How does the show work? “Episodes,” if they can be called that, start like those from the real show, with Jerry at the mic in a comedy club telling jokes, or at least an AI’s approximation of jokes. After the set, scenes play out featuring crude, blockish versions of Seinfeld’s four main characters engaging in sometimes humorous and sometimes incomprehensible conversations in robotic, stilted voices.
There are also occasional breaks where a very 1990s-looking Watch Forever channel, clearly modeled off the TV Guide channel of the era, scrolls a daily schedule of real and invented film and series titles.
Is it popular? At the time of writing, Nothing, Forever has 97,900 followers and an incredibly active chat log. It also offers a subscription option for $4.99/month, but subscriber numbers aren’t shared publicly by Twitch so we don’t know how many people are paying for the content.
What technology is used to keep the stream going? When the show was launched, Hartle’s Reddit post described the collection of programs used to generate Nothing, Forever:
Nothing, Forever is built using a combination of machine learning, generative algorithms (we use ‘generative’ here in a non-academic sense), and cloud services. Our stack is mostly comprised of Python + TensorFlow for our ML models, TypeScript + Azure Functions and Heroku for our backend, and C# + Unity for the client, with some neural voice APIs thrown into the mix.
Hartle also mentioned that there are plans to implement OpenAI’s Davinci models to spice up dialogue and Stable Diffusion for art generation.
Is Nothing, Forever any good? Viral now, the novelty of Nothing, Forever wears off fairly quickly when jokes are repeated and the visual humor of a low-polygon Seinfeld universe fades. It seems unlikely that this show is putting any human-made animated programs at risk of being replaced anytime soon.
