More important than AI skills: defining the problem

Learning to code once took years of study. It has moved from no-code tools to 'vibe coding,' where you build programs by describing the features you want to AI in plain language.

Vibe coding builds programs by having AI implement features described in words, and as the technology advances, agent automation too needs less and less human involvement.

AI slop is what happens when AI, filling in the blanks based on patterns in its training data, makes average choices.

It refers to output that looks plausible but shows little of a person's unique concern or distinctiveness. Because AI takes over the intermediate judgment that humans once made, the result is generic work.

It is hard to specify in a prompt a problem the user doesn't yet recognize. You can write down problems you already know about, but you cannot instruct AI about what you don't know yet.

Unexpected problems that show up while building a real product can only be discovered by doing the work. If AI handles most of the intermediate judgment, a fundamental question remains about how the role of the person who supplies the idea differs from that of the person who uses the finished product.

Writing a paper or building a program used to be proof of technical effort and time invested, and because of that scarcity, the mere existence of a finished work signaled some ability.

But as the cost approaches zero, the value of simply having produced something is falling. Simple repetitive computer tasks, and work once abandoned because it was too expensive, can now be done in volume with AI.

Much of what happens at work does not demand world-class creativity. It calls for reports or design programs that serve an ordinary purpose. Reliably producing a decent level of output was once considered sufficient expertise.

Now that reliably producing that kind of output has become something AI does best. Because AI has learned from an enormous amount of human work, it already knows the appropriate choice for many situations.

If you have several ideas, you can simply build them and check the results.

As the cost of failure drops and most failures become acceptable, the value of professions that only implement things is expected to decline.

The division of roles between planners and makers is breaking down. The AI skills you are struggling to learn today may be done by AI itself next year.

If you make mastery of each new tool your competitive edge, then as the tool gets easier, the scarcity of the skill you painstakingly acquired drops too. In other words, training in how to use AI tools has a short shelf life.

Knowledge is not becoming unnecessary. Rather, the value of knowledge for operating tools is falling, while the value of knowledge for judging the results is relatively rising.

Even when AI shows an impressive result, you need the ability to recognize that 'it doesn't work like that in reality,' and that ability requires domain expertise to judge AI output. Experienced users create strong synergy when they use AI, which can offer simulated experience that shortens what would have taken ten years.

The way we use computers is changing. We are in a transition from keyboard-centered work to a system of verbal instruction. Lately, when an idea comes to mind, instead of jotting it down or typing it into a phone as before, people keep talking their ideas through with ChatGPT.

This means ideas are passed to AI as they are, not as polished sentences, and work is done that way.

In the AI era, the productivity gap comes from defining the problem, not from tool proficiency. Simply asking AI to do what everyone else is already doing does not create value.

Conversely, if you find work nobody has done yet, or work that is needed but was never attempted because of cost, and you have AI do it, you can create a completely different level of productivity with the same AI. Ultimately, what matters is not whether you know how to give orders, but whether you can tell which work is worth doing.