On Friday evening the Wall Street Journal published a feature on the network most of the AI safety scene comes out of. The place is called Constellation, and it sits in Berkeley’s tallest office building. Vegan catering, couches with sweeping views of San Francisco Bay, monthly dinners and happy hours where the newest and most uncomfortable AI risks get worked through. People who work at Anthropic have their own office space there, according to people close to it.
The line that stands out comes later in the piece. Over the past few weeks, some of Anthropic’s earliest employees told an industry colleague they’re considering buying land in remote parts of the US. Somewhere to relocate to, if AI goes awry.
The idea isn’t new in this crowd
Former employees remember company dinners in Anthropic’s early days where a Manhattan Project scenario came up: they might be asked to move to the desert and keep building AI inside an electromagnetically shielded government facility. Back then that was a thought experiment over food. Now it’s property deeds.
The feature traces how tightly the early staffs of Anthropic and OpenAI are woven into effective altruism. Part of it comes from a Bay Area network that has spent years gaming out end-of-the-world scenarios. Eliezer Yudkowsky warned about uncontrolled AI in the mid-2000s and became the intellectual godfather of the subculture. Before AI reached the top of the list, the group obsessed over asteroid impacts and supervolcanoes.
Why this feature lands now
It doesn’t stand alone. Three days ago a memo went round the White House casting Amodei as the face of AI doom, and on Friday an Anthropic investor called the warnings a policy instrument. The effective-altruism roots of Anthropic’s leadership have gone from footnote to political argument, and it’s an argument being used against the company.
The Journal supplies the imagery for it. Vegan catering and a bay view in a co-working space where people discuss the end of humanity sticks in the mind more easily than any debate about compute thresholds.
Buying land is an honest statement
You can find the property hunt ridiculous. You can also read it as the most honest thing to come out of the industry this week. People who know the models from the inside are arranging their private lives around what they say in public. That’s more follow-through than most of the recent debate contributions offer.
What it doesn’t settle is whether the assessment is right. A plot in Montana is not evidence of a risk, it’s evidence of a conviction. Anyone who thinks Amodei’s slowdown call is tactical will read this and say the scene really does believe its own story. Anyone who thinks the call is warranted will say: exactly.