It’s time to take superintelligence seriously.

Two and a half years ago, our team published a report for the State Department. It was the first- ever U.S. government-commissioned investigation of the potential catastrophic risks of frontier AI, commissioned before ChatGPT existed. We spent more than a year on it, interviewing over two hundred people — from frontier lab executives and researchers, to national security officials and technical experts.

For years, we’d been working with AI safety researchers in the big labs and outside, hearing many of the same numbers that the public is now coming to terms with. 10% to 90% estimates of human extinction from AI. Catastrophically dangerous rogue AI by 2030. That sort of thing.

The investigation gave us the chance to show the government what we’d seen. It revealed that frontier AI development posed urgent and growing risks of weaponization and loss of control, that the U.S. government had no framework for handling either, and that some of the people building these systems were privately far more worried than anything they were saying in public. We had to argue with State to get the phrase “extinction-level threat” included in it once — that was considered alarmist at the time. To their credit, they eventually agreed.

Since publishing that report and a follow up investigation on frontier AI security, we’ve done a great deal of work with frontier AI labs, AI infrastructure builders, and U.S. government agencies. Most of that work we can’t talk about in detail, but it largely involved securing critical AI infrastructure against America’s most capable nation state adversaries, in collaboration with members of the U.S. intelligence and special operations community.

Because of what that work involved, it also touched on some aspects of AI containment and AI control. This has given us a fairly unusual perspective about what’s possible to achieve, on what timelines, when it comes to the U.S.-China endgame on AI.

Today, most people think of the U.S.-China endgame through a pretty narrow lens. AI lab researchers talk to other AI lab researchers. Members of the intelligence community talk to other people in the IC, plus the odd think tank or contractor. Most AI hardware security companies don’t regularly talk to cyber experts in the national security agencies. And just about no one talks to the builders or the neoclouds.

But each of these groups see things the others don’t. Often, they’re things that are strategically critical — things which, if they’re not accounted for, could put the U.S. and the world at far greater risk.

We no longer have time for these kinds of information silos. The decisions that will need to be made, now and in the next few months, will need to be made quickly, in the open, with as much information as possible.

That’s why we’ve decided the time has come to share what we know, so that we can have the best possible discussion in public about what to do about the AI endgame. Our hope is that this will encourage others — people who see other important things we don’t — to come forward and do the same.

For years, our approach has been to take the operator perspective: to work with the people at the front lines of developing the models and the infrastructure, of implementing the national security and diplomatic options that we expect to matter in the endgame. We visit datacenters, talk to whistleblowers, attend diplomatic engagements, and do technical research. Our focus is on the gears and wheels of what actually needs to happen, and how.

Over the next few months, we’ll be publishing a series of posts on topics ranging from U.S.-PRC AI agreements as informed by some of America’s top diplomats, to explainers of recursive self-improvement developed with input from current frontier lab researchers, to the future of cyber warfare, to how members of the general public can try to prepare themselves for what’s coming, to how to think about frontier AI datacenter security. Most of what we’re publishing has been circulated in private already, but has never before been made public.

With that out of the way, let’s get started. We’re in the endgame now.