Situational Awareness
My friend Leopold Aschenbrenner, who I got to know and respect on OpenAI’s now-disbanded Superalignment team before he left the company under disputed circumstances, just released “Situational Awareness,” one of the most extraordinary documents I’ve ever read. With unusual clarity, concreteness, and seriousness, and with a noticeably different style than the LessWrongers with whom he shares some key beliefs, Leopold sets out his vision of how AI is going to transform civilization over the next 5-10 years. He makes a case that, even after ChatGPT and all that followed it, the world still hasn’t come close to “pricing in” what’s about to hit it. We’re still treating this as a business and technology story like personal computing or the Internet, rather than (also) a national security story like the birth of nuclear weapons, except more so. And we’re still indexing on LLMs’ current capabilities (“fine, so they can pass physics exams, but they still can’t do original physics research“), rather than looking at the difference between now and five years ago, and then trying our best to project forward an additional five years.
Leopold makes an impassioned plea for the US to beat China and its other autocratic adversaries in the race to superintelligence, and to start by preventing frontier model weights from being stolen. He argues that the development of frontier AI models will inevitably be nationalized, once governments wake up to the implications, so we might as well start planning for that now. Parting ways from the Yudkowskyans despite their obvious points of agreement, Leopold is much less worried about superintelligence turning us all into paperclips than he is about it doing the bidding of authoritarian regimes, although he does worry about both.
Leopold foresaw the Covid lockdowns, as well as the current AI boom, before most of us did, and apparently made a lot of money as a result. I don’t know how his latest predictions will look from the standpoint of 2030. In any case, though, it’s very hard for me to imagine anyone in the US national security establishment reading Leopold’s document without crapping their pants. Is that enough to convince you to read it?