The Virtuous Cycle: On Learning and Building with AI

I have spent most of my adult life trying to read faster than the world can write. AI did not solve that problem. What it did was something I did not expect: it turned the gap between what I understand in the abstract and what I can build in practice from a wall into a bridge.

My bottleneck has always been reading and writing. Not the thinking — the throughput. As a teenager in Prague, I hauled books in and out of the city library in a backpack that left marks on my shoulders. When the first iPad came out, I was genuinely excited: all my books in one device, no more hauling. The constraint was always the same — more to read than hours to read it.

Over the years I found workarounds. Audiobooks while driving or hiking. Google’s NotebookLM to turn long documents and books into podcast-style summaries. Services like Listening.com that convert scientific papers into audio. Scite.ai for tracing citation networks back to the original source. I have always believed that access to first-rate, straight-from-the-source information has been democratized far beyond the original circle of the well-connected and well-funded. These tools confirmed it.

But consuming information faster is only half the problem. The other half is doing something with it.

I started using ChatGPT in 2023. The first uses were mundane and immediately useful: drafting assignments for my junior colleagues at the University of Washington, creating documentation, summarizing long reports. Tasks where I knew exactly what I wanted but the writing itself was the bottleneck. The machine was fast at the part I was slow at.

Since then I have tried most of what is on offer. Gemini, ChatGPT, Meta AI, Claude. Various versions of Sonnet, Opus, and the newer Fable. I have opinions. Opus 4.6 remains my favorite — when it first came out, I was genuinely impressed by its capabilities. The later versions seemed to consume my usage quota faster without a proportional increase in what I could do with them. But the specific model matters less than the pattern of use, and that pattern took me a while to recognize.

Here is the pattern. When you approach a new domain, you start by letting the AI do the heavy lifting. You do not yet know enough to evaluate the output critically, but you know enough — from adjacent experience, from general principles, from having read widely — to point it in roughly the right direction. You are, in the honest sense of the phrase, faking it until you make it.

But then you read. You pick up the manual, the textbook, the reference guide. And as you learn, something shifts: you start noticing where the AI was imprecise, where it took a shortcut, where it gave you a correct answer to a question you should not have been asking. Your prompts get sharper. Your feedback gets more specific. The output improves. And that better output teaches you faster, because now you are reading code and architecture that is closer to correct. The cycle accelerates.

This is not a metaphor. In my current role as an FP&A data engineer, I started using DAX and the Power Query M language — two languages I had never touched — in my first week. Within days I was building production artifacts: data models, transformation pipelines, reporting templates. Not because the AI wrote perfect code, but because I understood dimensional modeling, I understood data pipelines, I understood what the output should look like even when I did not yet know the syntax to produce it. The AI filled in the syntax. I supplied the architecture. And every day, the balance shifted a little: more of the syntax lived in my head, and the questions I asked the machine got more interesting.

I had tried something similar before, in a different context. I wanted to build a flashcard app that used an LLM as a learning tutor. I started with a Hugging Face model running locally, but quickly realized I was spending all my time wrestling with the model infrastructure and none designing the learning experience. I moved to the ChatGPT API. The pattern was the same — use the tool to handle what you cannot yet do yourself, and invest your attention in the part you understand best.

I think part of why this works for me is temperament. My training is in economics, which is in some sense a discipline of strategic thinking — it deals with aggregate phenomena, systems, incentives, second-order effects. It is a very big-picture subject. I have always been more comfortable thinking about what to build and why than about the specific implementation details of how.

I find that many people are the opposite. They are excellent tacticians — they know how to implement this, how to configure that, how to write production-grade code in a specific language. What they are working on, though, is not always the most productive use of their skills. The strategic layer — choosing the right problem, designing the right architecture, knowing which trade-offs matter — is a different skill, and AI does not replace it. If anything, AI makes it more valuable, because the cost of implementation has dropped. The bottleneck has moved upstream, to the person who can see the whole board.

This is where I have found AI most useful: not as a replacement for thinking, but as a bridge between strategic understanding and tactical execution. I understand data, communications, technology, and computer science at a high level. With AI, I can get the rest of the way quite effectively.

If I had to name the one skill that suddenly matters more than any specific programming language or software package, it would be critical thinking. Not critical thinking in the vague, resume-padding sense, but the real thing: how to source information, how to weigh evidence, whether a claim can be rejected, whether there is a ground truth or just a spectrum of opinions. Epistemology — the study of what counts as knowledge and why.

Karl Popper spent his career on these questions. Falsifiability, the demarcation problem, the logic of scientific discovery. For most of the twentieth century, this was academic conversation — the kind of thing that interested gray-haired men in philosophy departments. It is now, I would argue, the most practical skill set a working professional can have. When an AI gives you an answer, can you evaluate it? When it cites a source, can you trace the claim? When it produces code that runs, can you tell whether it is correct? These are not technical questions. They are epistemological ones.

My classical Czech education was built on a different model entirely: absorb information, reproduce it on the exam. That model was already being challenged by Google — why memorize what you can search? The arrival of AI has shattered any remaining illusion of its usefulness. If a machine can produce fluent text on any topic in seconds, the person who merely absorbed and reproduced has nothing left to offer. The person who can evaluate, question, and judge has everything.

The education system, conservative institution that it is, has responded the way conservative institutions do: denial, defensive posturing, suspicion, hostility. Banning AI in classrooms. Treating its use as academic dishonesty. Designing assignments around what machines cannot do rather than teaching students to work with what machines can do. It is a siege mentality, and it will not hold, because the walls it is defending were already breached by the search engine a generation ago.

Before landing my current role, I spent months interviewing across the Czech job market. What I found surprised me. Of the companies I spoke with, the vast majority either did not mention AI or actively prohibited its use.

At one consulting firm serving the energy sector, I shared my enthusiasm about a deep learning model I had read about that analyzed satellite imagery to infer electricity usage patterns. The interviewer spent the conversation steering me toward agreeing that AI was a hype. At another company, getting a Copilot license — twenty-five dollars a month — required a formal business case demonstrating at least five thousand dollars in expected returns. At a third, the IT business partner spent more time describing the dangers of AI agents accidentally deleting a database than discussing any potential productivity gains.

I do not think these people were wrong to be cautious. Security matters. Governance matters. But the pattern was striking: the conversation was almost entirely about risk, almost never about opportunity. In a technology that I believe fundamentally changes the ratio of what one person can accomplish, the default institutional response was to treat it as a threat to be contained rather than a capability to be developed.

There is a reaction I keep encountering — in interviews, in professional conversations, online — that using AI is somehow cheating. That the output does not count if a machine helped produce it. That real competence means doing everything yourself, from scratch, by hand.

I find this difficult to take seriously. Nobody argues that using a calculator is cheating at arithmetic, or that using a search engine is cheating at research. These tools changed what it meant to be productive, and the people who adopted them early were not cheaters — they were realists. The question was never whether to use the tool. The question was how effectively you could use it.

The same question applies now, and the answer depends on what you bring to the table. AI is a force multiplier, and a multiplier needs something to multiply. A decade of thinking about data architecture, causal inference, and systems design turns out to be exactly the kind of foundation that AI multiplies well. Not because the machine thinks for you, but because you have spent years learning what questions to ask — and asking good questions turns out to be the skill that matters most when the machine can answer them instantly.

I am still reading. I still listen to Popper on audiobooks and pick up reference guides on topics I know nothing about. The difference is that now, when I finish a chapter, I can sit down and build something with what I just learned — that same afternoon. The cycle between learning and doing, which used to take months, now takes hours. That is not cheating. That is what the technology is for.