I think this is, essentially, a wishful take. The biggest barrier to models being able to do more advanced knowledge work is creating appropriately annotated training data, followed by a few specific technical improvements the labs are working on. Models have already nearly maxed out "work on a well-defined puzzle that can be feasibly solved in a few hours" -- stunning! -- and now labs will turn to expanding other dimensions.

There are plenty of ways of writing more capable software stacks using LLMs which don’t rely on reinforcement learning. If anything, the AI labs have too much of a focus on building larger models with bigger or better sets of labeled data, where algorithmic changes will let you do more with the same tools.