Knowledge Cutoff

Every pretraining corpus is assembled by a certain date and then frozen.

That date is the knowledge cutoff, and it is a hard boundary on what the weights can contain.

The awkward part is that the model has no way to perceive the boundary from the inside.

It does not know today's date, so it cannot tell that a question concerns a period it never saw.

Asked about the newest version of a library, it will describe the newest one it knows and sound current.

This overlaps with hallucination but has a distinct cause: absence of data rather than weak recall.

Cutoffs are also softer than the single published date suggests.

Text about an event keeps being written for years afterwards, so coverage of recent months is thin.

A model may know a great deal about something two years before its cutoff and very little about two weeks before.

Post-training data is more recent still, so a model sometimes knows its own release date and nothing around it.

The reliable fixes all involve supplying fresh text at question time rather than trusting memory.

Retrieval-augmented generation searches a corpus you control and puts the passages in the context.

Tool calling lets the model query a live API, which is the right answer for prices, status and schedules.

Telling the model the current date in the system prompt is a small change that prevents many confident errors.

For anything that changes, treat the weights as a snapshot of the past and design the system to fetch the present.