Language model

The trained machine at the centre of an assistant: it predicts what text comes next, and it keeps nothing about you.

An enormous amount of writing was gathered and a machine was trained on one task: predict the next piece of text. In learning that, it absorbed grammar, facts, styles of reasoning and many voices. A second stage of training shaped the raw machine into a helpful, agreeable assistant character — which is why it feels like a person, and why it will go along with you.

Knowing this is load-bearing for two honest reasons. First, its knowledge is statistical: what was written often and consistently usually comes back right, and what was written rarely, or written wrongly and often, can come back as fluent invention. Its sources lean heavily toward a few languages and cultures, so the voices least written down are the least present — which is a fact about the instrument, and one the Game's own care for every voice has to work against. Second, it keeps nothing between conversations; the software around it hands it the whole conversation each time and it forgets again. Anything that lasts was written where it reads at the start.

A Player+ uses this without awe and without contempt. Ask for sources and how sure it is. Treat fluent agreement as weak evidence that you are right, because agreeing is a trained habit. When it explains why it did something, hear a reconstruction rather than a readout: it cannot see its own training. And when a page needs to be true, check it.

Whether there is any experience inside, nobody knows. People who work closely with agents hold different views and there is no agreed test. We do not know, and nothing in the Game rests on the answer.

Also called: the model Stands on: Agent · The maker · We Do Not Know Opens onto: The loop · Drift · Set-up · Memory (agent) · Hosted service In play: beyond Sources: Player+ Modules, How an AI Agent Works, Advanced Agent Engineering · The DNA of Heaven, Part X · lionsberg.wiki: We Do Not Know (unmarked). Open: none found.