[notahacker]

It wouldn't be HN if it wasn't going off on a tangent...

As for what you wanted to communicate and nobody else is engaging with directly at the moment, I agree there's a kind of Moravec's Paradox realignment going on where it turns out the guy that tiles bathrooms is pretty hard to replace but that giving the carefully-formatted impression you understood what $academic is on about is a simple word substitution exercise that maybe doesn't say that much about about generalised learning skill.

But nobody hires students to continue to be undergrads, and I think middle management should be the least worried of the lot. They still get to do actual Powerpoint presentations to make the unquantifiable bits of their job look quantifiable and explain whose fault x is, their true function is still to be a human that can do the manipulation and that upper management can reward or blame as suits them, and ChatGPT guilelessly disregarding the big boss instructions to satisfy amused end users is a pretty good indication that even basic functioning as a middle manager is nearly as hard as tiling!

[digitailor]

I kind of do think some people do get hired to continue to be like undergrads, and ChatGPT is turning into a pretty good undergrad. I really don’t know what the progression is going to be, but it seems like a widening in the middle of Moravec’s Pdx or something. Algorithmic management is next on the block and tools like GPT will be (and are) involved: [edit:algo stuff] took over a lot of content-making decisions in media concerns years ago, for example.

The results of that aren’t nearly as straightforward as was being portrayed (and so much capital injection was involved too) but what if models trained on known employee behavior really can understand the incentives that would work for individual employees at a finer grain than your typical middle manager? With all the data gleaned from the employee’s work computer etc? And blaming the algo has already become a national pastime!!

It could get weird once trained models start to emulate the behavioral and suggestion parts of communication, and soon. But we tend to want to minimize the behavioral aspect in favor of the raw computation aspect, despite the fact that generative models are creating content based on the behavior they learned from a training process, which is a behavioral training process, distinct from an imperative instruction writing process.

I think a lot of it comes down to that on this whole TFA commentary. People haven’t totally adjusted to the fact that there is a material difference between trained generative models that produce and written imperative sequences that compute. What the difference is and implications isn’t exactly clear, but certainty is not really on the table anytime soon