Frozen in time

When training ends, the weights freeze. From that moment the model does not learn anything new. It does not remember your last conversation. It does not update when you correct it. Every chat with it starts from the very same frozen numbers, like a clock stopped at one exact second.

So the model's knowledge has a cutoff: the point where its training data stopped being gathered. Anything after that simply does not exist for it. Recent elections, this morning's news, a result published last week, a company that changed hands yesterday. None of it is in the weights, because none of it was there when the dials were set.

You may have seen a model search the web or pull up a document mid-conversation. That helps, but it is a workaround bolted on the outside, not the model learning. The weights stay frozen; the tool just fetches fresh text and hands it to the model to read, the way you might pass a note to someone whose memory stops at a certain date.

There is a quieter version of this problem too. The model can be confidently wrong about things that were true when it trained but have since changed. Scientific consensus shifts. Laws are rewritten. The frozen model still answers from the world as it was at its cutoff, not the world as it is now. Knowing roughly when a model's clock stopped is basic hygiene for trusting what it tells you.