>As for the comparison to AI, the problem is, AI *must* be told what to do. It won't magically grow into a "mature developer." That's not a natural progression. It always assumes that the prompt accurately describes what it should do. It has no way to know that the prompt was wrong or incomplete in the first place.

This is wrong. You seem to be unaware that current sycophancy in mainline models is a specific choice made in AI model weights to maximize people returning to the model.

It's highly likely that one of the solutions that will be used in specialist fields where rejection of the input if it's insufficient in some critical way is reduction in pro-sycophancy model weighing. I.e. model will actually have a much greater ability to tell you "I can't do that Dave" and then explain why it can't do it.

Some narrow specialist models already do this through ControlNet style "AI that corrects and guides human AI prompts for optimal outcomes", where it will tell you in case of some of the common prompting errors before passing the input to the worker model.

So you do understand the problem then.

Would it then be fair in your view to reframe the specific problem you have into the two following components?

1. This is the worst AI will ever be at being manageable by people. It will continue to improve until it's better, just like what happened with everything where AI is already better.
2. You can manage AI current gen AI with similar methods you'd need in managing your average "yes saar, of course saar, I'll go do what you say right away saar" Indian developer stereotype.

Notably, once you accept the second one, you quickly realize that you can use ControlNet style methodology of "just use a specialized AI to curate your inputs into your preferred task specific model". And for even better results, you can add model alloying into this specialized AI, so it can utilize the best way to handle the sycophantic worker. "Have a different worker check entirety of his work to see where the failures lie and fix them".

I can tell you never had to do managerial work, as you're unaware that one of the most common stereotypes of a worker. The guy who will say "yes boss" no matter what is asked of him, and you'll find out you asked too much of him only when he fails to do the task correctly and this failure is reported on. Often by someone else.

This is even worse with people that come from Indian culture, where "yes boss" is the expected answer regardless of how impossible the ask is.

What you're describing is fundamentally a managerial skill set.

A lot of software developers struggle with those, and quite a few are borderline incapable of it. That's going to be increasingly a problem, unless we manage to get AI trained on individual preferences, and correcting their responses into proper AI prompts. I.e. narrow model that AI incapable worker can interact with you generate a prompt for the major model that will do the actual work.

Ironically to get your computer to do what you want to, you need to train on how to use it correctly.

Prompting correctly is a part of using it correctly.

Is it worker's fault when manager fails to explain the task correctly?

This is the part that most of the "I can't make AI work" crowd miss. When you give AI instructions, it's like giving instructions to a human worker.

You need to understand its strengths and its weaknesses at least to a reasonable degree, and you need to be careful delineating what task entails, what it doesn't entail, what's a priority and what is of low relevance.

A lot of very good experts at their specific field make for horrible managers because they don't know how to explain aforementioned things. They only know how to do it themselves.

This is the most common point of failure both when leading people and when prompting AI in my experience. Essentially all those leadership skills? They matter a lot now, even for mere subject matter experts, because prompting AI is leadership just as ordering a person to perform a set of tasks with a specific goal is leadership.

Before most subject matter experts didn't need any meaningful leadership skills. They just needed to do the things related to subject they're expert in, and leadership was handled by people with a different skill set (less deep and more broad expertise coupled with at least some leadership skills).

This is a training problem.

Not of AI mind you, but of human doing the prompting. When AI fails this badly, it means prompt has poor instructions.

This is fixed by training people to prompt better. It's similar to teaching people how to be good managers. You need knowledge in subject matter, but your primary expertise should in be identifying, segmenting and communicating each individual part of the task with high precision.

This is not easy, and it's not the same skill set as that of typical coder.

It's hatred horse breeders and horse messengers had for cars and their drivers.

Notably, this is the exact reason why they had to immediately start having exceptions in the law.

In this case they started with farmers, and you'll likely have problems across insurance industry, which will require way more exceptions added on later.

"Safety, security, buy a better rig to run my shit because I'm not optimizing this for peasants" being the clarion call of the bottom feeder tier software development that was massively encouraged over last decade or so is indeed the thing I'm decrying.

It's the IT version of "learn to code".

AI could solve this by bypassing this moat to enable translation to openCL.

Considering just how good AI is at this sort of work once properly trained, I would be surprised if this doesn't happen. Though Nvidia will certainly fight anyone trying to do this to slow it down.

Demand software developers start caring about memory print of their software again. Both in RAM and storage.

Unironically. We've lived out at least a decade and a half of "this software stack is utterly unoptimized garbage" "who cares, just slap bigger system requirements. We're not spending money on optimizing something that doesn't matter to anyone since hardware is advancing so fast".

It's good that every decade or so we get a memory and storage crunch and developers actually have to rediscover things like better compression algorithms and methods, proper garbage collection, and general software optimization.

Seriously, have you seen the size requirements of modern games? Have you seen the retarded chugging of modern office software running win11 on 8GB RAM machines when they have to actually start swapping? Have you experienced the joys of Chrome and all the memes about it being a ramvore?

What in the actual fuck are those tabs doing eating gigabytes of RAM? And why in the fuck are most Chromium based browser installs now almost a gig of storage?

You could do the same things a decade and a half ago on 4 gigs RAM and tiny SSDs that were less than one gigabyte and the system flew and most things except the porn torrents could be stored on it.

And then you consider "ok, what did we actually get for that insane increase in system demands?"

Built in always on spyware. Slightly redesigned UI according to the latest fashion trends. A few arcane additional features barely anyone uses. Games with "that unreal look" that look worse than unreal games a decade ago. And "modern" webpages that essentially ask you one question: "Would you like scrips with those scripts so you can enjoy scripts while you're enjoying scripts".

While reading a text based news article.

Just kidding. They don't ask.

To be fair, quite a few AIs already use reddit as a source of information.

They just don't elevate it to level of "trustworthy/expert/authoritative" in general.

That said, it's always funny when someone shows you "look AI supports my really stupid take" screenshot of AI answer, and AI shows that it used reddit as a source/reference for the answer.

Will this decision be used in court of law as a good argument for increased damages due to "gross incompetence" rather than mere "incompetence"?

They're experts in everything.

"I have a problem with my boyfriend/girlfriend/husband/wife". Expert advice: you need to break up with him immediately.
"How many genders are there?" Expert advice: no one knows. It's probably infinite.
"Are white men uniquely evil and should they be genocided?" Expert advice: Obviously, and you need to check your privilege for merely asking such a stupid question.
And my personal favorite:
"I am 12 and I think I want to fuck my toaster". Expert advice: You should dedicate your entire life and sexuality to exploring ways in which you can fuck a toaster.