Is that understanding?

A question has been building through this chapter and it deserves to be asked out loud rather than left to sit.

The space may put grief near loss, mourning, and bereavement. The model can learn that people are described as stricken by grief and rarely as purchasing it. Some vector relationships can also line up.

That is a great deal of structure, and every bit of it was worked out from the company those words keep. No one wrote a definition. The model has never lost anyone, never sat with anyone who had, never been to a funeral or felt the specific weight of a Tuesday afternoon three weeks afterwards.

So: does it understand the word grief?

There is a case for a limited yes. Much of language use depends on relationships. A model that captures those relationships has learned something more useful than a dictionary label.

There is also a case for no, or at least not in the same way. Human meaning is tied to bodies, perception, action, memory, and other people. A text-trained model learns from traces of those experiences in language, not by living through them.

This course will not settle the word understanding. The argument is older than this technology. You met one version of it with the Chinese Room.

What has changed is that you now know exactly what is in the machine when the question is asked. The final module returns to it with more to work from.