The ASI-is-imminent argument has a problem: The Smuggled Oracle.
Building on Ben Mathes's
Smuggled Infinity frame.
Everyone assumes that because we've gotten to recursive self improvement for coding that general ASI takeoff is also close.
But coding is a verifiable domain.
Our current training approaches are excellent at absorbing the hidden patterns in any oracle.
Over time they can get asymptotically close to fully capturing the oracle's judgement.
Even before they do that, they can often interpolate valid states that weren't known to the oracle but also work.
The oracle has to improve faster than the models to still give them gradients to climb.
Self-ratcheting oracles are cheap in formal domains like coding or formal math proofs.
But they are impossible in human-in-the-loop domains or ones that require a physical loop.
The "ASI is imminent" argument relies on a smuggled oracle as a skyhook to pull all of the models up.