Why LLMs Feel Different
(the payoff)
Every AI tool before the LLM did one thing. A spam filter filtered spam. A translation model translated. A recommendation engine recommended. Each was trained separately, for one task, on one kind of data. LLMs broke that pattern. One model. One set of weights. Translation, writing, summarisation, code, reasoning, all from the same machine.
The one trick
Predicting the next word turns out to be a proxy for understanding language. To predict well, the model has to learn grammar, facts, logic, tone, structure, and intent. It does not learn these because anyone asked it to. It learns them because they all help it guess the next word better.
By the time the model is good at predicting text, it has absorbed most of what humans know and write about. Ask it to translate, and it draws on what it learned from translation examples. Ask it to write code, and it draws on the code in its training data. The same weights serve every task.
Input
Bonjour, comment ca va?
LLM
same weights
Output
Hello, how are you?
No retraining. No different model. The same weights handle all three.
Three completely different tasks. One model. Switch between them. The weights never change.
A familiar example
A person who has read widely can do more than someone who has only studied one subject. They can write clearly, follow an argument, spot a logical flaw, explain a concept to a child. They did not study each of those skills separately. Reading taught them all of it. An LLM trained on the full breadth of human text gets the same general capability. Training data alone produced it.
The breaking point
The model has no concept of truth. It learned to predict what humans write. Humans write things that are true, and things that are false, and things that sound true but are not. The model learned from all of it. When it produces a confident, fluent, well-structured wrong answer, it is not lying. It is predicting the kind of text that usually follows this kind of prompt. The fluency is real. The accuracy has to be checked separately.
Your takeaway
For the first time, a single tool can help with almost any language task. That is new. Every specialised AI tool you used before required a team to build, train, and maintain it for one job. LLMs collapsed that to one model, one API call, and whatever you type in the box.