Feedback Is Evolution

I’ve always sensed that feedback matters, but I never really understood why.

Over the past couple of years, a mix of reading and lived experience has gradually reshaped my thinking. Feedback isn’t an evaluation. It isn’t some workplace technique. It’s the only way we extract information from the real world, the foundation of every action we take — and arguably one of the most fundamental mechanisms by which the world operates.

Starting with AI “agency”

One of the hottest concepts in AI recently is agentic — getting AI to autonomously complete an entire task, rather than following your instructions one prompt at a time. How do you get there?

You might assume the key is making the model itself more powerful. But in practice, the decisive factor often isn’t model capability — it’s whether you’ve given it a feedback mechanism.

Take AI writing code. If you simply say “build me feature X,” what it produces will almost certainly have issues, requiring you to manually adjust it over and over. But if you give it a test suite — letting it run the tests, see the gap between output and target, and correct itself — it can converge toward what you want, step by step.

This is really the same logic as AI training itself. During training, a model needs a loss function — it adjusts its parameters by continuously comparing its output against the target. Without that feedback signal, no model, however large, can learn anything. It’s just making random guesses.

So the core of making AI “autonomous” isn’t making it smarter — it’s giving it the ability to perceive gaps and self-correct. In other words, it’s building a feedback loop.

This made me pause and think for a long time. Why does such a simple mechanism produce such an enormous difference?

We were never taught “feedback”

To answer that question, you first have to confront something: in the education most of us grew up with, there’s simply no such thing as a “feedback” mindset.

School taught us to master the material first, then prove ourselves in a single shot on the exam. Bombing a test wasn’t called “receiving feedback” — it was called “not studying hard enough.” The entire education system rests on an implicit assumption: there exists a correct answer, there exists a path to it, and your job is to walk that path to the end — not to fumble your way forward through trial and error.

That inertia doesn’t vanish after you start working. I used to approach everything with a one-shot mentality — not because I thought revisions signaled incompetence, but because I was in the habit of polishing things on my own, always feeling that if I just thought a little more, revised a little more, I could make the work good enough before handing it over.

But I was overlooking something: as long as the work stays in your own hands, it hasn’t truly received feedback. Real feedback must come from the real world — from users, from the market, from external environments you don’t control. Revising a draft ten times on your own and putting it in front of someone once yield entirely different orders of information.

Netflix co-founder Marc Randolph has a formulation I deeply agree with. He repeatedly urges entrepreneurs to do one thing: collide your idea with reality as quickly as possible¹. Don’t refine it in your head. Don’t model it in a slide deck. Get something out there — however rough — and see how the real world responds. His reasoning is simple: whether an idea is good isn’t something you can judge yourself. “Nobody knows anything” — only real-world feedback can give you the answer.

The issue isn’t just that getting it right on the first try is impossible. It’s that you must recognize that obtaining feedback from the real world is, in itself, more valuable than spending three extra days thinking in a room.

Evolution: a profoundly underestimated word

So why does feedback carry such power?

What truly brought clarity for me was a handful of seemingly unrelated books.

The one that influenced me most was David Deutsch’s The Beginning of Infinity². He advances an audacious thesis: evolution isn’t merely a subtopic of biology — it’s the shared logic running through the physical world, the living world, and the world of knowledge.

When we learn Darwinian evolution, we usually understand it as “how species came to be.” But Deutsch shows that evolution’s reach extends far beyond that. The growth of knowledge, the transformation of cultures, even the formation of physical structures — all follow the same generative mechanism. Once this insight enters your thinking, everything looks different.

Popper arrived at this territory even earlier. Reading his Conjectures and Refutations and The Logic of Scientific Discovery³ hit me hard. He describes the growth of knowledge as a cycle of “conjecture and refutation”: propose a hypothesis, attempt to falsify it, revise in the face of failure, then propose a new hypothesis. This process is structurally isomorphic with Darwinian natural selection — not as a metaphor, but as a shared logical structure.

Add in Jared Diamond’s The Third Chimpanzee and some readings in evolutionary psychology, and a complete picture emerges: evolution describes not a biological process, but a universal generative mechanism — the basic way complex structures emerge from simple rules.

Break that mechanism apart and you find three core elements: variation, selection, and replication.

Variation is the source of new possibilities. New genetic mutations, new theoretical hypotheses, new ways of working.

Selection is the real world’s feedback. What works and what doesn’t isn’t something you reason out — it’s something the environment tells you.

Replication preserves the solutions that survived selection and spreads them.

All three are indispensable, and they must operate in a loop. One cycle isn’t enough. Two aren’t enough. The loop must keep turning.

Feedback is evolution’s prerequisite

Once you grasp this framework, “feedback” takes on an entirely different weight.

It’s the necessary condition for evolution to occur. Without feedback, there’s no selection pressure. Without selection pressure, variation is nothing but blind random drift. And random drift doesn’t produce any meaningful structure. A system without feedback cannot evolve, and a system that cannot evolve will ultimately be eliminated in a world that never stops changing.

This also reveals the deeper meaning behind Marc Randolph’s words. Why collide your idea with reality as soon as possible? Because only collision generates selection pressure. No matter how perfectly you model things in your head, that’s just variation — it hasn’t been through selection. And variation that hasn’t been through selection, within the evolutionary framework, amounts to nothing.

Every intricate, complex structure in this world — from the optical design of the human eye to the ecosystem of a tropical rainforest, from the grammar of human language to the price mechanism of a market economy — not a single one was “designed in one shot.” They are all products of evolution, the result of countless rounds of variation, selection, and replication. If even nature can’t get it right on the first try, what makes us think our work plans can?

Two worldviews

Your attitude toward feedback reflects one of two fundamentally different worldviews.

One holds that the world’s basic mode of operation is cycling, iterating, evolving. People with this view treat feedback as a necessity — not because they aren’t smart enough, but because this is simply how complex structures come into being. “Not yet finished” is the normal state, not a failure. Every collision with the real world accumulates the information evolution needs.

The other holds that the world’s basic mode is design, execute, achieve. People with this view treat feedback as a remedial measure — ideally you shouldn’t need it, and needing it means something went wrong earlier in the process.

Popper has a famous metaphor: science isn’t built on solid bedrock but on piles driven into a swamp. We don’t stop driving piles because we’ve found firm ground — we stop because the piles feel stable enough for now³. We will never reach the moment of “finally done.” We only reach stages where “good enough for now” holds.

Build the feedback mechanism from the start

With all this in mind, one thing has become increasingly clear to me in practice: you can’t bolt on a feedback mechanism after the fact. It must be built in from the very beginning.

The essence of business is a ceaseless loop — ship the product, the market responds, you adjust, ship again. How fast and how accurately that loop turns matters more than any individual’s ability on the team.

I’d even go so far as to say that a team with a few brilliant people but no good feedback mechanism will be outperformed by an average team with a well-functioning feedback loop. The reason is simple: the former is making one-time guesses; the latter is evolving. One-time guesses rely on talent. Evolution relies on mechanism. Talent has a ceiling. Mechanism doesn’t.

So the most important question may not be “how do I get this right on the first try,” but “have I thought, from the very beginning, about where my feedback will come from?” Is it real usage data from customers? Price signals from the market? Direct reactions from colleagues? Or some measurable target you’ve set for yourself?

Work without a feedback mechanism is like model training without a loss function — you might be working hard, but you have no idea which direction you’re heading.

Building a feedback mechanism is what gets the wheel turning. And only when the wheel turns can evolution happen.

Feedback is evolution

Back to the original feeling. What is feedback, really?

It’s not an evaluation. It’s not “getting other people’s opinions.” It’s the concrete manifestation of “selection” in evolution — the real world’s genuine response to your actions. From DNA replication error-correction to the immune system’s antibody screening, from the conjecture-and-refutation cycle of scientific theories to the price signals of a market economy, from the loss function in AI training to the customer data driving a business — the same logic appears again and again.

Without feedback, variation is blind, replication is directionless, and evolution cannot occur.

We aren’t optimizing a process. We are participating in evolution.