1. What a Model Actually Does

A large language model (LLM) does one thing: it predicts the most plausible next piece of text. Give it a sentence and it calculates what words are most likely to follow, then repeats that process token by token until it has a full answer. That single trick, scaled up with enormous amounts of training text, is what lets it summarize, translate, write code, and answer questions.

Critically, it is not searching a database of facts. It has no inner encyclopedia it consults. It produces whatever continuation sounds right — which is why it can be impressively helpful one moment and confidently wrong the next. Once you see it as a text predictor rather than a knowledge engine, both its strengths and its failures start to make sense.

Models read and write in tokens — small chunks roughly equal to a short word or part of a word (about 4 characters in English). "Hello" is one token; a long word like "anticonstitutionnellement" is several.

Every request has a context window: the total number of tokens the model can consider at once. This window holds your message, the conversation history, any attached documents, and behind-the-scenes instructions. Think of it as the model's working memory for that single request. When the window fills up, the oldest parts are dropped. The model cannot remember anything that has left its window unless something puts it back in.

Because the model generates plausible text rather than verified facts, it will sometimes state false things with total confidence — a fake reference, an invented statistic, a feature that doesn't exist. This is called a hallucination. It is not a bug you can switch off; it is a built-in property of how prediction works.

The remedy is grounding: giving the model verified material to work from. That can be your own documents (a knowledge base), a live web search, or a tool that returns real data. Grounded answers can point to where a claim comes from; ungrounded answers cannot. Whenever accuracy matters, ground the model — and still verify the output yourself.

One common misconception: talking to the model does not teach it anything. A model is fixed once trained; your conversations don't change its weights. It also has a knowledge cutoff — a date after which it knows nothing because its training data stops there.

So how does the experience improve over time in OUPI? Not by changing the model, but by enriching what goes into the context window: your saved memory, your skills, your uploaded documents. The model stays the same; the context it receives gets smarter. And your data is never used to train anyone else's model.

Finally, remember the difference between a model and a product. The model is the prediction engine. The product wraps it with instructions, memory, tools, safety rules, and an interface. Two products using the same model can behave very differently.

In OUPI, the same models power the chat, the agents, and Conductor (the agent layer). A chat answers once; an agent plans, uses tools, checks results, and iterates until a goal is reached. But underneath, it is always the same principle: predict the next token — just with different instructions and capabilities plugged in.