For predicting feasible scaling investment, drop-in replacement for a significant portion of remote work that currently can only be done by humans seems important (some of which is not actually done by humans remotely). That is, an AI that can be cheaply and easily on-boarded for very small volume custom positions with minimal friction, possibly by some kind of AI on-boarding human professional. But not for any sort of rocket science or 90th percentile.
(That’s the sort of thing I worry about GPT-5 with some scaffolding turning out to be, making $50 billion training runs feasible without relying on faith in heretofore-unseen further scaling.)
For predicting feasible scaling investment, drop-in replacement for a significant portion of remote work that currently can only be done by humans seems important (some of which is not actually done by humans remotely). That is, an AI that can be cheaply and easily on-boarded for very small volume custom positions with minimal friction, possibly by some kind of AI on-boarding human professional. But not for any sort of rocket science or 90th percentile.
(That’s the sort of thing I worry about GPT-5 with some scaffolding turning out to be, making $50 billion training runs feasible without relying on faith in heretofore-unseen further scaling.)