noun · AI concept
Knowledge cutoff
also called training cutoff
The date the AI stopped studying. It knows a lot up to then, and nothing it was not shown after.
The point in time after which a model's training data contains little or nothing, so its built-in knowledge stops there.
Training data has an end date, so events after it are absent from the weights, and the most recent months before it are often thin because less had been written about them yet. Fresh information has to come in through the input.
Real-life example
You ask which version of a framework is the newest and get a confident answer. By the time you read it, two newer versions are out. The model was right about the world as it was when it finished reading. Give it the release notes and it will happily talk about the new one.
Where the analogy breaks
-
The model cannot always tell you its own cutoff. It may guess a date that is a bit off, because it has no clock and the latest period is lightly covered.
-
The cutoff is not the release date. A model usually ships months after its data ends, and people keep using it for a long time after that.
-
A search tool or pasted document can fix it for one request. The model itself does not update.