistemic distance from the implementation layer is mostly what leads people today to talk about AI (in general) and AI tools as something “alive”, a “mystery”, a self-aware entity with its own desires and goals. I say not exclusively because there is also the narrative to justify impossible budgets and investments. And yet, in general, it holds that the more distant one is from the building blocks of a technology, the more “transcendental” it will look.

There are times when researchers and engineers work on physical analogies and other illustrations to help communicate complex subjects, which is useful and, many times, even required for a better absorption of dense technical matters. But at the interpretability layer, mental models are often wrong. And wrong mental models bring more “wrongs” with them as they help create wrong expectations, direct resources in the wrong direction, push to production the wrong applications for specific problems, and in the long term, lead the market to invalidate and reject the technology when the technology was never the problem.

AI Unmasked

I am well aware that “one swallow does not make a summer,” but perhaps, one day, there will be enough swallows to change the season. I can't help but have a positive, highly enthusiastic view of the future of all things AI, despite the deservice caused by the AI hype.

So, since the AI hype is highly organized and orchestrated, there must be an antidote. In the signal vs noise dispute, I choose signal every day.

I also find that engaging in endless arguments (arguing for the sake of arguing) is a tremendous waste of time. It consumes precious resources like time and energy while producing nothing useful. How can I meaningfully contribute to producing more signals about AI?

That's why I decided to build AI Unmasked, a series of courses, tips, insights, hands-on experiments, labs, and other resources to help more people to deepen their knowledge on all things AI. My journey in the AI field is an ongoing venture, as I am learning something new every day. All of this and more will be incrementally available on AI Unmasked.

Courses and premium materials are already available to paid subscribers from Substack. However, I will also keep free resources available for anyone. So these two types of resources will be published regularly.

What the Materials Will Cover

We start with an overview to get started. Think of it as a “know before you deep dive.” We will first examine highlights of everything we will later crack open. We will look at the theory, run experiments, build AI tools from scratch, code our solutions, and more.

I tried to keep the prerequisites as minimal as possible. If you are comfortable using a computer, you should be good to go.

Give it a Try!

The frisson around large language models (LLMs) is real. And that is due to a killer consumer application: a chat interface. It makes it possible for anyone to type anything and obtain an answer. This makes it a wonderful magic box. But what if there is no magic at all involved?

We can't get to the core LLMs until we start with the “M” part: what is a model?

The first course released on AI Unmasked is named “The Foundation.” If you want to enrich on knownledge on all things AI, I would say this: give it a try! You might end up liking it:

I would love to hear your thoughts about how I structured each module. There is enough there now to get everyone entrained and much more coming soon.