Lesson 0.2
Why Models Make Things Up
A language model stores facts the way a sponge stores water: squeezed in, lossy, and with no receipt. That is why it guesses, and why a source fixes it.
The friend who never says "I don't know."
Picture someone who has read an enormous amount, remembers the gist of most of it, and is physically unable to admit a gap. Ask them anything and they answer smoothly, in the right shape, with the right tone. Often they are right. Sometimes they are confidently, fluently wrong. A language model is that friend.
The reason is in how it stores what it knows. During training, billions of sentences get compressed into the model's numbers (its parameters). Facts are not filed away as exact records you can look up. They are blended into the weights, the way many photos averaged together blur into a face that is no one in particular. The common, repeated facts survive that blending. The rare or specific ones get smeared.
A model's memory is lossy and has no receipt. It cannot point to where a fact came from, because it never kept the source. So it cannot tell its solid knowledge from its blurry guesses.
Two gaps memory cannot cover.
Even a perfect memory of training data would still miss two things, and these are exactly where real questions live:
The knowledge cutoff
Training ended on some date. Anything after it (a release last week, a price that changed yesterday) simply is not in the weights. The model will still answer, by guessing in the right shape.
Private and specific facts
Your company's docs, this customer's order, that internal wiki page: the model never saw them. They are not rare in its training data, they are absent from it.
Ask "what shipped in our Q3 launch?" and a pure-memory model has no honest answer available. It produces a plausible one anyway. That is not a bug to be patched out; it is what answering from a blurry, source-less memory looks like.
Retrieval gives it a receipt.
The fix is not a smarter model. It is a source. If you find the passage that answers the question and place it in the model's context, the task changes from "recall this from memory" to "read this and summarize it." The second task is far easier and far safer, and the model can now cite where the answer came from. That is the entire promise of RAG, and the rest of the course is how to do the finding well.
Hallucination is the cost of asking memory to do retrieval's job. Hand the model a source to read and most of it goes away.
Key takeaways
A model stores facts blended into its weights: lossy, and with no source to check against.
It will confidently answer past its knowledge cutoff and about private data it never saw.
Retrieval reframes the task from recalling a fact to reading one, which is what stops the guessing.