[jncfhnb]

> Generative AI is an instantaneous push-button solution. It generates a legal brief or construction plan from scratch. Interacting with software is like chatting to a friend. You don’t have to re-learn your whole process – you simply remove tedious tasks from your to-do list.

We’re getting sued for the homes we built that have unreachable floors and outdoor facing windows placed inscrutably on load bearing interior walls but I feel really good about our legal defense

[Workaccount2]

The core tool being offered here is allowing experts to focus more on checking work than solving work. It's way faster to check if a Sudoku puzzle is correct than it is to solve the Sudoku puzzle.

[nzach]

But this doesn't apply to all domains equally.

Debbuging, for example, is harder than programming. And the difficulty is related to the 'amount of code' you are debbuging.

This means it is pretty easy to verify the Copilot's output when he just spits out a couple of lines at a time. But you will have a pretty bad time when you ask ChatGPT for a complete service. Things will not work out and the time you take to fix it will be greater than what you saved in the first place.

[RhodesianHunter]

Absolutely, which is why you have the LLM generate one method for you at a time.

This still increases your productivity dramatically, and shift you into the role of editor rather than writer.

[jncfhnb]

That depends entirely on how quickly you can write the method on your own.

[tivert]

> The core tool being offered here is allowing experts to focus more on checking work than solving work.

That sounds like a nightmare, not an improvement. Really, think about it. It's the same shit as "self driving" cars that need continuous human monitoring. You're taking an relatively engaging task and replacing it with a mind-numbing slog that humans are particularly bad at.

Not to mention the skill of being able to "check work" usually flows from deep experience of "doing work."

> It's way faster to check if a Sudoku puzzle is correct than it is to solve the Sudoku puzzle.

Yeah, and which of those tasks do humans choose to do? I don't see many "100 Solved Sudoku Puzzles To Check" books in the bookstore.

[RhodesianHunter]

/Not to mention the skill of being able to "check work" usually flows from deep experience of "doing work."

Precisely, but you need far fewer of these people, which is why this is being so heavily pushed.

[tivert]

>> Not to mention the skill of being able to "check work" usually flows from deep experience of "doing work."

> Precisely, but you need far fewer of these people, which is why this is being so heavily pushed.

That sounds like extremely specious reasoning to me, probably due to working backward from technology to application in order to hype the former.

Firstly, is it's anyone experience that it's easier to understand an unreliable system that was barfed out of some unreliable process (doesn't have to be an LLM, could be a bad offshore team), than it is to try to build it right from the start? It's still garbage out. It's like abusing the QA process by saying quality is only their job, then carelessly pumping out crap work and expecting them to catch all the mistakes.

Secondly, where are these "fewer" skilled people supposed to come from? The technology, if embraced this way, will have the effect of cutting off the the skils pipeline. That would work in the short/medium term, but in a generation when you start to see lots of retirements, you'll hit a skills dead end.

[RhodesianHunter]

>The technology, if embraced this way, will have the effect of cutting off the the skils pipeline. That would work in the short/medium term, but in a generation when you start to see lots of retirements, you'll hit a skills dead end.

When have corporations ever cared about the next generation, let alone anything beyond the current quarter?

[jncfhnb]

The sudoku puzzle analogy has a single correct answer and many easily assessed errors.

That does not map to instantiating a construction plan from scratch.

[Workaccount2]

But it's literally how construction projects work. You have a bunch of junior engineers who hammer out the work, and then a master PE who reviews and signs off on it (who also takes full legal responsibility for all the work).

The master engineer isn't the one who does all the hours and hours of nitty gritty design work. She just does review and makes adjustments as needed.

[jncfhnb]

Yes but the junior engineers aren’t the dumbest morons you’ve ever seen. This isn’t a realistic depiction of how the tools would get used anyway. It’s not “generate a plan from scratch” and submit for review. It’s “carefully iterate small parts over and over”.

The junior engineers would be the ones reviewing the outputs. The senior engineer would be the one reviewing what is still constructively the junior engineer’s work.

[beeboobaa3]

It's also way easier, and much less responsibility, to miss a mistake than to make one.

"Oh, yeah, that was a fuckup but the AI did it, not me! What, I was supposed to catch it? Well, I blame the AI!"

[notnaut]

Insurance exists. Money points to something in reality (usually).

Negligence and duty don’t change and the implementing human (the person checking for mistakes) will be just as liable as the human implementing someone else’s work today.

But true… it is unlikely the AI firms or the suits that force it into every nook and cranny they can will ever be held accountable for the mistakes it will inevitably make. Not without a couple catastrophes first.

[ben_w]

I was thinking that, but then I remembered how often humans make mistakes and and don't check their own work. For me, a common example is loose/lose, which I only find in my writing by assistance from text-to-speech. Did you spot the deliberate mistake in this comment? Because that too is one I miss, when the two words are separated by a line-break.

[Workaccount2]

Engineers already sign off on work done by others all the time. This isn't some new or radical concept. For civil engineering projects, you can face jail time if you signed off on bad work regardless of who did it.

[htrp]

> We’re getting sued for the homes we built that have unreachable floors and outdoor facing windows placed inscrutably on load bearing interior walls but I feel really good about our legal defense

This seems oddly specific, has this already happened in real life?

[nicklecompte]

Not in the sense of an architect actually building a generative AI blueprint. But transformer ANNs are simply incapable of designing physically plausible construction plans. Theoretically, not being able to solve graph connectivity seems like a big issue. Empirically, not being able to count is a fatal flaw.

[drbig]

> Theoretically, not being able to solve graph connectivity seems like a big issue. Empirically, not being able to count is a fatal flaw.

Thank you for capturing the state so eloquently. I've asked Bard about how to dilute hydrogen peroxide from 10% (do not touch) to 3% (medical) while producing a specific quantity of the dilute... Results were harmful.

[langsoul-com]

This is happening a lot in Australia, very dodgy builders resulting in illegal (from the building code) houses. Some are unsafe, others are failing apart way quicker due to the shortcuts taken.

[jncfhnb]

There have been some gen ai architecture design tools posted here with such flaws. They were not so bold as to call them construction plans.

[bende511]

GroverhausAI, coming to a neighborhood near you soon!

[RhodesianHunter]

I read it as obviously and entirely tongue-in-cheek.