I’m not sure I agree, but one moderate piece of evidence for this view is that the distribution of skill profiles for current-gen AIs seems to be sharper in the “bad at hard-to-verify tasks compared to tasks with easily-verifiable tasks with good feedback loops,” compared to humans. So you might expect future AIs under this current trajectory to share this relative skill distribution, at least in the near future.
Optimistically we (with or without near-future AI assistance) may in the future discover other architectures and training processes that have the opposite distribution, at least within remote-only tasks: more like a Descartes-bot that’s unusually good at legibly reasoning and making progress with informal arguments, compared to going hard on empirical metrics like programming or persuading humans.
I’m not sure I agree, but one moderate piece of evidence for this view is that the distribution of skill profiles for current-gen AIs seems to be sharper in the “bad at hard-to-verify tasks compared to tasks with easily-verifiable tasks with good feedback loops,” compared to humans. So you might expect future AIs under this current trajectory to share this relative skill distribution, at least in the near future.
Optimistically we (with or without near-future AI assistance) may in the future discover other architectures and training processes that have the opposite distribution, at least within remote-only tasks: more like a Descartes-bot that’s unusually good at legibly reasoning and making progress with informal arguments, compared to going hard on empirical metrics like programming or persuading humans.