[koolba]

> The act of typing isn't the hard part - its understanding what's going on, and why you're doing it. Using AI to generate code is only faster if you try and skip that step - which leads to an inevitable disaster

It’s more than just typing though. A simple example remembering the exact incantation of CSS classes to style something that you can easily describe in plain English.

Yes, you could look them up or maybe even memorize them. But there’s no way you can make wholesale changes to a layout faster than a machine.

It lowers the cost for experimentation. A whole series of “what if this was…” can be answered with an implementation in minutes. Not a whole afternoon on one idea that you feel a sunk cost to keep.

[lelanthran]

> It’s more than just typing though. A simple example remembering the exact incantation of CSS classes to style something that you can easily describe in plain English.

Do that enough and you won't know enough about your codebase to recognise errors in the LLM output.

[groby_b]

That's a bold assertion without any proof.

It also means you're so helpless as a developer that you could never debug another person's code, because how would you recognize the errors, you haven't made them yourself.

[mrklol]

imo a question is, do you still need to understand the codebase? What if that process changes and the language you’re reading is a natural one instead of code?

[lelanthran]

> What if that process changes and the language you’re reading is a natural one instead of code?

Okay, when that happens, then sure, you don't need to understand the codebase.

I have not seen any evidence that that is currently the case, so my observation that "Continue letting the LLM write your code for you, and soon you won't be able to spot errors in its output" is still applicable today.

When the situation changes, then we can ask if it is really that improtant to understand the code. Until that happens, you still need to understand the code.

[kjksf]

The same logic applies to your statement:

> Do that enough and you won't know enough about your codebase to recognise errors in the LLM output.

Okay, when that happens, then sure, you'll have a problem.

I have not seen any evidence that that is currently the case i.e. I have no problems correcting LLM output when needed.

When the situation changes, then we can talk about pulling back on LLM usage.

And the crucial point is: me.

I'm not saying that everyone that uses LLM to generate code won't fall into "not able to use LLM generated code".

I now generate 90% of the code with LLM and I see no issues so far. Just implementing features faster. Fixing bugs faster.

[lelanthran]

> The same logic applies to your statement:

>> Do that enough and you won't know enough about your codebase to recognise errors in the LLM output.

> Okay, when that happens, then sure, you'll have a problem.

It's not exactly the same: how will you know that you are missing errors due to lack of knowledge?

> I now generate 90% of the code with LLM and I see no issues so far.

Well, that's my point, innit? "I see no errors" is exactly the same outcome from "missing the errors that are generated".

[pdimitar]

You do have a point but as the sibling comment pointed out, the negative eventuality you are describing also has not happened for many devs.

I quite enjoy being much more of an architect than I could compared to 90% of my career so far (24 years in total). I have coded my fingers and eyes out and I spot idiocies in LLM output from trivially easy to needing an hour carefully reviewing.

So, I don't see the "soon" in your statement happening, ahem, anytime soon for me, and for many others.

[lelanthran]

> I have coded my fingers and eyes out and I spot idiocies in LLM output from trivially easy to needing an hour carefully reviewing.

This is exactly the opposite experience of sibling, who reports not seeing any issues in the generated code.

You report spotting idiocies, he reports seeing nothing, and you are both making the same argument :-/

[pdimitar]

Not really. We are both reporting our own anecdotal evidence.

There are common threads though. LLMs do terribly in certain areas. They also do terribly when not supervised well.

[cassianoleal]

What happens when your LLM of choice goes on an infinite loop failing to solve a problem?

What happens when your LLM provider goes down during an incident?

What happens when you have an incident on a distributed system so complex that no LLM can maintain a good enough understanding of the system as a whole in a single session to spot the problem?

What happens when the LLM providers stop offering loss leader subscriptions?

[peyton]

AFAIK everything I use has timeouts, retries, and some way of throwing up its hands and turning things back to me.

I use several providers interchangeably.

I stay away from overly complex distributed systems and use the simplest thing possible.

I plan to wait for some guys in China to train a model on traces that I can run locally, benefitting from their national “diffusion” strategy and lack of access to bleeding-edge chips.

I’m not worried.

[the_af]

> What if that process changes and the language you’re reading is a natural one instead of code?

Natural language is not a good way to specify computer systems. This is a lesson we seem doomed to forget again and again. It's the curse of our profession: nobody wants to learn anything if it gets in the way of the latest fad. There's already a historical problem in software engineering: the people asking for stuff use plain language, and there's a need to convert it to a formal spec, and this takes time and is error prone. But it seems we are introducing a whole new layer of lossy interpretation to the whole mess, and we're doing this happily and open eyed because fuck the lessons of software engineering.

I could see LLMs being used to check/analyze natural language requirements and help turn them into formal requirements though.

[locknitpicker]

> But it seems we are introducing a whole new layer of lossy interpretation to the whole mess (...)

I recommend you get acquainted with LLMs and code assistants, because a few of your assertions are outright wrong. Take for example any of the mainstream spec-driven development frameworks. All they do is walk you through the SRS process using a set of system prompts to generate a set of documents featuring usecases, functional requirements, and refined tasks in the form of an actionable plan.

Then you feed that plan to a LLM assistant and your feature is implemented.

I seriously recommend you check it out. This process is far more structured and thought through than any feature work that your average SDE ever does.

[the_af]

> I recommend you get acquainted with LLMs and code assistants

I use them daily, thanks for your condescension.

> I could see LLMs being used to check/analyze natural language requirements and help turn them into formal requirements though.

Did you read this part of my comment?

> Take for example any of the mainstream spec-driven development frameworks. All they do is walk you through the SRS process using a set of system prompts to generate a set of documents featuring usecases, functional requirements, and refined tasks in the form of an actionable plan.

I'm not criticizing spec-driven development frameworks, but how battle-tested are they? Does it remove the inherent ambiguity in natural language? And do you believe this is how most people are vibe-coding, anyway?

[locknitpicker]

> Did you read this part of my comment?

Yes, and your comment contrasts heavily with the reality of using LLMs as code assistants, as conveyed in comments such as "a whole new layer of lossy interpretation. This is profoundly wrong, even if you use LLMs naively.

I repeat: LLM assistants have been used to walk users through software requirements specification processes that not only document exactly what usecases and functional requirements your project must adhere to, but also create tasks and implement them.

The deliverable is both a thorough documentation of all requirements considered up until that point and the actual features being delivered.

To drive the point home, even Microsoft of all companies provides this sort of framework. This isn't an arcane, obscure tool. This is as mainstream as it can be.

> I'm not criticizing spec-driven development frameworks, but how battle-tested are they?

I really recommend you get acquainted with this class of tools, because your question is in the "not even wrong" territory. Again, the purpose of these tools is to walk developers through a software requirements specification process. All these frameworks do is put together system prompts to help you write down exactly what you want to do, break it down into tasks, and then resume the regular plan+agent execution flow.

What do you think "battle tested" means in this topic? Check if writing requirements specifications is something worth pursuing?

I repeat: LLM assistants lower formal approaches to the software development lifecycle by orders of magnitude, to the point you can drive each and every single task with a formal SRS doc. This isn't theoretical, it's month's old stuff. The focus right now is to remove human intervention from the SRS process as well with the help of agents.

[sublinear]

This is not correct. CSS is the style rules for all rendering situations of that HTML, not just your single requirement that it "looks about right" in your narrow set of test cases.

Nobody writing production CSS for a serious web page can avoid rewriting it. Nobody is memorizing anything. It's deeply intertwined with the requirements as they change. You will eventually be forced to review every line of it carefully as each new test is added or when the HTML is changed. No AI is doing that level of testing or has the training data to provide those answers.

It sounds like you're better off not using a web page at all if this bothers you. This isn't a deficiency of CSS. It's the main feature. It's designed to provide tools that can cover all cases.

If you only have one rendering case, you want an image. If you want to skip the code, you can just not write code. Create a mockup of images and hand it off to your web devs.

[slopinthebag]

Eh, I've written so much CSS and I hate it so much I use AI to write it now not because it's faster or better at doing so, just so I don't need to do it.

[ufish235]

So AI is good for CSS? That’s fine, I always hated CSS.

[CivBase]

> But there’s no way you can make wholesale changes to a layout faster than a machine.

You lost me here. I can make changes very quickly once I understand both the problem and the solution I want to go with. Modifying text is quite easy. I spend very little time doing it as a developer.

[EagnaIonat]

> It lowers the cost for experimentation. A whole series of “what if this was…”

Anecdotal, but I've noticed while this is true it also adds the danger of knowing when to stop.

Early on I would take forever trying to get something exactly to whats in my head. Which meant I would spend too much time in one sitting then if I had previously built it by hand.

Now I try to time box with the mindset "good enough".