I talk about artificial intelligence often.
I talk about agents, software, programming, models, and how these tools are changing work. Yet I rarely publish the typical tutorial about “how to use this new feature,” “the best prompts,” or “how to configure a tool step by step.”
I do not make a video every time a company announces a model either.
This is not accidental. It is a conscious decision about the content I want to create and, increasingly, about how I think we should learn to relate to this technology.
Tutorials are useful, but they age very quickly
The simplest reason is that thousands of excellent tutorials already exist. Entire channels are dedicated to testing every new feature, model, configuration, and interface.
I use them myself. When I need to configure something specific, I will probably look for a tutorial.
But producing one means researching the tool, testing it, recording the screen, editing, and explaining every step. Two weeks later, the interface may have changed, the feature may have disappeared, or a task that once needed seven steps may now need one.
Every update requires another video. Every new tool replaces part of the previous one. As a creator, you end up chasing a wheel that never stops.
Another question interests me more.
Not so much “where do I click today?”, but “what is actually changing?”
Tools change quickly. The ideas underneath them change much more slowly.
We are approaching probabilistic technology with a deterministic mindset
For decades, we learned to use computers in a particular way.
We give them an instruction and expect execution. If we repeat the exact same operation, we expect the exact same result. A function receives certain parameters and returns a predictable output.
Then a different technology appears.
We can send the same prompt twice and receive different answers. A model solves one problem brilliantly and fails on an almost identical one five minutes later. We can copy a workflow shown on YouTube and get another result.
Still, we search for the perfect prompt, perfect agent, and configuration that will always do exactly what we want.
We are trying to domesticate a probabilistic machine as though it were deterministic software.
A tutorial can teach a procedure and offer good practices. It cannot guarantee that seven steps will always produce the same output, because these systems do not work that way.
The skill we need, therefore, is not merely following instructions.
Following a recipe is not the same as learning to cook
A recipe can specify exact quantities, times, and temperatures. If we follow it, we will probably get the expected dish.
But knowledge appears when something changes: the dough is too wet, the oven runs differently, an ingredient is missing, or the result looks nothing like the picture.
That moment reveals whether we memorized a sequence or understand what we are doing.
AI works similarly. A tutorial demonstrates the happy path. Real work with these tools often means knowing what to do when that path does not appear:
- The model misunderstands the task.
- The answer looks correct but is shallow.
- The agent takes an unexpected direction.
- The solution works technically but starts from the wrong idea.
- More context, a reframed problem, or a rollback is needed.
The magic prompt is not very useful there. What we need is judgment.
Asking well is only half the skill
Since ChatGPT appeared, people have repeatedly said that the future belongs to those who know how to ask good questions.
That is true but incomplete.
We also need to evaluate the answers. Doing that requires at least some understanding of the domain in which we work.
I can ask an AI to design a software architecture. It may return microservices, queues, databases, caches, and a diagram full of convincing terms.
Is it a good architecture, or does it merely look like one?
Without foundational knowledge, I may not be able to tell the difference.
This reveals an important paradox: AI can dramatically reduce the knowledge required to produce something, but it does not necessarily reduce the knowledge required to judge whether it is good.
The more convincing outputs become, the more important our ability to evaluate them becomes.
Fundamentals do not lose their value
This is obvious in programming.
An agent can create components, configure frameworks, write tests, fix errors, and deploy an application. That is an extraordinary expansion of execution capacity.
But when the architecture becomes chaotic, the data model is poorly designed, a vulnerability appears, or the system grows unsustainably, someone must recognize the problem.
How do I know what to correct if I do not understand what is failing? How do I know what to ask the tool if I cannot evaluate what it did?
The same applies to writing, research, design, studying, and building a business.
AI increases our capacity to produce. Fundamentals allow us to direct that capacity.
That is why I do not believe they have stopped mattering. They probably matter more.
A capability prosthesis
I find it more useful to think of AI as a capability prosthesis than as something that does the work for us.
A telescope does not choose what an astronomer should investigate; it expands what they can observe. An excavator does not decide where to build; it expands the operator's strength. A calculator does not define the problem; it expands our ability to calculate.
AI expands many capabilities at once:
- Writing and programming.
- Researching and analyzing.
- Organizing information.
- Exploring ideas.
- Proposing and comparing solutions.
But direction still has to come from us: what we want to achieve, what we consider good, which constraints exist, what must be verified, and when we should stop.
AI does not eliminate our judgment. It amplifies it.
If that judgment is good, it can multiply it. If it is poor, it can multiply that too.
We are making new technology imitate old work
When a new technology appears, we initially use it as a version of the previous one.
Early automobiles looked like horseless carriages. Early websites resembled printed brochures. Early printed books imitated manuscripts.
We are doing something similar with AI.
We take a human process made of steps A, B, C, D, and E. Then we insert AI into step C or automate B and D.
That is a reasonable beginning, but perhaps the most important question is not “how do I do my current work with AI?”
It might be:
“If we designed this work from scratch knowing AI exists, would we do it the same way?”
The answer will often be no.
The system around the model matters
When building agents, we usually surround the model with tools, context, memory, permissions, validation, retries, output formats, budgets, and limits.
This environment is often called a harness: the set of controls that makes a model more useful and predictable.
Those controls are necessary. Much of the progress in agents comes not only from better models but from learning to build better systems around them.
The mistake would be assuming that the final objective is to turn AI into traditional software.
If we specify every possible move, we reduce its ability to explore, work with uncertainty, and find paths we did not anticipate. We need reliability, but we should also preserve what makes the tool interesting.
Today's cost shapes the workflows we design
Imagine for a moment that using AI were nearly free and instantaneous.
Instead of asking for one answer, we could request twenty solutions, ask several agents to criticize them, discard the weakest, and retry until clear criteria were satisfied.
We do not work like that all the time today because inference costs money, energy, compute, latency, and human supervision.
That is why we build limits and budgets. We try to obtain the best result with the least reasonable amount of inference.
This makes sense under present constraints. But those constraints may change.
In a few years, we may look at some current workflows as we now look at old software obsessed with saving a few bytes. It was not poorly designed; it responded to the price of memory at the time.
Many methods presented today as permanent rules of AI may only be temporary adaptations to the current cost of using it.
Automation and intelligence are not the same
If I can specify every step of a task exactly, I can probably solve it with traditional software:
first A → then B → if C happens, execute D → validate E
That is automation, and it remains extremely valuable.
AI becomes especially interesting when we do not know the complete path. We can describe the current state, the objective, the constraints, and how to recognize a good result; then we allow the system to find a path.
For decades, programming largely meant telling a machine exactly how to do something.
Increasingly, our work may consist of telling it what we want, where the boundaries are, and how we will know it finished well.
That is a profound change in our relationship with computers.
Less control of the path, more control of the boundaries
Directing intelligence does not necessarily mean specifying every move.
It may mean designing a safe space within which it can act:
- This is the objective.
- These are the constraints.
- These actions are forbidden.
- This is the budget.
- These resources are available.
- These criteria determine whether the work is complete.
The system then retains some freedom to find the route.
Less control of the path does not mean no control. It means concentrating control on boundaries, permissions, validation, and outcomes.
This approach requires much more than a good prompt. It requires designing new systems and processes.
From doing the work to directing systems that work
Much of our professional value has historically been in execution: writing, programming, researching, analyzing, and producing.
As artificial systems perform more of that execution, our role may shift toward another kind of work:
- Defining the problem correctly.
- Providing context.
- Designing constraints and permissions.
- Choosing tools.
- Evaluating results.
- Giving feedback and correcting direction.
- Taking responsibility for the final decision.
Memorizing the interface of one particular tool is not enough to do this well.
We need fundamentals, critical thinking, domain understanding, abstraction, and continuous learning.
Content that should outlive the tool
That is why this space will probably continue to discuss these ideas more than buttons.
I will keep testing tools and discussing news when it reveals something important. But the news will often be an entry point for understanding the direction of change.
If ChatGPT, Claude, Codex, or any current product disappeared tomorrow, I would like most of what we discuss to remain useful.
I do not want to teach only how to use this week's tool.
I want to understand how we work when tools stop behaving like the computers we used to know.
Perhaps the great change is not learning how to use artificial intelligence.
It is inventing a new way to work with it.
This article develops the editorial decision behind the channel and the kind of learning about AI I want to explore. You can watch the original video on YouTube.