> ...the wolves of white collar job automation are closing in for the middle class. You believe that we'll all become cyborg centaurs, while the managers believe we'll all become redundant.
I think he doesn't think it's possible to actually make white collar workers redundant because we don't have AGI, since AGI is precisely defined as that — continually deferred, likely impossible for current technology — end goal that we know we haven't achieved yet.
And FWIW, I think he's right. The fact that LLMs are inherently stochastic, cannot reason or plan sufficiently by themselves, and do not have a world model, means that you will always need humans in the loop not just to oversee, verify, and act as an accountability sink (which by itself could be pretty bad), but also to break problems down, plan, and architecture for them, and, when possible, design automated verifiability systems so that the LLM can act as the core of a cybernetic feedback loop, like a sort of linear genetic programming algorithm (which is what a Ralph loop does, incidentally). This last especially, this act of figuring out how to, either by hand or in a very tight supervised loop with an LLM that relies heavily on human expertise and judgement, specify the desired behaviors in a machine-verifiable way, looks a lot like just a higher level of programming to me. It's just red-green BDD.
> . You think people will care about the sideslop everyone will build, not seeing that 'everyone will build' means 'no one will care'. Worse, means no one will buy (knowledge| skill|creation).
I think this is again assuming AGI, where AI slop sort of reaches and becomes indistinguishable from the designs of people with good taste, architectural knowledge, experience, and care for the craft of actually making good, reliable things — but we're not there yet, and as I said above, I don't know that we'd ever get there. So yes, everyone will be able to make things, but not all of it will be of the same quality even if they're all using AI to do it! See, for example, the kind of thing you get if you put an agent in a Ralph loop to make a terminal emulator, versus what Mitchell Hashimoto is able to do using AI on GhostTTY.