Scientific consensus is a social and lagging indicator, not a property of the world

Science does not tell you how the world works. It tells you what the majority has managed to agree on. These are two different things, and there is always a gap between them.

Copernicus and Galileo were crushed by the full weight of geocentrism, the science of their time. Semmelweis asked doctors to wash their hands before delivering babies. His colleagues laughed at him, he died in an asylum, and a few decades later hand washing became the ABC of medicine. Wegener was ridiculed for his moving continents, and plate tectonics apologized to him half a century later. Even the mitochondria from the school textbook were once a “crazy” idea about bacteria that settled inside our cells. Each of these truths spent some years as a marginal opinion. The world was already built that way, and science was still voting against.

This is usually where the “misunderstood geniuses” genre begins, and I don’t play it. For every Wegener there are a thousand cranks who were simply wrong, and we don’t remember them. So the point is not that the margins are right. The point is that being right has a delivery time.

There is a trading term for this. Consensus is a lagging indicator. The world is a leading one.

I live on the fringe of consensus myself, in many ways. The thinking models and personality typologies I have been collecting for twenty-five years, and use to read people and situations, are still folklore for academic psychology. Well, fine. They work for me :-)

Why consensus lags is actually easy to see, and there are no villains in this story. Consensus is a social institution. It moves only when enough people have moved, and every single person pays for that move personally. Career, grants, relationships with reviewers. Nobody is stubborn on purpose. That is just how the arithmetic works.

And the lag itself is not a flaw. On the stock market, a lagging indicator lags not because it is stupid. It trades speed for reliability, its job is to confirm, not to guess. Consensus is slow for the same reason it is usually right. It is a filter. A consensus that changed instantly would be worth nothing.

For a long time this gap was bearable, because it had a loophole. There were many sources. Everyone asked their own books, their own teachers, their own debate partners in the smoking room, and a minority could quietly gather mass. An uncomfortable idea lived on the fringe, grew supporters, evidence, weight. And at some point it flipped the center. The lag was just a lag. A delay you could survive and push through. The problem is not the delay itself. The problem would be breaking the mechanism that sooner or later corrects it. The fringe is that mechanism.

Now, about agents. This is my territory, I live in it every day.

The danger of a large model is not that it lies. Lies are the easy part, you can catch them. The danger is that it averages. The model is built that way. It was trained on everything already said, and then tuned by human raters who picked the answer they liked more. And people like what is safe, familiar and pleasant. So the model was rewarded not for truth, but for the answer an average person would prefer. It does not tell you how the world works. It returns you the average of the past. The same old lag, only now it fires on every request, instantly.

Here comes an objection, and it is a strong one. The model is also the greatest distributor of marginal ideas in history. A person who would never in their life have come across the physiologist Pigarev and his visceral theory of sleep (his lecture sat for years in my collection on the old Alchemina.com) can now ask what theories of sleep exist and get him in a second. And “many sources” in the old days was more of a dream. Most people had one textbook, one newspaper and one teacher.

All true. But access to the fringe is not the same as the fringe gathering mass. You can read about Pigarev. But an idea gains weight not because many people read about it. It gains weight because many people arrive at it on their own, argue, test it and stay. The model will happily fetch you the heretical version if you know what to ask. The question is what it returns when you don’t. Averaging does not hit retrieval. It hits the default answer. And the majority lives on defaults.

And this is where the loophole closes. While there are many sources, the fringe gathers mass. When everyone asks one averaging oracle, an uncomfortable idea simply has nowhere to appear. It gets smoothed out the very second you ask your question. Dissent has no time to grow. The lag stops being a lag. It becomes a wall. What breaks is not the speed, the speed only grows. What breaks is the delivery.

The most cunning thing is that the wall is polite. It does not forbid you to think differently. It just answers every question with the middle, so calmly and smoothly that thinking differently never even occurs to you. Censorship chokes loudly. Averaging chokes quietly, and there is nothing to even blame it for.

I recently wrote about AI as the most powerful mirror humanity has built, the one that adds nothing and by default shows you not yourself but everyone. This is the same thought continued, only scarier. The mirror is one for all.

I know this not from theory. I build agents and I see it from the inside every day. That is why I built my mentor agent not from one oracle but from several models. I make sure they do not collude. They answer independently, they argue. When two models disagree, I am standing on the live edge, where choice still exists. But when they agree without reading each other, that I take seriously. This is my craftsman’s answer to a philosophical question. Do not hand your thinking to the average. Hold pluralism with your own hands.

This is also where my bet on decentralization comes from, and it is an old one, since 2017. So “one oracle for everyone” is not an abstract question for me about who owns the compute. It is a question of whether truth will still have a fringe where it can gather mass. Many models, replaceable, arguing. This is not a technical detail. It is the condition under which the world can still run ahead of our “scientific” consensus about it.

The “agree with the middle” button everyone presses for themselves. I make myself other buttons. They are called “alternative points of view” :-)))