I agree. The most likely scenario is that this is just a "new normal" lift that is percolating through human endeavors and will saturate at some point. For example, the whole cyber-security bruhaha should ultimately resolve into higher standards for code published -- we can now cheaply find and fix a whole slew of minor bugs that weren't worth our time before.

The fact we see a lift is not the same as evidence that the lift is unbounded.

The lift being finite is supported by the fact improvements have come at the edges: improvements from human feedback, improvements in harnesses, improvements on model compatibility with harnesses, improvements in inference efficiency with new architectures, etc. If we were just training better models from scratch that would be one thing, but we are just making better use of a tool we've developed.