Each character had all their Bethesda-written dialog fed to the LLM as a starting prompt, so each one communicated with the player within the bounds of their authored identity, and knew what that character was written to know.
The human player used freeform in-game text-to-LLM as input to talk to the people in the town square. The latency was large, but not more than a few seconds.
He was able to walk up to someone and say things like "Where do you live? Who is your father?" .... latency pause ... "Oh me? I live in the Cloud District. It's where all of us Battleborn clan live, of course!"
The real kicker of this demo was that they passed the LLM-generated response dialog in to a deep-fake model trained on each character's original voice actors (since the corpus in Skyrim is vast, this ended up quite accurate).
All of this was a major generational shift for gaming as a whole. I had just witnessed the first time a video game player could actually TALK to an NPC, in a fully unstructured and unscripted fashion, while each character remained as in-character as the original game's authors intended.
So maybe this purchase by OpenAI is about something like that. Building NPCs for Biomes.gg that you can actually talk to.