[prplfsh]

I really love using AI to code but more and more I wonder ... Are things really that different? So I guess I'm the business as usual type.

I think on the frontend side we're going to see a lot more scope for teams.

On a backend infra side it seems as hard as ever. Still have to think really hard problems, think deeply about data structure and flow, and deal with second- and third-order effects. Or even harder because the models like to confidently lie.

The harder question is how we train people but that doesn't seem insurmountable either. Most of us cut our teeth as junior engineers somewhere, implementing tasks that Claude can now do without breaking a sweat but was that really the most efficient way to train and learn?

[englishspot]

I work on backend infra. I touch a lot of things and have multiple instances of claude code running at once. I do feel like I'm doing more simultaneously. but that's still more that I have to keep track of and make sense of in my head. claude is making me way more productive, for sure, but at the same time I feel way more overwhelmed than I've ever been.

idk how anyone else is doing it and managing all of this. supposedly there's people with large teams of agents that they can just trust to do everything end-to-end.

[steve_adams_86]

I’ve been finding I can mitigate the sense of being overwhelmed by trying to keep my tasks within the same domain or part of the code base. I know it’s not always an option, though.

I guess the goal is to stack context rather than spread it.

I find it incredibly exhausting otherwise. It’s a skill I don’t have in spades. I’ll have to develop it, though.

I don’t think the people trusting agents end to end are getting the quality they think they are. I’m bullish on AI as a permanent and genuinely useful tool in our kit, but I’m not seeing signs that you can actually let them loose at all. Looping on a very well defined problem, sure. But open-ended tasks across large infrastructures and complex domains, no, it’s going to be duct tape and sprawl as far as the IDE can see. It’s going to sprawl faster than context limits can grow, and the mess will only get worse.

From what I can tell, the solution is an ever-expanding roster of agents to make piece work of these immense tasks. The token expense for this approach is insane, though. I used deep-research in Claude Code today and it dispatched 103 agents and consumed something like 3.5MM tokens in ten minutes. That can’t be the future, can it?

[gxs]

Right there with you - this has been my experience as well, to a tee

I can do more now…so I do - it’s really that simple

And it’s way more exhausting because there’s no room to breathe - fresh code to work with every few minutes with prepping for the next set of tasks in between

On the one hand the dopamine’s got me hooked on this like a video game

On the other…I’m as overwhelmed as ever line you said, even if it’s my own doing

[AaronAPU]

It is definitely a different type of stress. I used to lock in on one problem in deep work. Now it’s a constant juggling of 7 problems. It feels good to close one out, but your other 6 keep you perpetually locked in the stress loop.

[englishspot]

idk if anyone here is a gamer, but it feels like I'm playing Dynasty Warriors. endless wave after wave after wave of things to do, notifications about fires to put out all over the map, all at once.

and that's just on my terminal.

[slopinthebag]

Yeah I mean, as far as I can tell the result of the agent mania is the same amount of software, but an acceleration in the decline of performance and quality. I'm also increasingly seeing early adopters going back to a more "traditional" approach to development. So idk, maybe the result will be more jobs fixing up all the vibe code while we transition to a more mature implementation of language models into our workflows.

[glouwbug]

I’ve noticed a massive reverse in AI sentiment in the last 3 weeks here on HN.

It’s not that I don’t disagree, but I wonder what’s going on. Maybe it’s the IPO

[hootz]

Reversing to which direction? Because what I've always seen here is a pretty good mix of positive and negative sentiments. Usually we get a lot of AI related submissions, but with skeptics/opposers in the comments.

[glouwbug]

I’m not sure. I’ve been reading death-of-the-software engineer for years, but recently the -vibe- feels different. I don’t have anything anecdotal to back it up so take it with a grain of salt. I might be reading what I want to see

[slopinthebag]

I'm assuming it's a turn to the negative and not more positivity you're seeing? Geohot's article and Hasimoto's tweet about AI psychosis kind of made me pay attention.

https://news.ycombinator.com/item?id=48263238

https://xcancel.com/mitchellh/status/2055380239711457578

[glouwbug]

Yeah more or less

[BobbyTables2]

In my opinion, as AI was oversold for too long, it was was easy to dismiss it. Classical image processing was marketed as “AI”. Doomsday predictions about AI seemed laughable, just as SkyNet in the Terminator seemed unrealistic.

The early ChatGPT versions were also pretty silly and equally oversold.

At this point, the popular messaging of AI is still 90% fiction but the remaining real 10% is now a force to be reckoned with.

Companies laying off Indian call center employees to replace with AI is something I never would have dreamed of.

My experience of using AI as a search engine has surprised me. I never expected an overgrown pile of matrices to work that well.

[nextaccountic]

> My experience of using AI as a search engine has surprised me. I never expected an overgrown pile of matrices to work that well.

The first version of Google was also surprising. Mind blowing use of linear algebra (also a pile of matrices, but this time sparse matrices mostly) to rank websites

So maybe the search business was always meant to use pile of matrices

[gymbeaux]

LLMs seem to be best at writing web apps intended to be internal tools. I don’t need to ever really read the code because the functionality is relatively simple- but still valuable. I had Claude Code build me a CLI tool for running common kubectl commands against our EKS cluster. With EKS (AWS) you generally use the AWS CLI to authenticate and then choose the correct EKS cluster and then you can run your kubectl commands. So this CLI tool remembers all the AWS accounts and EKS cluster names and namespaces and pods that I care about. It’s been a huge time saver, and I never would have had the time to build it as part of my day to day. And because it’s an internal tool, I don’t have to worry about things like security/authentication, and any bugs aren’t as big a deal as if they were customer-facing. I wouldn’t want to use it for generating an app that is going on the public internet and doing anything sensitive/important/valuable with my or other people’s data (I’m sure that doesn’t stop others from vibe coding commercial products).

[Fire-Dragon-DoL]

My feelings right now can be described as follows:

If AI progress stalls now or grinds to a halt, we get to keep a lot of new jobs that are going to show up (it opens a lot of doors), while maintaining software devs as an interesting career.

Egoistically, I would love that.

If it keeps going and software devs get replaced, so many jobs are going to disappear.

[gymbeaux]

Even though the synthetic benchmarks paint a picture of LLMs coming a long way since 2022, my practical experience has been that they aren’t tangibly better. No doubt someone reading this will chime in and say LLMs are way better at writing code or whatever, and maybe that’s true, but there’s no difference between ChatGPT 3.5 and Claude Opus 4.8 as far as my trusting the output. Opus 4.8 still messes up plenty. It’s particularly bad with identifying and fixing CI yaml, but it struggles in the usual areas too.

So I’m thinking we’ve just about reached apex with LLMs, and they have failed at replacing software engineers (companies can freeze hiring juniors at their own, future peril using any excuse they like).

[bigstrat2003]

Yep, that has been my experience as well. There hasn't been any meaningful improvement in LLMs since ChatGPT first launched. They still fall over, in the same ways, and with more or less the same high rates.

[Fire-Dragon-DoL]

The difference for me has been that you can use llm as your main typing interface, you couldn't do that (without being annoying) before I think opus 4.5

[whateveracct]

i've heard of principle level engineers saying their typing is bad now.

skill atrophy is real. especially for the stupid.

[Fire-Dragon-DoL]

I'm typing to talk to the LLM, so definitely not lost.

The main atrophy I'm concerned about is keeping in mind the state of a massive piece of code executing in my mind.

Now with LLM I just ask "show me the state by putting comments in the code" and I just read it

[nextaccountic]

Software developers are in the business of making other people unemployed, so... the second scenario is poetic justice in a sense

[Fire-Dragon-DoL]

Every job that's not art-related is about making other people unemployed (by improving efficency)

[stlark]

In my experience I'd have to agree. I'm shipping more and I'm onboarding onto domains faster but the same bottlenecks exist, the same complexities creep up and the same talents that help individuals push through are still relevant.

[sam0x17]

What I like is it's still hard, but you can do things like "prove this using game theory" or "find the optimal value for this and a proof" without having to know game theory or deep math really

[stvltvs]

How do you ensure that the output isn't BS if you don't know those areas? What value does it add to have a mathematical proof of questionable validity?

[nh23423fefe]

if you have an intuition something is true, you can verify the proof. even if you couldn't construct the proof. allowing skepticism to terminate thought is a dead end.