Frozen in time

The second edge is about time, and it is the simpler of the two.

Training ends, and the weights stop moving. They took months to settle and they are now simply a file, the same on every copy, unchanged by anything that happens next. Talking to the model does not adjust a single number. Ten million people talking to it does not either.

So whatever the model knows, it learned before you arrived, from text collected up to some particular moment. That moment has a name, the training cutoff, and the world has carried on past it without the weights noticing.

The consequence is easy to miss. The model does not experience its knowledge as old. Ask about something that changed last month and it will answer from the world as it stood at the cutoff, in exactly the tone it uses for anything else, because from the inside there is no difference between the two.

Applications work around this by fetching current text and putting it in front of the model, which is a good solution and a genuinely different thing from the model learning. The next module is about that outside layer and what it can and cannot fix.