For some time now, OpenAI models have appeared, in one way or another, in almost everything I set out to do.
I have used them to research subjects that initially felt far beyond me. To answer everyday questions I might otherwise have postponed. To prepare scripts, ideas, and assets for my YouTube channel. To work through programming exercises, understand unfamiliar code, and turn ideas that were still messy into prototypes that actually worked.
If I look only at the list of results, it already feels like a lot.
But that would be a poor way to tell the story. The best thing they gave me was not a collection of answers, texts, or lines of code. It was everything I learned while trying to build with them.
From curiosity to something concrete
I have always been curious. The problem is that curiosity alone is not enough. Between a question and a useful answer, there are often books, papers, prerequisite concepts, terms you do not yet know, and a considerable amount of time needed to organize it all.
AI shortened that distance enormously.
When I begin researching something, I can ask for an initial map of the subject, identify the concepts I am missing, compare explanations, and discover better questions than the ones I started with. It does not replace primary sources or decide for me what is true. In fact, I learned quite quickly that a confident answer can be wrong and that anything important still needs to be checked.
But it lets me enter the subject.
And entering changes everything. A subject that once looked like a wall begins to look like a sequence of steps. Sometimes I end up reading material I would never have found on my own. Other times, I discover that my original question was poorly framed. Even when the first answer is not useful, the conversation helps me understand why.
Something similar happened with everyday questions. I no longer have to choose between ignoring a small doubt and opening twenty tabs to resolve it. I can explore it in the moment, ask for a different explanation, and keep going until I genuinely understand.
I do not feel this has reduced my curiosity. It has removed friction from it.
Creating for YouTube without waiting to know everything
The channel has been one of the places where I have experimented the most.
I tried scripts, structures, titles, visual assets, and formats ranging from avatars to audio-only videos that felt almost like radio. Many of those attempts were imperfect. I would do some of them differently today. But I was able to make them.
That matters to me because an idea used to sit in my head for weeks, waiting until I knew how to write, produce, and present it perfectly. Working with AI brings the idea to a first version much sooner—a version I can see, criticize, and improve.
The model does not know what I mean until I explain it. It does not have my experience or my judgment either. It can write something technically correct that does not sound like me, exaggerate a conclusion, or turn a simple reflection into prose that is overly polished and lifeless.
Learning to notice that was part of the process too.
Over time, I began giving better instructions, but more importantly, I learned to recognize my own voice more clearly. Which words I would never use. Which claim needs more care. Where I want to leave a question open instead of closing the subject with a perfect sentence.
In a strange way, working with a machine that can produce so much text forced me to decide more clearly which part of the text was truly mine.
Programming stopped beginning with a blank screen
The change was even more visible in programming.
I was able to complete exercises, explore technologies I did not know well, diagnose errors, and build prototypes without having every technical detail resolved beforehand. This website, the blog, and several ideas I have tested grew out of that kind of collaboration.
It was not a matter of pressing a button and receiving finished software.
I had to explain what I wanted, make decisions, test, find what did not work, backtrack, review code, and learn enough to distinguish a reasonable solution from one that merely sounded convincing. Sometimes AI saved hours. At other times, it sent me in the wrong direction and I had to rebuild the context.
The difference is that I was no longer alone at every roadblock.
I could ask what was happening, request alternatives, compare the costs of a decision, or have it explain a part of the system until I understood it. That did not remove the need to learn. It made learning part of the work itself.
And the more I built, the better my questions became.
The answers were not the most valuable part
There is a seductive idea around AI: that its value lies in doing things for us.
My experience has been different. The greatest value appeared when I used it to help me do things that once took too long, intimidated me, or sat beyond my immediate reach.
I learned to break down large problems. To separate a need from the first solution that comes to mind. To ask for evidence. To make constraints explicit. To recognize uncertainty. To give concrete feedback instead of simply saying, “I don't like this.” To review an answer without assuming that good writing means good thinking.
I also learned that the model improves considerably when I do my part better.
If I give it a vague idea, it may return an elegant version of the same vagueness. If I provide context, criteria, and clear intent, the conversation changes. Not because there is a magical prompt, but because building something with another intelligence—human or artificial—requires learning to communicate what is still only half-formed in our own mind.
That is why I do not see these months as a story of delegation. I see them as a process of expansion. I was able to research more, test more, and finish more things. Along the way, I developed capabilities I did not have when I began.
How I choose between Sol, Terra, and Luna
For a while, I searched for the best model, as if one model had to be right for everything. That question now seems less useful.
I do not always need the most capable model. I need the model's capability to match the difficulty, ambiguity, and cost of being wrong in the task in front of me.
My practical rule today is simple:
- If a task is important, difficult, or still poorly defined, I choose Sol. That is where I would place demanding research, complex programming, a decision with several implications, or work that needs extensive polish.
- For normal day-to-day work, I start with Terra. It works well for writing, analysis, organizing ideas, and general programming, with a strong balance between depth and speed.
- If the task is perfectly defined and speed is the priority, I use Luna. Summarizing, extracting, classifying, or repeating a clear operation does not always require the largest tool.
Reasoning effort is a separate choice. For most work, Terra with medium reasoning feels like a strong starting point. If the problem needs greater depth, I increase the effort or move to Sol. There is little value in spending more time and compute when the task is already solved well; there is equally little value in saving a few minutes if an error will force me to redo everything.
This is not a law or a benchmark. It is a personal heuristic. The models will change, and this choice will probably change with them. But it helps me stop choosing by impulse and start choosing according to the actual shape of the problem.
OpenAI's official documentation suggests a similar distinction: Sol as the highest-capability option, Terra as the balance between intelligence and cost, and Luna for efficient or high-volume work. I find that a useful foundation, as long as each person tests it against their own work.
Keep learning without surrendering judgment
I do not want to turn this into a naive celebration. Models make mistakes. They invent things. They misunderstand. Sometimes they produce a lot when what you needed was an honest question. The more important the result, the greater our responsibility to review it.
Even with those limits, the results they have given me so far have been extraordinarily good.
Not only because of how much I was able to produce, but because they expanded what I was willing to attempt. A piece of research could become a script. The script could become a video. A technical question could become an exercise. An idea that felt too large could become a prototype. Each result left knowledge behind for the next one.
The models did not make the journey for me. They allowed me to cover much more ground.
Perhaps that is my most sincere recommendation: do not use them only to finish sooner. Use them to approach questions you still do not know how to answer, build something you do not yet know how to build, and pay attention to everything you learn while trying.
Because, at least in my experience, the best result was not an asset, a text, or an application.
It was becoming someone capable of asking better questions and carrying more ideas into the real world.
References
- OpenAI, GPT-5.6 model guide.
- OpenAI, Text models.