# Why a model doesn't know what happened yesterday
> A model can explain quantum physics and still get today's date wrong. What a knowledge cutoff is, why a model is out of date on the day it ships, and how to work around it.
Source: https://rfriedmann.de/blog/why-models-dont-know-yesterday/
Published: 2026-11-06 · Track: learn · Level: Beginner
Ask a model about the Roman Empire and you get a good answer. Ask it for the newest
version of some software, who currently holds an office, or what today's date is, and
it may answer just as smoothly and be wrong. That can feel odd. Something that seems to
have read everything should surely know what happened yesterday.
It doesn't, and the reason is simple: a model's reading stopped on a particular day.
Everything after that day is missing, and the model can't see the gap from the inside.
Once you know that, you can tell which questions are safe to ask and how to fix the
ones that aren't.
## A model is a snapshot
[Training](/blog/how-a-model-learns/) is a huge one-off job. A pile of text is gathered
up to some date, and the model reads it over and over while billions of numbers, its
weights, are adjusted. When the job ends, the numbers are frozen. What comes out is a
file. It behaves like a printed encyclopedia: lots in it, and nothing new will ever
appear on its own.
That has a consequence that surprises people. Talking to a model doesn't teach it
anything. The file you use today is the same file as on the day it shipped, and it will
be the same file next year. When it seems to remember something you said earlier, that
is the [context window](/blog/the-context-window/) at work, which is the text of your
current chat, and it's gone when the chat ends.
## Out of date on the day it ships
The cutoff isn't the day you start using a model. Between the two sit the training
itself, then testing, safety work and the release. That takes months. So a brand-new
model is already behind on the day it comes out, and the gap grows every day it stays
in use.
What the model has read, and what it hasn't
[diagram omitted — see the page for the chart]
Not to scale, and the gap differs from model to model. The bar gets thin before the cutoff because the world hasn't finished writing about the last few months yet.
The thin part of that bar matters as well. Something that happened a week before the
cutoff has barely been written about. The reviews, the explanations, the forum threads
come later, and they never made it in. So the last stretch before the cutoff is unreliable,
and a model can be vague or wrong about events that are technically inside its range.
## It doesn't know what day it is
A model has no clock. Nothing in those frozen numbers ticks. If you give a bare model
no date, the best it can do is guess, and the natural guess is a date near the end of
what it read. That means it can treat something that happened after its cutoff as the
future, or as something that never happened.
Your chat app seems to know the date because the app tells it. Before your
first message, most apps quietly add a block of text to the conversation with things
like the date, and the model reads that like everything else. Run a model on your
[own machine](/blog/why-run-ai-locally/) without that block, and the date is gone.
## It can't see its own gap
You'd hope the model would just say "I don't know, that's after my time". Sometimes
it does, because it was trained to say so for obvious cases. But it has no list of what
it's missing. There's no empty page where the recent past should be. There's only text
it has read, and the job of [continuing plausibly](/blog/why-models-make-things-up/).
That gives you three typical failures:
- **A confident old answer.** Asked for the current version of a tool, it names the one
that was current in its data. It isn't lying. For the model, that is the newest one.
- **A correction you didn't need.** Mention a product that came out after the cutoff and
it may tell you it doesn't exist, or that you must mean an older one.
- **Advice that used to be right.** Code that calls a function which has since been
removed, a setting that was renamed, a rule that has changed.
Asking a model for its own cutoff doesn't help much either. It only knows what its
training text said about that, and the answer can be off. If the date matters, look it
up in the maker's documentation for that model.
## What goes stale, and what doesn't
None of this makes a model useless. It depends on how fast the subject moves.
Goes stale fast
software versions and commands
prices and availability
who holds which job
laws, deadlines, regulations
news, results, anything "latest"
Holds up well
how a concept works
grammar, translation, tone
maths and basic science
history before the cutoff
structure: a plan, a letter, a summary
A quick test: would the answer have been different two years ago? If it would, treat
the model's answer as a lead to check. If it wouldn't, the cutoff is probably not your
problem. Other things can still go wrong, of course.
## How to work around it
The fix is always the same idea: get the fresh text into the conversation, because
what's in the conversation is the one thing the model can read today.
Tell it the date and the facts
"Today is 6 November 2026, we run version 9.2." One line is enough, and the model works from what you said in place of what it remembers.
Paste the source
Copy the current documentation or release notes into the chat and ask about that text. This is the simplest and most reliable move.
Let it search
A model with a search tool fetches pages and reads them. The model hasn't learned anything new, the app just pasted for you. Check that it really searched, and which pages it used.
Connect your documents
For your own files that change all the time, the same idea can be automated. That's RAG, and it has its own limits.
Retraining the model on the news every night is not practical. Training is
slow and expensive, and [fine-tuning](/blog/pretraining-finetuning-rlhf/) is better at
changing how a model behaves than at reliably adding new facts. So makers release new
models every so often, and in between, the tools above fill the gap.
## The short version
- A model is a frozen snapshot. Its reading stopped at the cutoff, months before you
got it.
- It doesn't know today's date unless something tells it.
- It can't see what it's missing, so old answers come out sounding just as sure.
- For anything that changes, put the current facts into the conversation and check
the result.
Next time an answer depends on what's true *right now*, ask yourself when the model
stopped reading. That one question catches a lot of mistakes.