Can a model know about something that happened after it was built?
A model’s training data does not include events that happened after that data was collected. A language model without retrieval or other tools cannot fetch newer information on its own; newer facts can be supplied in its prompt or retrieved by surrounding software.
The cutoff describes the training data
A knowledge cutoff describes the latest period covered by a model’s training data, which may be assembled across different sources and stages. Newer information can still be supplied in a prompt or local file; fetching it from an online source requires a connected tool.
The cutoff is a slope, not a clean line
A cutoff is usually reported as a single month, but the data behind it was collected over a window, and coverage thins as you approach the end of that window. The world takes time to finish writing about an event: the follow-ups, the corrections, the reference pages. There is far less text about last month than about last year.
The practical shape of it is a slope. A model is solid on things well before its cutoff, patchy in the months just before it, and blank after. Fast-moving subjects feel the edge sooner than slow ones, since a field that publishes constantly has more coverage go stale in the same stretch of months. Later stages of training can also run on newer text than the main pass, so one model can hold knowledge of several different ages at once.
Does a model know today’s date?
Not by itself. A model has no clock and no calendar. Ask a bare model what day it is and it will produce a plausible date drawn from the text it saw, usually landing somewhere near the end of its training data. It is guessing, in the same way it guesses the next word.
Assistants normally get the date right because the software around the model writes it into the prompt before your question reaches the model. The date is an input, not a memory. The same is true of anything else current: the model is not recalling it, it is reading it.
Anything newer has to be handed over as text
There is one channel, and it is the prompt. Paste an article, attach a document, or let software run a search and paste the results back in; either way the model is reading text placed in front of it. None of that changes the weights, and when the session ends, the information is gone unless something outside the model stored it.
How much fits at once is set by the context window, which behaves like working memory for one sitting rather than a place where knowledge builds up. Without a connection, information that is not already local or supplied in the prompt cannot be fetched from an online source.
A cutoff describes the coverage of training data. Newer facts can come from your prompt or local files; fetching them from an online source needs a connection.