They don't really know what hardware they're running on, or what load that hardware is under at the time you ask.
I use these "human time required" estimates a different way: compare against the METR time horizon, and breakdown tasks which are too long and therefore unlikely to succeed.
> Why do models do this?
> Likely the pretrain datasets are at fault
They don't really know what hardware they're running on, or what load that hardware is under at the time you ask.
I use these "human time required" estimates a different way: compare against the METR time horizon, and breakdown tasks which are too long and therefore unlikely to succeed.