Are people seriously getting use out of LLMs? Everything they produce is extermely sloppy in my experience, like actually mostly useless. I really dont understand the hype. Its very confusing.

I work 40 hours a week at a company I care very little for. Despite that, I still have to make 40 hours. I also have a side-project that I do in my free time. I do not have the energy to sit behind my desk for >40 hrs a week

So I walk, program, sit, get coffee, read a bit, come back and review the code. Most of the time it's fine, sometimes I had forgotten to mention something, and have to correct it, but this step doesn't take more than ~15 mins.

I then have a feature that would've taken me multiple days. Not because I need multiple days, but because I do not have the 8 hours of continuous time to work on it.

People have forgotten that when you start your programming day, you have to get up to speed which takes me longer than others. Let's say this takes ~15 mins. That means that if you spend 2 hours programming, ~12% of that was just getting back into the groove. LLMs do this instantly, it reads context, you can ask it questions, and you're up to speed again in less than a minute.

The point here is not that LLMs provide high quality code, but they do save you a bit of time and energy, which is worth a lot in my opinion.

A lot of inventions haven't been that ground-breaking; only there to save time. You can wash the dishes by hand, but you can also have a machine do it while you go watch TV.

I am convinced the people who swear by them either have very different work tasks from me, or they have very different ideas about what a job well done looks like.

I feel like everything I apply these things to sends me up a much messier and long winded route to a useable result, when compared to just doing it myself from the jump. Even the things they're ostensibly good at like sorting data comes out so messy it's practically net zero by the time you're done with quality control.

Every time I have actually engaged in conversations with people making the same claim you do, it turned out they did not really invest time into learning how to work with coding agents: They assumed, given that they’re developers, the know how to code, and thus would knew how to steer a coding agent.

That is a wrong assumption, however. An agent is an entirely new tool in your toolbox, with no similarity to any of those you already have. You will need to learn how to wield it, like a new programming language or technology you’re unfamiliar with. You will need to do some small side projects to learn. You will need to develop a feeling for how it reacts to your inputs, when to reset the session, pass it links to documentation, or interrupt it.

None of this comes intuitively. It takes time and effort, and if you’re not ready to consciously invest that, coding agents are not going to work efficiently for you. That doesn’t take away from their utility though.

Without you telling what your job entails and how the pristine artifacts of your crystalline mind fit into the larger picture it's very hard to have a conversation about this.

But yes, many developers I know haven't coded manually in months and that includes me. That doesn't mean I drink coffee and take walks while the agent codes however. I'm now in the driving seat instead of wrangling syntax and waiting for my hands to type something.

I manage the rules, the intent, the structure. If I don't like what it does, I update the coding standard, the docs, the specs, you name it. If I want a different architecture I can actually get it done in 10min instead of a week - if at all. It's more effort than coding for me, because coding is slow and methodical and incentivizes pedantry, but for me it's a massive improvement. I have been bored out of my mind for a decade now, but I'm relatively good at what I do so I just stuck with it. I'm one of those "middle management types stuck in developer role"-type of guys so I guess my personality and proclivities have something to do with it. I also know quite a few grumpy old-school types that don't get anything done with LLMs. They can't communicate, they don't understand psychology, they don't understand architecture and have a debilitating case of missing the forest for the trees. The spread, in my experience, is massive.

Just out of curiosity, did you actually use claude?

Otherwise have a look here:

https://news.ycombinator.com/item?id=48406174

Yes. Most of the people I see are on the extremes of either hating LLMs or fanatically loving it, but I am somewhere in the middle ground, I think, where I see it as a tool that sometimes helps to some extent. I don't depend on it, it almost never gives me new ideas or ways to do things, but for trivial tasks it helps me relax supervising it - usually writing the first draft for documentation or suggesting initial updates, even filling in some semi-repetitive code patterns in web apps ("create empty functions and link the buttons to it"). It is just a tool, not too bad to ignore, not too good to get more productivity, sometimes useful. Using it is like using an IDE 30 years ago or Intellisense 15-20 years ago.

LE. I see current versions of LLM like an intern that helps me doing work. We work together, I give directions and supervision and I am responsible for the results. I cannot give complex tasks and I cannot skip checking everything, but it usually helps.

I am getting serious use out of LLMs, and everyone else who does knows that those claiming otherwise are only getting slop because they’re not giving it enough guard rails, likely because they’re uninitiated, likely because they hate AI in the first place.

is a sometimes useful agent, and a sometimes useful slightly smarter search engine.

it can be useful for some automation... but its also dangerously dumb for that.