Hacker News Reader

Comment by sho | Hacker News Reader

So, would anyone like to take a stab at the stack/process/setup enabling this? Because I seriously wouldn't even know where to start.

View on HN

In their Show HN post they mentioned some bits about the tech stack: https://news.ycombinator.com/item?id=33833645

The visuals are rendered in Unity (I guess in real time), maybe their models spit out dialog including stage directions?

Other links related to the show: https://linktr.ee/nothingforever

Wow that's post had 3 comments when it showed up 2 months ago.

There could be a huge business model here, cut costs massively on new TV shows or cartoons getting c churned out.

Have a look at the comments and submissions of the user "skytrue". They seem to be one of the creators of this.

https://news.ycombinator.com/threads?id=skytrue

Probably it's just a little unity world with a visual filter, some models and movements that those models make, and then you train GPT-3 by giving it the API of what scenes, actions, and camera angles are available, and have it generate those in combination with a script and laugh tracks.

i dunno how they do the animations. Haven't watched enough to see if it repeats, or it is generated dynamically.

The script is probably written by just asking chatGPT (for a script). Might even have some prompts that seed the script.

Then the output of the chatGPT is fed into a TTS (text to speech) and played.

I think the laugh track is randomly played (or played in between different scripts that GPT generates). And i can't tell if the music is generated, or just repeats and is only for some scenes.

Quite genius actually.

For the animations, they could probably write it with traditional artificial intelligence. It's as simple as having events that trigger based on whos talking, such as sit on couch, walk over to kitchen, wave hands while talking, etc.

Right, it might be possible to have chatGPT produce a set of movement scripts that run as part of the text of the show.

Tho i would imagine trying to get that working is a lot more work than just randomly choosing a set of pre-animated videos. I suppose if the video doesn't need to sync with the text, then you can create a large bank of them, and choose based on keywords from the text.

Scenes and animations seem to be picked at random. Just watched George microwave something in the kitchen while having a conversation about something totally different. They really nailed it. Seinfeld is truly a show about nothing. Thinking back to countless scenes of Kramer wandering into the kitchen to pour a bowl of cereal and spout some nonsense that would loop back on itself in the last scene of the episode in some tangential way

Reply on news.ycombinator.com