The Great Consolidation
I’ve noticed something when using GitHub Copilot. If I start writing code to analyze data, it almost always suggests Python and the pandas library. Usually a good suggestion. But it is always that suggestion. It feels less like magic and more like walking a path worn smooth by thousands of people before me.
This is a sign of a bigger shift happening right now. Large language models are powerful agents of consolidation. Trained on a vast snapshot of the world, they reflect it back to us and amplify its most dominant ideas and habits. They freeze our culture in digital amber.
You see this most clearly with language. LLMs learn from the internet, books, and articles, so they learn the slang we use, the way we argue, the stories we tell. Ask one to write a casual email and it might throw in an “LOL” or a phrase that feels distinctly like Twitter in 2022. It captured the linguistic fashion of a specific moment.
On the surface this seems harmless. But most of the internet is in English, and much of its cultural gravity is American. Ask an LLM an abstract question about “freedom” and its answer will be shaped by Western philosophy, particularly American. Getting a perspective from an Indigenous tribe in the Amazon takes real work. The model presents a default view, and that default is whatever is most common in its training data. Globalization in overdrive, smoothing out the weird and wonderful edges of culture into a single global dialect.
The same thing happens with code. Copilot suggests pandas because most public code uses pandas. It suggests React because React is all over GitHub. This creates a feedback loop: the most popular tools get suggested, which makes them more popular, which ensures they dominate the next model’s training data.
This could make it harder for new ideas to spread. A breakthrough programming language or a brilliantly efficient web framework might struggle to get noticed, because the AI assistant millions of developers use daily never mentions it. Like trying to find a new restaurant when your map only shows McDonald’s.
Are we all doomed to use React and say “yeet” forever? Probably not. Culture moves too fast. By the time a massive model is trained and deployed, the slang it knows already sounds dated. New ideas, memes, and ways of speaking bubble up from the edges. An LLM is a lagging indicator of culture, not its source.
The same is true for programming. Programmers are restless innovators. Rust happened because developers wanted something better and safer than C++ and pushed for it. A good programmer can ignore Copilot’s suggestion and try that new experimental library. Human curiosity is a powerful antidote to consolidation.
And the models themselves aren’t perfect monoliths. Their training data is huge and contains multitudes. Ask, and they can often write in a rare dialect or generate code in a niche language like Haskell. The seeds of diversity are in there, even if the main path is paved with whatever is most popular.
What we’re seeing is a new tension. On one side, LLMs push for centralization, for a shared standard. On the other, human creativity pushes for decentralization, for variety, for change. The future probably won’t be a bland monoculture. It will be a landscape with a huge six-lane superhighway running through it, built by AI, and thousands of smaller, more interesting dirt roads built and explored by people.
The most interesting thing about these models might be what they tell us about ourselves. They are a mirror. What they choose to consolidate is simply what we, collectively, are already doing the most. If we don’t like the reflection, we’re the only ones who can change it.