Rocket

The Metaculus/Manifold questions for AGI aren't very helpful! The most popular questions focus on passing a Turing test, but this is likely to be a lagging indicator. As long as AI capabilities remain "spiky", then an experienced interlocutor can focus in on whatever areas AI is known to be weak at. Passing an adversarial Turing test with an expert judge thus requires an AI which is at human level even in its weakest domains, at which point its other domains will probably be wildly superhuman and it will be closer to ASI than AGI.

The current focus of the frontier labs is on automating coding and R&D, and so the most important questions... (read more)

Making an account because I had also been wondering this! The impression I got is that they seem quite loosely committed to that prediction; in Machines of Loving Grace, Dario says that powerful AI "could arrive as early as 2026" but could also take much longer, which seems a much more defensible position. Later, this seemed to firm up to "most likely around late 2026 or early 2027", but that claim stopped appearing in their posts for a while, and the post from Jack Clark that you linked is the only recent reference to it that I'm aware of.

Given how surprising the claim is, how no-one else can figure out their reasoning... (read more)