Homogenization | Avraam Mavridis<br>Homogenization<br>June 21, 2026·5 min read<br>AIphilosophyfoundation models

Hiddensee, Baltic Sea<br>In Greek we have a word "Ομογενοποίηση" that is a combination of 3 words:

όμοιος , meaning similar

γένος , meaning kind/type

ποιώ , meaning, the act of making

lately this word keeps coming to my mind, especially during working hours. Everyone is sharing documents generated by AI, everyone summarizes the AI generated documents with AI, everyone reviews the generated documents with AI and leaves comments with AI, same thing is happening to code, PRDs, RFCs, presentations etc etc. The documents get longer, the PRs are getting longer, and although I am tired of reading longer and longer "things" what annoys me the most is the lack of "character". Everything becomes the same and is hard to know if the author really stands behind the deliverable, homogenization (Ομογενοποίηση).

I also stumbled upon a book by Stefan Zweig, Η ομογενοποίηση του κόσμου ("Die Monotonisierung der Welt" — the German literally means "the monotonization of the world"), and there he wrote:

«Το ιδιαίτερο λεπτό άρωμα της ιδιαιτερότητας κάθε πολιτισμού εξαχνώνεται μέρα με τη μέρα, τα χρώματα ξεφτίζουν όλο και πιο γρήγορα∙ και κάτω από το μαδημένο σοβά αρχίζουν να διακρίνονται τα ατσάλινα έμβολα της μηχανής, της σύγχρονης παγκόσμιας μηχανής.»

(The delicate, distinct fragrance of each individual culture evaporates day by day; the colors fade ever faster. And beneath the flaking plaster, the steel pistons of the machine—the modern world machine—are beginning to show.)

It feels this homogenization is happening on both sides of the screen. Inside the machine and outside of it. Thirty years ago, a spam filter and a chess engine shared almost nothing. Different math, different assumptions, different people. Then everything started to merge, first everything became machine learning, then deep learning, then the transformer. Now everything runs on a handful of foundation models, adapted at the edges, every software 1.0 somehow calls somewhere a foundation model, no matter if the use case is useful or not.

A radiologist's tool and a lawyer's tool and a programmer's tool are the same.

The selling point is "Improve the base model and you improve everything downstream", but any flaw in the base is inherited by everything built on top. You get a single point of failure for an entire civilization's worth of applications.

And there is the homogenization to the other side of the screen, outside of the machine, where a few models become the tool and the way of how millions of people write, think, summarize, and decide, the homogenization does not stop at the model. It crosses over into the people.

I notice it in my own writing first. I sit down to draft something, reach for the model, and what comes back is competent and clean and average. Not wrong, but kinda average. It is the center of mass of everything ever written on the subject, and the center of mass has no edges. If I am not careful, I start writing toward it. My sentences drift toward the shape the model expects, because that is the path of least resistance. I am being fine-tuned by the thing I am using.

We are not all consulting different oracles. We are consulting the same one. A hundred million people asking the same model get variations on one answer, sampled from one distribution. The output feels personal because it arrived in a chat window addressed to you. It is not personal. It is the mean.

Machine learning already has a word for this, mode collapse. A model trained on enormous variety stops producing it. The range collapses and the edges go quiet. The term was coined for what goes wrong inside a model. I think it also describes what goes wrong outside. Its particular flavor of reasonable answer quietly becomes everyone's answer.

If the base model has a blind spot, and we all think through the base model, we all share the same blind spot. Worse, we cannot see it, because the instrument we would use to look is the same instrument with the blind spot. I know, the models now come with "self-reflection" but how many times did you push back on a model's answer and it replied "You are right yada yada...". Side note: It's funny that we use more and more terms that try to humanize the machine (Hallucination, Reflection, Attention)

Diversity of thought is how a field finds the answers no single mind would reach. It is friction, disagreement, conflicts, the person in the room who read the weird book or who had the particular past experience no one else in the room had and is valuable to share. Homogenization is the slow removal of that person.

There is a line about an earlier era of technology: the best minds of a generation were spent figuring out how to make people click on ads, now it seems we have another, the best minds of newest generations, all thinking through the same models, slowly thinking alike. There is a real danger that we will be...