It Wasn't the AI. You Were Trained to Be Replaceable

Every week there is a new culprit for programmers losing value, and it is almost always AI. But AI only does very well what we were trained to do for years. The real value is in what nobody taught us to cultivate.

English translation of an article originally written in Spanish. Leer el original en español.

Illustration of a column of checked, verified tasks on the left and, on the right, an open circle with a question mark and an arrow leaving it
In this article
  1. Software no longer has a final version
  2. Where there is a judge, AI already innovates
  3. Where there is no judge, the human begins
  4. The only way to reinvent the wheel is not to use it
  5. Fundamentals for thinking, not for producing
  6. Stop being a doer

Every week I read a new headline about AI coming to replace programmers. And the culprit is always the same: the machine. I think we are looking in the wrong direction.

For years we were trained to be doers. University, bootcamps, job ads and technical interviews always measured us by the same thing: which language you master, which framework you know, how many years of experience you have in some tool. The ideal programmer was the one who executed the recipe fast and without errors.

Now think about what an AI does well: executing recipes fast and without errors. Writing the endpoint, applying the pattern, passing the tests. Exactly what we were trained to do.

AI did not come to replace developers. It came to replace the profile the industry manufactured for years. The difference matters, because what it cannot replace is precisely what nobody taught us to cultivate. And to understand what that is, we first have to accept how far AI goes, which today is much further than most people think.

Software no longer has a final version

Let's start with something that is already happening. In April 2026, Anthropic launched Project Glasswing, an initiative in which companies that maintain critical software use Claude Mythos Preview to look for vulnerabilities. In its first report, published in May, Anthropic said that in just one month it and its roughly 50 partners had found more than ten thousand high- or critical-severity vulnerabilities in some of the world's most important software. One example: Mozilla fixed 271 vulnerabilities in Firefox 150, more than ten times the number it had found in version 148 with the previous model.

But what interests me most is not the number but the report's own conclusion: progress in security is no longer limited by how fast vulnerabilities are found, but by how fast the ones AI finds can be verified, communicated and patched. The bottleneck stopped being the machine and became the human. In June, when it expanded the program, Anthropic said many partners already use the model to write the patches and to run pre-release checks that stop vulnerabilities from appearing.

That is the beginning of what I call living software. For decades, software was something that got finished: version 2.0 came out, went into production and someone reviewed it until the next one. What is coming is a continuous flow: AI detects a hole, writes the patch, tests it and deploys it. No final version and no human reviewing line by line.

To be fair: today almost nobody lets that happen without human approval. But the direction is clear, and it is also a necessity. If the attacker is going to be an AI, the defender cannot be a person reading a diff.

So, if AI writes, audits and patches, what is left for us?

Where there is a judge, AI already innovates

Here I would like to say that AI creates nothing new and only recombines what already exists. It would be comfortable, but it is not true.

In 2025, Google DeepMind presented AlphaEvolve, an agent that improves algorithms by modifying code and evaluating the results in cycles. In matrix multiplication, a problem studied for more than half a century, it produced algorithms that compute 14 different matrix sizes faster than any known method. That is not "another similar wheel". It is something no human had found.

The key detail is in the system's own description: it can be applied to any problem whose solution can be expressed as an algorithm and verified automatically.

That is the frontier. Where there is an automatic judge (faster, fewer operations, the tests pass, the vulnerability is gone), AI explores millions of options and finds new things. And it will do it better and better. Competing there means losing.

And notice the coincidence: that is exactly the terrain where we were trained to compete.

Where there is no judge, the human begins

Now think about problems where there is no automatic judge. Which product is worth building? What does it mean for something to be "better" for this company, these users or this moment? Which assumption of the industry is wrong?

There is a study that shows this difference very well. At Stanford, more than 100 natural language processing researchers blind-evaluated research ideas written by experts and by an AI. The AI's ideas were judged as more novel than the human experts', though slightly less feasible. Point for the machine, so far.

Then they ran the second part: they had researchers execute those ideas. And on many metrics the ranking flipped. The AI's ideas looked new but performed worse when put into practice.

AI can generate a thousand ideas that sound original. What it cannot do on its own is decide which of those thousand is worth it, because it does not want anything. We provide the intent (this matters to me, I want it to exist, I am going this way).

The only way to reinvent the wheel is not to use it

I have a saying of my own: the only way to reinvent the wheel is not to use it. And it turns out there is science behind it.

In 1991, researchers David Jansson and Steven Smith described a phenomenon they called design fixation. They asked students to design solutions, and showed some of them an example beforehand. Seeing that example restricted their ideas and even kept them from solving the problem. Whoever sees the wheel ends up drawing wheels.

Years later, Lars Bo Jeppesen and Karim Lakhani analyzed 166 scientific challenges posted on InnoCentive, a platform where large companies publish problems they could not solve. They found that the probability of winning increased with the distance between the participant's specialty and the problem's field: the further away, the better the odds. Their explanation is that outsiders have a "useful ignorance" of the dominant assumptions.

Careful: useful ignorance does not mean knowing nothing. Those outsiders knew a lot, but about something else. Whoever reinvents the wheel knows perfectly well what a wheel is. They are simply not tied to it.

And here a weakness of AI shows up. By how it works, it tends toward the most probable within everything that already exists. A study published in Science Advances measured it with writers: those who used AI ideas wrote stories that were rated better, but those stories were more similar to each other than the ones written by humans alone. Each one improves, but everyone ends up writing alike.

That is why, instead of asking AI to "reinvent the wheel", the human contribution is to pose a hypothesis that seems absurd: "I think this can be solved without wheels, help me find how". It is like asking it to make 2 + 2 equal 5. Orwell used that phrase in 1984 as a symbol of accepting an imposed absurdity. I use it the other way around: as the challenge of finding the system where what seems absurd makes sense.

It has happened in history. For more than two thousand years, that two parallel lines never meet was an unquestionable truth. Someone asked "what if they do?" and non-Euclidean geometries were born, which much later turned out to be necessary for relativity. The absurdity was not in the question, but in not daring to change the axiom.

Fundamentals for thinking, not for producing

This does not mean you should stop studying. It means you should study for a different reason.

You used to learn how a database works so you could build one. Today AI builds it. But if you do not understand how it works, you cannot imagine another way of solving the problem or notice which assumption could be broken. Nobody has an original idea about something they do not understand.

Technical knowledge shifts from a production tool to raw material for thinking. You do not need it to explain to AI what you want: it will understand you better and better, whatever you say. You need it to solve the problem in your head before asking for it.

Stop being a doer

The specialist in a single tool, the one who knows how to do X in a certain way, breaks when a problem outside their domain appears. And it is the first profile AI replaces, because it works exactly where there is an automatic judge: the code compiles or it does not, the test passes or it does not.

The developer who is coming looks more like a researcher than an operator. They understand the fundamentals, move between domains, distrust "that's how it's done" and have something they want to build. AI brings the capacity to execute. They bring the direction.

The question I leave you with is uncomfortable: if tomorrow AI does everything you know how to do, what is left that is yours?

If the answer is "nothing", it was not the AI. It was how you were trained. And that, unlike AI, does depend on you.

Sources mentioned in the article

ShareLinkedInX

Written by Sebastián Larrauri

Independent Claude consultant. Claude Certified Developer – Foundations (CCDV-F), Claude Certified Associate – Foundations (CCAO-F) and Claude Code Partner Badge holder, with the Architect certifications in progress.

Related articles