It stopped reading at some point

A model’s training text ends on a date. After that date it knows nothing, and it cannot tell.

Part of the What is this thing? track on lAItest.

Ask a model what happened last week and you can get a confident, detailed, completely invented answer.

Its reading stopped, and nobody told it that it had.

The cutoff

Training text has to be collected before training can start, so every model has a knowledge cutoff: the point past which it saw nothing. Events after it are not merely unknown — they are absent in a way the model cannot detect from the inside. There is no sense of today’s date in there, and no gap where the missing months should be.

A model ships long after its reading stopped

A common misconception

Commonly believed: A newer model always knows about more recent events than an older one.

Actually: Release date and cutoff are two different dates. Two models from the same maker released weeks apart can share an identical cutoff, or differ by months. The version number tells you nothing about it — only the model’s own documentation does. Guessing from the number is how people end up trusting a stale answer.

The fix is to put it on the page

The prompt is the only other place information can come from, so current information has to be pasted in: a search result, a document, a lookup from a database. This is called grounding, and a whole later track is about doing it well. The part that matters now is that the model is not recalling the news. It is reading it, in the prompt, like anything else.

In one sentence

A model’s knowledge has an end date that it cannot feel. Anything after that date has to arrive on the page.