[asdfman123]

Yep, that's definitely true... for now.

The thing that concerns me, however, is that just a year ago it was only useful for doing things like reminding me how to read in a file. Now it's at the level of a competent junior, who can almost do the whole task from start to finish but needs a little supervision.

All the AI companies are retaining the messages we're sending to their agents to train the next models. I wouldn't be surprised if the agents will be able to do what I'm doing in a year or two.

UPDATE: I'm getting a lot of replies to this thread and while I'm not trying to be argumentative here, I have to ask if a lot of them aren't rooted in denial... it's been a hard adjustment process for me too.

[runarberg]

People were saying the exact same thing exactly a year ago. “Last years models were no good, but this year we can do anything; so what about next years models?” the people said in July of 2025.

Well... Now we have next years models, and people are still saying: “Last years models were no good, but this year we can do anything; so what about next years models?” And I am pretty sure in July 2027 (if this whole thing hasn’t collapsed in on it self by that point [very likely]) you all will be saying: “Last years models were no good, but this year we can do anything; so what about next years models?”

[orangedog]

Who is talking like this? Presumably the people looking forward to next year's models find some value out of this year's.

You seem to want it to collapse, but that's not the argument you person you're responding to is making.

[runarberg]

It is in fact very common:

https://hn.algolia.com/?dateEnd=1753056000&dateRange=custom&...

In particular those of us rooting for the collapse of AI were complaining about this speech a year ago as well:

https://news.ycombinator.com/item?id=44525074

[CuriouslyC]

We had O1 and Gemini 2.5 Pro last year, which were both very good models, just not for long horizon agentic tasks. They were surprisingly capable, just not without a lot of hand holding.

Current models are very good for long horizon agentic tasks, but they lack taste and higher level organizational principles. They can get a tremendous amount done with limited hand holding, but they still tend to build messy slop unsupervised/without intervention.

By this time next year hand writing code will be dead outside the rarest domains, and agents will be able to build larger projects coherently, and in two years human software taste and architectural guidance will be basically redundant. I predicted the complete automation (3+ 9s) of software engineering in 3 years back in mid 2025, if anything we're slightly ahead of schedule.

[webdood90]

You must really have your head in the sand if you truly believe AI will not soon be very close to, if not better than, humans at writing code.

[runarberg]

That is simply not technologically possible. You are speaking in science fiction here. You cannot train a neural network on existing things and then by some magic have the model be better then the training data. The laws of statistics prohibit it.

[CuriouslyC]

For code you can do better than humans on a lot of metrics pretty easily bro. This is because you can RL on objective verified outputs. Kind of like how RL was used to make agents wipe the floor with humans in Dota2, chess and go.

RLing taste and discernment are harder, but don't doubt that researchers can get a model that produces a more performant solution than you, while using less code, and being more secure/robust/etc, in a fraction of the time. Your moat for the moment is alignment with stakeholders and high level taste, enjoy it while you can.

[runarberg]

All your examples used little to no human created data and a very simple success criteria. LLMs on the other hand take massive amount of human created data (like almost all the human created data there is) and their success criteria is how well it predicts the next token. In other words, the success criteria of vibe coding models is how much it resembles the human created data. The more it resembles, the better the model.

[pydry]

For anything that isnt generic I find the effectiveness drops off very quickly.

It's already 100x better at writing todo apps than the average senior developer but when I get it to try and achieve something unusual using a relatively obscure library it flails.

[flyinglizard]

I find myself just jumping up the abstraction ladder, aiming for more and more ambitious projects. It's like electric mountain bikes; at first there was a lot of objection across the purists (and there still is), but eventually people understood you put in the same effort just go longer and faster. I work the same hours and strain my thought just as much but my output is significantly more ambitious. Left to their own devices even SOTA models like Fable would produce a complete mess over time, fine locally but awful globally. Admittedly, I don't know if it's the nature of the beast or if I'm just not thorough enough with my prompting and let the LLM guess what I want to end up with; but clearly some directional expertise is required.

[georgemcbay]

> UPDATE: I'm getting a lot of replies to this thread and while I'm not trying to be argumentative here, I have to ask if a lot of them aren't rooted in denial...

I do think there is a very common denialism that is rampant causing many people to believe on what is essentially faith that AI is going to get just good enough to be a perfect assistant to them, but then magically halt on progress right before it entirely replaces them.

And to be clear I believe LLMs are likely to plateau before they replace literally everyone working in software development, but I also think that a lot more than n% of developers believe they will be part of the n% who will survive unscathed, where n ends up being quite a small number.

[gofreddygo]

in a decade, we're all going to say how developers job is so different and yet still the same as it was 10 years ago.

[georgemcbay]

If in a decade we are discussing that online as opposed to tips on how to survive the most recent wet bulb event, then... wonderful, I'll take it.

I am less optimistic than you, but would love to be wrong.

[skydhash]

> Now it's at the level of a competent junior, who can almost do the whole task from start to finish but needs a little supervision.

Define tasks!

The most important aspects is that they’ve always been good at applying patterns, but have no discernment into why and the specific adjustments that needs to be done for a particular contexts. It’s rare that I can’t find discrepancy at the edges if several patterns have been used.

We already have good tools for patterns. It’s called code reusability. So frameworks, libraries, code generators, paradigms, snippets, and the mighty copy and paste (there’s a reason vim has like 26 registers).

Software development is all about managing complexity, which LLM are notoriously bad at. A little supervision won’t cut it.

[LinuxAmbulance]

> All the AI companies are retaining the messages we're sending to their agents to train the next models. I wouldn't be surprised if the agents will be able to do what I'm doing in a year or two.

Some things are very easy to solve via code (see https://xkcd.com/1425/), and others are extremely difficult.

For LLM based AI, it can be trained to write functionally correct code easily - there's tons of pre-existing code and documentation that can let AI learn patterns (LAMP stack, messaging queue, etc) and syntax.

Judgement is much more difficult. For some things "Should I go with a message queue for this problem?", there's a known set of solutions as to what works best in a specific situation, sure, LLMs can answer that easily.

But when it comes to turning human goals into code that does the job correctly and is maintainable and stable?

That's a heck of a lot different. Without human guidance and having tons of context gained by years of experience dealing with the very nuanced and situational judgement calls needed, LLMs are going to fail at doing a decent job.

LLMs (and junior people) don't know what they don't know. I think it will be really difficult to get a LLM to work well at a senior / staff level where it's more about coordination, working with human beings, handling situations based on legacy code, knowing what will and won't be fragile, etc.

[bnfcl]

Exactly, tacit knowledge is not easily added as AI training data.

[asdfman123]

> Some things are very easy to solve via code (see https://xkcd.com/1425/), and others are extremely difficult.

Note that comic is about a task that AI has already conclusively solved...

[LinuxAmbulance]

The irony is not lost on me!

[bnfcl]

Understanding the fundamentals of anything will always be a valuable skill, and that only comes from hard work and experience. And much of the experience lies in tacit knowledge, not easily added as AI training data.

That, combined with knowing what to build, and more importantly what NOT to build, I don’t think AI ever will take away from us.