Modbox OpenAI-GPT3/Replica AI Scripting [video]

youtube.com

29 points·aniijbod··4 comments

It's great to see the whole chain of "speech to text to model to text to speech" in action. The huge amount of computation that has to be done shows in the delay during the API calls.

Before this is actually usable in a game or product the models must be made cheaper to compute and smaller in size.

That’s a cool demo (starts at 4:24). I assume these NPCs don’t actually know anything about the game, so there actually isn’t a “City Hall Street Location” in the game, and you can’t actually pay 10 coins and get a sweet hot dog. What would it take to hook this up to the reality of the gameplay?

The way to do it is to set the prompt to things based on the game world - so instead of 'City Hall' actually have that as a quest location to have the AI tell the player about. Could even dynamically change the Prompt based on the player's clothing / actions.

The Hot Dog vender in this example was just prompted as a 'hot dog vender' - but we could have told it what prices hotdogs were beforehand, then do gameplay actions based on the player's response (right now it only checks if 'bye' was said to end the conversation)

It depends, you prompt GPT with a (limited) amount of information and it will most likely keep it's answers consistent. I think that's what they did for the three types of hot dogs and the price. You can't prevent it from producing inconsistent answers though, especially if you ask nonsensical questions, so usage is very limited as of now. But awesome idea, hope we'll see more of it.

Reply on news.ycombinator.com