[lsy]

AI coding isn’t an abstraction, though. You can’t treat a prompt like source code because it will give you a different output every time you use it. An abstraction lets you offload cognitive capacity while retaining knowledge of “what you are doing”. With AI coding either you need to carefully review outputs and you aren’t saving any cognitive capacity, or you aren’t looking at the outputs and don’t know what you’re doing, in a very literal sense.

[IceDane]

It's staggering to me how many times I've heard this argument that LLMs are just the next level of abstraction. Some people are even comparing them to compilers.

[SkyBelow]

As much as I use AI, even for coding, I really do not like the argument. They are too chaotic to be compilers. The descent from prompt to code has far too many branches, and even small requests begin to build up bad patterns.

There is some fun to consider when sufficiently advanced AI allows this in areas where we are okay with things going wrong, but that seems a very limited domain for fun and games and not for serious software that needs to be correct as possible.

I can see vibe coding building very simple systems, and it likely will get better with systems that are one off throw aways where edge cases don't matter because we have a one off need of turning input X into output Y, but when it comes to people using AI in systems where correctness matters, long term support must be provided, and ease of adding new functionality is a serious consideration, it seems we are as far from having prompt as code as we are from AGI.

[girvo]

> Some people are even comparing them to compilers.

A lot of people are using them as such too: the amount of people talking about "my fleets of agents working on 4 different projects": they aren't reviewing that output. They say they are, but they aren't, anymore than I review the LLVM IR. It makes me feel like I'm in some fantasy land: I watch Opus 4.7 get things consistently backwards at the margins, mess up, make bugs: we wouldn't accept a compiler that did any of this at this scale or level lol

[IceDane]

It's awful, and seeing even engineers I respected become so AI pilled they're shipping slop without review has made me lose respect for them. It also can't help but make me wonder: what am I missing? Am I holding it wrong? Am I too focused on irrelevant details?

So far, my conclusion is that while LLMs can be s productivity boost, you have to direct them carefully. They don't really care about friction and bad abstractions in your codebase and will happily keep piling cards on top of the crooked house of cards they've generated.

Just like before AI, you need a cycle of building and refactoring running on repeat with careful reviews. Otherwise you will end up with something that even an LLM will have a hard time working in.

[habinero]

Right? People have put in decades of work to make them extremely reliable, they didn't magically start like that.

[Krssst]

Non-determinism is not as much of a problem as the lack of spec. C++ has the C++ norm, Python has its manual. One can refer to it to predict reliably how the program will behave without thinking of the generated assembly. LLMs have no spec.

[hirako2000]

The two come in hand.

Non determinism is what conveniently feels the gap of having no spec.

In fact turn temperature to 0. And it will be virtually deterministic. It exacerbates the problem that LLMs, as you rightly point out, have no spec.

[lukan]

"You can’t treat a prompt like source code because it will give you a different output every time you use it"

But it seems we are heading there. For simple stuff, if I made a very clear spec - I can be almost sure, that every time I give that prompt to a AI, it will work without error, using the same algorithms. So quality of prompt is more valuable, than the generated code

So either way, this is what I focus my thinking on right now, something that always was important and now with AI even more so - crystal clear language describing what the program should do and how.

That requires enough thinking effort.

[lelanthran]

Didnt work for the prod data that the AI nukes in spite of prompts saying "DON'T FUCKING GUESS", just like that in all caps: https://news.ycombinator.com/item?id=47911524

What makes you think it will work for you?

[lukan]

That I don't let agents run wild in a production environment?

[ubertaco]

You let them write code that runs in prod, which is the same thing with extra steps.

Unless you review that code carefully, and then we're back to the point about it not saving you any cognitive overhead.

[lukan]

Of course it saves me overhead by not having to read all the necessary docs etc myself and just check the resulting code and not having to type all myself.

[lelanthran]

> Of course it saves me overhead by not having to read all the necessary docs etc myself and just check the resulting code and not having to type all myself.

That link I posted upthread, the "developer" did not read the docs "because the LLM wrote the code" and hence did not realise that the token they had was a full access token, not a limited access token.

You're now boasting that you also don't have to read the docs because the LLM wrote the code?

I mean, it's literally the cause of their irreversible production deletion, in a link that you replied to, and you state you do the same?

[enraged_camel]

>> You let them write code that runs in prod, which is the same thing with extra steps.

The “with extra steps” is doing a lot of work in that sentence.

[habinero]

> if I made a very clear spec - I can be almost sure

That "almost" is doing a lot of heavy lifting here. This is just "make no mistakes" "you're holding it wrong" magical thinking.

In every project, there is always a gap between what you think you want and what you actually need. Part of the build process is working that out. You can't write better specs to solve this, because you don't know what it is yet.

On top of that, you introduce a _second_ gap of pulling a lever and seeing if you get a sip of juice or an electric shock lol. You can't really spec your way out of that one, either, because you're using a non-deterministic process.

[lukan]

Well, unfortunately it is the same with real humans who happen to be non-deterministic as well. If I give them a task, I can be allmost sure, they will do it. But even humans can have unexpected psychotic breakdowns and do destructive stuff like deleting important databases.

So right now, humans are for sure more reliable. But it is changing. There are things I already trust a LLM more than a random or certain known humans.

[byzantinegene]

your spec is a guideline, not something the LLM has to adhere to. it is definitely not guaranteed to work without error

[lukan]

Are humans guaranteed to work without error?

[xnx]

> AI coding isn’t an abstraction

Isn't it an abstraction similar to how an engineering or product manager is? Tell the (human or AI coder) what you want, and the coder writes code to fulfill your request. If it's not what you want, have them modify what they've made or start over with a new approach.

[habinero]

No, because software engineering is more than <insert coin, receive code>. I've never had a full spec dropped on my desk lol. There's no abstraction.

Software engineering is a lot more social and communication-heavy than people think. Part of my job is to _not_ take specs at face value. You learn real quick that what people say they need and what they actually need are often miles apart. That's not arrogance, that's just how humans work.

A good product manager understands the biz needs and the consumer market and I know how to build stuff and what's worked in the past. We figure out what to build together. AIs don't think and can't do this in any effective way.

Also, if you fuck up badly enough that you make your engineers throw out code, you're gonna get fired lol

[skydhash]

With an abstraction, you literally move your thinking up a level. So you move up a floor up the tower and no longer have to think what's happening below. The moment something leaves your floor, its course is set. If a result come back, its something familiar, not something from the lower floor.

A human coder can be seen as an abstraction level because it will talk to the PM in product terms, not in code. And the PM will be reviewing the product. What makes this work is that the underlying contract is that there's a very small amount of iterations necessary before the product is done and the latter one should require shorter time from the PM.

We've already established using a LLM tool that way does not work. You can spend a whole month doing back and forth, never looking at code and still have not something that can be made to work. And as soon as you look at the code, you've breached the abstraction layer yourself.