A 24/7 Twitch Stream of AI-Generated Celebrities Is Even Weirder Than It Seems

A 24/7 live stream featuring AI-generation versions of real streamers, internet personalities, and comedians is a mesmerizing look at the potentially dystopian future of live-streaming. But that’s not the weirdest thing about it: the stream is the product of volunteers working commune-style for an organization called The Singularity Group, which was founded by a former pro gamer who has been described as founding a “pseudo-religion.” 

Bachir Boumaaza, the former-pro-gamer-turned-Twitch streamer in question, goes by Athene online and his face and voice serve as an AI-generated host on stream, asking questions of similarly deepfaked guests on the Athene AI Heroes channel. Sometimes Boumaaza streams as himself live, while talking to generated versions of guests on his main channel, including AI parody versions of Ricky Gervais and Bill Burr. The Athene AI Heroes channel, which shows a deepfaked Boumaaza picking questions for AI-generated guests from a live (human-populated) chat, has been streaming non-stop for days. It features dozens of guests, including virtual simulations of Joe Rogan, Amourath, Alinity, PewDiePie and Hasan Piker.

Reese Leysen, a spokesperson for Boumaaza’s philanthropic venture Gaming for Good, told Motherboard that this project started on February 5 with Boumaaza’s main Twitch and YouTube streams, and the 24/7 fully automated stream began a few days ago. 

The stream is run by volunteers from The Singularity Group, according to its description. The group—which lives together in a compound in Germany—exists to further the ideals of a philosophy Boumaaza created in 2011 with a YouTube documentary laying out its ideas called “Athene’s Theory of Everything.” The philosophy can be called “neuro-Spinozism”, or, as it was initially, Athenism. “Athene’s Theory of Everything” ends with a link to Athenism.com, which now redirects to The Singularity Group’s website. A YouTube video posted by Boumaaza aimed at attracting more volunteers to the compound shows him describing The Singularity Group as a “movement.”

“With the incredible pace at which AI technology is evolving, we started to realize at some point last year that 'AI influencers' would soon become a real thing,” Gaming for Good’s Leysen said, “and we started working on our own in-house tech to see what we could come up with that would make for great entertainment both for the Athene Twitch & YT channels as well as for our Athene AI Heroes game.” 

The Athene AI Heroes game, which is heavily advertised on the AI-generated channel, is a gacha-style crypto game that’s part of Boumaaza’s charity funding for the philanthropy startups he’s founded. Attempting to download the game takes the user to the App Store page for Mobile Minigames, which is developed by The Singularity Group under the auspices of a firm called The Naughty Cult. Leysen is also director of that firm, according to his LinkedIn. 

One of the Singularity Group’s goals is to  pave the way to a Universal Basic Income through crypto-token-based play to earn mobile games. In 2019, Boumaaza and The Singularity Group were accused of “manipulation, misogyny, emotional abuse and a lack of accountability at the top of the organisation,” according to Kotaku, which referred to the movement as a “pseudo-religion.”  

In 2021, Boumaaza posted a self-promotional documentary on YouTube that told the story of his rise to fame and the formation of The Singularity Group, framing it as being merely philanthropic and denying allegations of being a cult or a scam. A crypto game developed by the team—Clash of Streamers, which now also redirects to Mobile Minigames—is described as a “shameless cash grab” and “trash” to fund philanthropy. 

Leysen said that no part of the stream is pre-recorded, and because no one else that they’re aware of is doing anything like this, many viewers were skeptical that it was really AI-generated, and assumed that some of it was scripted.

“The entire setup behind it is something that took quite some development time and combines many different AI technologies as well as a lot of creative tweaking of the machine learning process to make sure the result is something that is an entertaining parody rather than just an imitation,” Leysen said. The answers are generated using a combination of GPT-3 and their own datasets, they said, and the deepfake videos are generated using a system that they built in conjunction with open-source components. “For now we don't want to give away too many details yet regarding the secret sauce that is allowing us to generate the responses on-the-fly,” Leysen said.  

The voices are sometimes glitchy and the video is choppy, but the AI characters’ responses do sound convincingly like the real humans they’re parodying: AI Chapelle frequently starts its answers with “Let me tell you something folks,” and AI Rogan goes on about elk being a great example of gene-edited deer and asks Jamie to pull up a video of a bear doing jiu-jitsu. AI Jordan Peterson (who, in real life, has gotten very mad about AI generated versions of himself in the past) works the phrase “woke moralists” into almost every reply it gives.

In one of his own streams, Boumaaza says that the questions chat asks the AI guests are training them to give the answers they’re giving. 

Leysen told Motherboard that during the streams, moderators stand by to filter out questions in the chat that might lead to offensive or less entertaining answers from the AI guests. 

“During the 24/7 stream we combine very heavy multi-layered AI moderation with human moderators who also try to keep an eye on things so that we minimize the chances of the channel getting in trouble,” they said. “We also very much go out of our way to fine-tune the way the models are trained to make sure the general tone is light-hearted and we also moderate with the streamers in mind. This means questions that we think the streamers might be uncomfortable with get rejected.”

In the Discord, admin have created spaces for people to help train the AI guests with answers to prompts, like “opinions and beliefs” and “jokes and catchphrases” for each personality. “To make an AI streamer say believable sentences, one of the things we need is personal information about the streamer. You can imagine you are explaining the streamer to one of your friends that doesn’t know the streamer,” the admin wrote, and ask viewers to fill in details about each person’s life. 

For example, under Elon Musk’s history prompt, someone wrote, “He is addicted to tweeting and uses twitter when he is on the toilet.” These answers from viewers, in theory, will be input into the ChatGPT3 model that the Athene AI Heroes stream is using, to teach it what Musk might typically say to a given question. There are also sections for people to submit video clips of each person to help train the AI audio and deepfake visuals. 

AI generated content has quickly become the genre du jour for experimental Twitch streamers, as well as popular personalities who have spoken out against reckless uses of AI. In January, Twitch suspended AI-generated Seinfeld parody “Nothing Forever” for telling a transphobic joke, and in early February, streamer Brandon “Atrioc” Ewing confessed to watching deepfake porn of his colleagues, prompting many female streamers to speak out against image-based abuse. 

This AI-generated mega-show, including parodies of some of the most naturally annoying online personalities and streamers out there, is hard to look away from. The faces aren’t very realistic and the audio is frequently glitchy, but it’s easy to watch for question after question just to see what absurd catchphrase or disturbing glitch they’ll come up with next. At one point, the chat asks the AI version of streamer Sniperwolf to sing a wolf howl, and it obliges, then proceeds to take several long seconds to moan like a demon. “That’s probably the closest I can get to singing,” the AI Sniperwolf says. “You like?”