Large language models feel magical at first.
You can ask them almost anything. History, programming, psychology, cooking, physics. They answer instantly. The experience gives the impression that the model knows everything.
But that impression is slightly misleading.
What the model really knows is something closer to the average knowledge of the internet.
And average knowledge is not the same as expert knowledge.
The Averaging Problem
Imagine you trained a model on the world's best books.
Carefully researched history books. University textbooks. PhD-level material written by experts who have spent decades thinking about their subjects.
That sounds like a recipe for intelligence.
But now imagine you mix those books with the rest of the internet:
- Online forums
- Social media posts
- Comment threads
- Wikipedia edits
- Random blogs
Now the model is learning from both.
When this happens, something subtle occurs.
The expert information doesn't disappear. But it gets diluted.
If ten people online say A, and one expert book says B, the model doesn't necessarily know that B is correct. It only knows that A appears more often.
So it learns the pattern:
Most people say A, therefore A is probably true.
The model isn't checking truth. It's checking probability.
And probability reflects what people say, not necessarily what is correct.
So the model begins to mirror something very familiar:
The collective knowledge of the internet.
Which is useful. But imperfect.
Why This Is Still Useful
Despite this limitation, large language models are still incredibly valuable.
Because most people don't need expert knowledge about everything.
In fact, no one does.
Human knowledge has always been distributed.
We rely on specialists:
- Engineers build bridges
- Doctors study medicine
- Historians study the past
- Programmers build software
No single person understands everything.
Civilization works because we build on top of each other's expertise.
Large language models tap into that shared knowledge pool.
For many topics, they give you something like the community answer.
And the community answer is often surprisingly good.
Better than nothing. Often much better.