Inside Ai Safety’s Hottest Club

BERKELEY, California — If the world was going to hash out the biggest policy dilemmas surrounding AI, a converted inn in Berkeley, California seemed as fitting a place as any.

Decoded by POLITICO got an inside look at The Curve, the invite-only conference last weekend that hosted a rare convergence of AI influence and offered a microcosm of the state of current global safety talks. It’s the kind of gathering where Virginia Democratic Rep. Don Beyer asks famed AI pioneer Yoshua Bengio about open-weight models, while attendees muse in the bathroom line on the odds of California Gov. Gavin Newsom calling a special session for AI. (Hint: they are low.)

Other industry luminaries including OpenAI co-founder Wojciech Zaremba, Anthropic co-founder Jack Clark and Google DeepMind’s head of policy Owen Larter roamed the summer-camp-like campus with representatives of both red and blue state governments, Congress, foreign leaders, political groups, civil society, major donors and federal agencies.

Yet for all the expertise assembled, the three-day retreat broke up with the regulatory path forward still decidedly murky, between Congress being on recess and midterm elections approaching that could make the politics around AI even more factional. It highlights the challenges ahead for AI safety advocates now that OpenAI’s Hugging Face hack and other incidents have finally given them the broader public and policymakers’ attention.

“We just need to be uber-pragmatic and action-oriented, and I don’t think we have time for a lot of screwing around,” one AI industry employee told Christine. “That is, come in and say, ‘what the hell's the plan?’”

The event captured a growing, politically potent movement as it confronts a strange irony in its own mainstream ascendance. The future risks that AI “doomers” and safetyists long warned about sound grimmer than ever, but the campaign and policy worlds are bending their way and giving them leverage to do something about it sooner than expected.

“There's a new layer of pressure, like a kind of pall that’s cast over the whole thing by the outside world having gone completely insane over AI,” said Joshua Achiam, OpenAI’s former chief futurist. “It's the subject of every major political debate, and no one feels particularly good about the trajectory that it's on in terms of beneficial impacts and risks.”

Case in point: New political organizations and PACs keep continually forming to help elevate AI onto the agenda for 2028 candidates. Organizers with more established political groups also checked out the Curve this year, intending to sharpen their focus on the issue.

The evolution is notable compared to 2025, when the event felt more like AI safety advocates needing to court politicians to their cause. Safety advocates, saddled with a reputation in policy corners as an eccentric, technical subculture, were trying to find allies with populists on the right, China hawks, Hill staffers and other creatures of official Washington — some of whom showed up in Berkeley in full suits and ties.

Now, safety groups have multiplied and gained political weight, arriving with better staffing, a record of state regulatory wins, viral researchers’ warnings and AI CEOs themselves warming to tougher regulations.

Several attendees argued industry self-regulation is insufficient and government needs to play some oversight role. They contested the exact division of responsibility with others, as well as an intermediary option, like using third-party monitoring organizations, which has been catching on with lawmakers.

Other than frontier AI safety, there are “enormous policy questions open” around cybersecurity, AI’s economic impact and privacy issues, said Taren Stinebrickner-Kauffman, a Curve host. “I hear policy leaders worried about that a lot. It’s not just that there's no consensus, but there's not even that much fleshed out on the bone of the policy ideas either.”

Beyer blamed Congress’ inaction on Republican House leaders, saying he was “confident that will change, and that the next speaker, the next majority leader, the next vice chair of the Democratic Caucus will be very committed to putting something on the ground.”

Much of the conference unfolded under Chatham House Rule, protecting speakers’ identities, and through impromptu chats like breakout groups in the attic and gossiping around fire pits to encourage candor between unlikely bedfellows or usual adversaries.

And not everyone was as optimistic as Beyer. “There is no legislature in the country, and there certainly is not a Congress that is at this moment equipped with the fundamental understanding about how to create a durable legislative framework that will last any more than a couple of months at a time,” said one attendee who works in government.

A second government official remarked how changing that seemed out of reach: “I know this conference is not a random sample of humanity, but I feel like the consensus here is we don’t even have the time for that.”

Enough time before what? Attendees were also divided on this question. Recursive self-improvement — AI’s ability to autonomously upgrade itself, potentially leading to a loss of human control — was a frequent hot topic. One AI engineer predicted a future with a Hugging Face-like incident “once every week.”

Several people expressed that an even larger AI security incident could happen within the year or the next one and may force President Donald Trump to rethink his AI positions.

For the first time, the Curve plans to host an event in D.C. next year. And the changing relationship between the AI labs and government is a dynamic to keep watching.

“The labs would just say stuff last year that was the stuff you would say to a crowd that already agrees with you,” Neil Chilson, the Abundance Institute’s head of AI policy and an AI accelerationist, said after his Curve experience. “Their messaging was much more disciplined this year.”

That’s not to say the cultural differences have disappeared. At one point, a political operative turned to a friend to debrief a striking claim: that machines may represent the next stage of human evolution. And Beyer and Democratic Rep. Sara Jacobs of California were asked about their p(doom) — the probability they assign to an AI-driven doomsday — but neither was willing to say.

A version of this story first appeared in Decoded by POLITICO, the must-read briefing on the people, power moves, policy battles and market shifts defining what’s next in global technology. Sign up to subscribe to the newsletter and listen to the podcast.