原推：I don’t think it’s possible to pre-predict the performance upper-bound, given capability is emergent and discontinuous.
But declining cost to train models — a prospective halving every 9 months through 2030 — will allow teams to more aggressively search the capability-space.
