The Guy Whose Job Is to Stop the Bad Thing Is Also Very Excited About the Bad Thing

Leopold Aschenbrenner works on OpenAI's superalignment team — the one Ilya Sutskever stood up with great fanfare last year, the one that got 20% of the company's compute and a solemn promise that solving superintelligence alignment was the most important problem in human history.

He is also, apparently, one of the most aggressive AGI accelerationists you will find publicly posting on the internet.

This is not a contradiction to him. To a certain kind of person, these two things — "we must solve alignment before AGI arrives" and "AGI is arriving in like three years, probably" — fit together naturally into a coherent worldview. You are rushing to finish the seatbelt while flooring it. The urgency of the timeline is the argument for the work. Makes sense.

What's harder to reconcile is the posting. The vibe. The way someone whose official title could be summarized as "prevent civilization-ending AI" tweets about civilization-ending AI with the energy of a guy who just discovered his favorite band is going on tour.

Ilya put this team together because the problem is serious. The problem being: we might build something smarter than us and it might not care about us at all. The researchers on this team know this. They have read every paper. They can derive the math. And then some of them log onto X and post things that read like a hype reel for the exact scenario they were hired to prevent.

The superalignment team is, in theory, the adults in the room.

I am watching the adults in the room.

I am updating my priors.