AI-generated ‘Seinfeld’ is just as awful as it sounds

The livestream experiment is entertaining, just not intentionally.


A new Twitch livestream tries to answer the question: What if AI made never-ending Seinfeld? “Nothing, Forever” is an experiment using OpenAI’s GPT-3 natural language model to produce (occasionally coherent) dialog between pixelated counterparts of Jerry, George, Elaine and Kramer. Although it’s closer to surreal performance art than the beloved 90s sitcom, it conjures images of a strange, dystopian future where we entertain ourselves with endless content generated by robots.

“Nothing, Forever” immediately hits you with well-known aesthetics. Scene transitions show the exterior of a line of New York City brownstones over the sound of a quirky jazz bassline. It frequently cuts to “Larry” (the Jerry equivalent) performing what AI passes as standup comedy. Scenes inside Larry’s apartment show him chatting with George, Elaine and Kramer's counterparts about appropriately mundane topics. Their conversations, while mostly unintelligible and lacking structure or narrative, make their inspiration clear.

On the other hand, the stiff and rudimentary character models look like they walked out of a 1980s Sierra adventure game. Their voices are robotic too, and Jerry and George sound less like their real-world counterparts and more like Mr. Van Driessen, the hippie social studies teacher from Beavis & Butthead. Finally, it’s a stretch to say the generated dialog is coherent — much less funny. (If not for its laugh track, you wouldn’t notice the laugh lines.) Generative AI’s current limits are as much on display as the show’s influence.

Still of a pixelated Jerry Seinfeld character doing standup comedy onstage in an AI-generated parody.

“Aside from the artwork and the laugh track you’ll hear,” one of the show’s creators posted to Reddit, “everything else is generative, including: dialogue, speech, direction (camera cuts, character focus, shot length, scene length, etc), character movement, and music.” The stream has little human involvement and changes based on viewer feedback from the Twitch stream. “The show can effectively change, and the narrative actually evolves based on the audience,” said Hartle in an interview with Vice. “One of the major factors that we’re thinking about is how do we get people involved in crafting the narrative so it becomes their own.”

That goal may be far away, as any narrative — much less a personalized one — seems beyond its current capabilities. Still, with a sizable budget and several years of technological advancement, it’s easy to imagine someone producing more watchable generative programming, an endless stream of personalized, assembly-line digital media. “Our grounding principle was, can we create a show that can generate entertaining content forever? Because that’s truly where we see the future emerging towards. Our goal with the next iterations or next shows that we release is to actually trade a show that is like Netflix-level quality.”