[dodu_]

>master the tools

Except the entire value proposition of these tools is that there is no skill or mastery to be built.

The entire slop factory workflow, or sorry I mean "AI-native" workflow is:

"Woah, I cajoled a chatbot into building something I don't understand at all, I'm so good at my job!"

It's the participation trophy of building. Something else builds it, I take credit for it despite not understanding much about about it. There's no compounding return on my effort. No lessons learned. No understanding built. No insights gleaned for possible future innovation. No differentiation. Just mind-numbingly screaming into a void until the slot machine shits out some slop amalgam that seems "good enough", and then I do it all again the next day.

If that's the game, count me out. It's nice that others apparently enjoy it, I guess. But to think there's any sort of mastery here is delusion. The only requirement to be "successful" with these tools is to stop giving a shit and surrender to it.

[jeremyloy_wt]

I think there will always be some sort of mastery element to communicating with AI. You can see that nowadays with juniors who blow through tokens and still don’t produce good results. There will always be some group of people using AI better than others. Maybe the results will be equal in the far future, but some other dimension like cost may be better used by some people

[enraged_camel]

>> Something else builds it, I take credit for it despite not understanding much about about it.

How much you understand what was built is entirely up to you. Literally nothing is stopping you, or anyone else, from having the AI walk you through it, or reading the code yourself if you don’t trust the AI.

[dodu_]

> Literally nothing is stopping you

What about the threat of unemployment due to not meeting AI usage/output metrics? I've personally found it has effectively coerced me to stop trying to understand pretty much anything, and instead just send out whatever passes basic test to "keep up".

Unless you want to just play bad-faith word games and say that "technically it's still not stopping you" in which case yeah man you got me good job buddy.

[teravor]

these tools cannot read your mind. every time you prompt it you condition what it will give you on what you gave it. there will either always be a limit to how good that will be or none of this will matter because no human will ever matter again.

so far the skill is to condition it to give you the best results.

there used to be a time where you had to hack together silly idiosyncratic prompts to get the model to do what you wanted. now you just go into the engineering and describe the object you want it to conjure for you in as much detail as possible (including the high level description of the internals if able) and any constraints you want on it.

[CamperBob2]

"Woah, I cajoled a chatbot into building something I don't understand at all, I'm so good at my job!"

That ship sailed when people started using compilers and stopped learning assembly language.

[dodu_]

Never really understood this comparison, as it always felt intentionally obtuse to me, but thanks for replying, friend!

[CamperBob2]

Your confusion probably comes down to a well-intentioned but now-obsolete focus on determinism.

[dodu_]

Why is it a mistake to value determinism, or in your words "focus on" it?

Why is this alternative better?

[CamperBob2]

Because your competitors are using the alternative methodology, however scary it may be, to beat you.

[californical]

Are there really companies losing right now for using less AI?

Think - would you rather your telecom company’s customer support be AI-forward? Would you pay an extra $5 per month to ensure that you get humans solving your problems immediately when you call with an issue?

What about your backup software? Would you rather choose the company that comes out with new innovations in backing up your data and tons of features, but occasionally breaks everything? Or would you want to choose a company for backup software that is slow at adding anything new and reliable? Isn’t it good if this is deterministic?

What about even a fitness tracking watch. Are there really that many missing features that need to be released way faster? Or is it better if it just tracks your heart rate and workouts well and then gets out of your way? Same here, don’t you want the features to be reliable and deterministic?

[dinkumthinkum]

So, this is a complete nonsense take. Why do people keep making this kind of comparison? Is there really such a lack of critical thinking being taught these days?

[CamperBob2]

What you were taught no longer matters, compared to what some of the rest of have learned over the past couple of years. Sorry. Shoot the messenger if it makes you feel better, but it won't change anything.

[tavavex]

This comment has no substance to it. It's the same as posting "I am right and you are wrong. 'Sorry' if me being right makes you feel bad, but you are wrong."

People are asking legitimate questions whenever this is brought up, because the comparison is wrong. Compilers are an optimization of the previous paradigm, they let you do the exact same thing as what was done in the preceding decades, just faster. LLMs are not that, they are just a completely separate thing that exists alongside regular programming. Arguing that their use isn't just another option, but a superior and total replacement doesn't explain why you think that randomness is a perfect substitute for determinism. The only people I see promoting this view are ones who only want to look at graphs where lines are going up without caring about anything else. Because I think that even if LLMs get 10 times better, we'll still have industries and use cases where determinism will dominate, ones where you get one take to get things right. For instance, I would not fly on a plane with firmware that was created by a guy hitting Generate, going for a coffee break, coming back and saying "yea looks good".