Comment by demibabs

Suppose the proliferation of LLMs unlocks some new kinds of work that models are not good at, and thus it makes more sense for humans to do.

Then what is preventing the owners of AI companies from collecting and training on tons of examples of this new work, until AI are equally good at it as humans?

The issue is that unlike the technology that automated farming or the like, AI is a general technology. So not only could it theoretically automate the work humans are currently doing, they could also automate any future human work, even if there’s some degree of lag.