Ubisoft revealed an AI NPC prototype at GDC and everyone online made fun of it

rockpapershotgun.com:

It's GDC week, and conversations are being had about AI. Ubisoft have blundered into this, in classic Ubi style, by revealing that they've had an R&D team beavering away on a project called NEO NPC. It's the usual pitch. "Have you ever dreamed of having a real conversation with an NPC in a video game?", asks Ubi's official post about it. And what all people who make this fail to understand that the answer most normal people give if they think about it is "Erm, probably no, actually?"

Still, the internet made good hay from the prototype image Ubisoft shared on Xitter, as many people took the opportunity to make fun of the dialogue from Bloom, a prototype man in a prototype beanie who wants to be your friend.

How NEO NPC works in practice is that writers create the backstory and personality of a character, and input this into the language learning model. Players talk to the NPC by actually talking, out loud, into their mic (which is the bit I think average consumers do not actually want to do - like, sitting alone in a room talking to thin air). The player speech is then used as a prompt for the AI to spit out a response theoretically in line with the character defined by the writer. Sometimes, if you ask the right question, it will trigger a pre-written piece of plot thread rather than an AI response.

Currently NEO NPC uses Nvidia’s Audio2Face application and Inworld’s Large Language Model to do al this. Ubisoft's post also gestures vaguely at the concerns people have with AI. VP of Production Technology Guillemette Picard says that "we know that developers and their creativity must still drive our projects. Generative AI is only of value if it has value for them.”

There's some interesting stuff in there to interrogate, which the post doesn't really spend any time on, such as trying to filter out "toxicity and inappropriate inputs on the part of the player" and figuring out how the AI will react to antagonism. There's also a small section on "catching bias" that