AI安全网红训练营PDKU：用短视频给AI安全注入趣味

Oliver Kemp for Transformer

“Would you like a bite of cake? It’s for my time lapse.”

A petite woman with metallic silver face paint points to the remains of a days-old wedding cake. A tripod-mounted iPhone stands at the end of the picnic table. Another woman is about to start a 24-hour livestream in the building behind us. Around the firepit, several others squint at laptops, fighting the sun’s glare.

I politely decline the cake and make myself a press badge out of construction paper and glitter glue. Usually at Lighthaven, a converted Berkeley hotel that serves as the default conference center and clubhouse for the rationalist community, women are few and far between. Cool women are even rarer. Today, I’m quite literally surrounded by women who’d fit in with Kylie Jenner. A guy with a septum piercing and green hair walks on his hands across the fake lawn.

Cut through the noise.

At the start of July, nearly 60 content creators — some seasoned, many first-timers — packed their bags and moved to Berkeley for Plz Don’t Kill Us, or PDKU, a month-long bootcamp meant to inject fun and sex appeal to AI safety comms. It’s the brainchild of Aella, Substacker/amateur data scientist/sex worker/orgy-nizer, and Ronny Fernandez, who pitched “anthropomorphic raccoon of Lighthaven” as a fitting job title.

The rules of PDKU are simple. Fellows live at Lighthaven rent-free for the month of July, and have to post one short-form video every day. The video must be at least 15 seconds long, and it can be about anything that isn’t “neglectfully misinformed” or needlessly cruel. Anyone who fails to post something by 11:59 pm gets kicked out of the house.

It might sound like the setup for a reality show (several fellows joked that it ought to be), but it’s partially funded by the Machine Intelligence Research Institute (MIRI), one of the most established organizations in the world of AI existential risk. Throughout the month, fellows were visited by a rotating cast of AI safety experts and internet celebrities including video essayist CJ the X, rationalist luminary and MIRI founder Eliezer Yudkowsky, as well as a video call from synth-pop artist (and Elon Musk’s ex) Grimes.

The PDKU house, based at Lighthaven

PDKU is a bet that if AI safety nerds, zoomer edgelords and fashion girlies hang out for a month-long sleepover, united by shared challenges and friendly competition, they’ll rub off on each other. It’s an experiment in throwing proverbial spaghetti (TikToks) at the wall (the general public) to see what sticks. Maybe, just maybe, a star will be born, and fifteen seconds of fame will shape the public’s AI discourse more than fifteen years of blogging on LessWrong, the prominent forum for reason- and first-principles-obsessed rationalists, ever could.

The theory of change at PDKU is multi-tiered and a bit nebulous. On the surface, it’s about science communication. “Last year, there were a couple of instances where I watched nerds try to communicate, and oh my god, they don’t know how to talk or get the concepts across,” Aella told SF Standard back in April.

PDKU isn’t the first attempt by the AI safety-minded to inform regular people, of course. There’s BlueDot Impact, which runs beginner-friendly courses that help people get into the field. There’s the Tarbell Fellowship, run by Transformer’s publisher, which trains and funds journalists to work at major newsrooms. There’s even the Frame Fellowship, another AI safety content creator bootcamp in the Bay Area. But compare Frame’s website to PDKU’s, and the vibe difference will smack you in the face. The former features polished white-on-black sans serif text, reminiscent of an OpenAI product announcement. The latter has animated rainbow Comic Sans against a backdrop of floating rainbow sparkles.

This aesthetic contrast is by design. CJ the X, a PDKU mentor who’s advised MIRI on its comms strategy, thinks that AI safety advocates are too close to Silicon Valley for comfort. “The same guys that are running this camp, that are like, ‘AI might kill everybody,’ are the same people that use AI more than anybody I know,” they told me. “They’re obsessed with it. They love it.”

The PDKU site

This breeds distrust and resentment of both the AI industry and those who think the industry is destroying humanity. “People look at Mark Zuckerberg and they’re like, I don’t believe anything this guy says, he talks like a robot,” said Wyn, a PDKU fellow. “A lot of the AI safety people are also like that.”

Wyn is not like that. He’s a long-haired theater kid with video editing experience and a “rationalist-adjacent” brother who exposed him to AI safety canon. “People are pretty hungry to hear about what the hell is gonna happen with AI,” he told me. “But all the people who actually know what’s going on with AI are not, you know, goofy and engaging. So this is sort of filling a gap there.”

“People told me when I got there that [the organizers] said Plz Don’t Kill Us happened because of a rant I gave at a MIRI retreat, where I said ‘everyone hates you guys,’” CJ told me. (They added that Fernandez overheard this once and clarified, “it’s about 40% true.”)

PDKU would make more sense if it was just about getting goofy, engaging outsiders to communicate about AI safety better than AI safety people can. But, counterintuitively, the organizers tried very, very hard to convince fellows that they didn’t have to post about AI at all. “They really feel like, ‘Oh my god, this AI safety program wants me to make AI safety content.’ And we’re like, no, no, do silly things!” said keltan, a PDKU organizer and MIRI employee (who prefers his name is written in lowercase). To make authentic content that doesn’t stink of AI safety, the organizers figured, creators need space to be themselves.

“It’s really hard to run a fellowship that’s not an indoctrination camp,” Fernandez said. “If you don’t try really hard to not make an indoctrination camp, you will on accident.” PDKU, he argued, is trying its best to prevent fellows from feeling pressured into posturing as doomers to impress their doomer-y mentors. Because a deeper layer of the program’s mission, organizers told me, is to seed a cohort of creators — ideally, some of whom gain traction — who can build trusted relationships with their own niche audiences.

“What you should really do, especially if you want to be influential in AI safety, is to focus on the things that make you passionately human,” said Tyler Alterman, a fellow who spent the 2010s as a professional effective altruist (more recently, he co-founded Fractal University, an “improvised college” in New York City). I asked him, hypothetically, what might happen if someone like Ariana Grande were to become a PDKU fellow: someone with 363m Instagram followers and absolutely zero preexisting connection to Silicon Valley or AI safety. “It’d be jarring, it would alienate a bunch of her fans … She’d be taking away the thing that we’re fighting for,” he said. Alterman thinks that thing worth fighting for is something like human excellence and creative expression. But, he added, a lot of AI safety-focused effective altruists alienate people “by paving over their subjectivity with a cause. Just being a cause robot, a cause zombie … Which I did to myself.”

The program’s most galaxy-brained goal, though, is to create a celebrity network that can be called upon in future times of need. As anyone who’s tried to organize an event knows all too well, cool people are more likely to go where cool people already are. “If you were like, hey, I can give you a warm intro to Ariana Grande, do you want to hang out with me and Ariana Grande at a bar? I’d be like, absolutely,” Fernandez said. “It’s valuable to us if there’s more AI safety content out there, but it’s also valuable to us if we have connections with large creators who we helped emerge into the world,” keltan said. “If we ever need to call on them, then we can.”