There is an amazing quirk of the LLM, whenever I don’t know about the topic, and refuse to google, it gives me some useful answers, but if I ask it something I know about, the answers are always stupid and wrong. I asked a computer about it but it said that everything is normal and I should buy better subscription, so there’s that.
It’s not a quirk of LLMs, it’s a quirk of human cognitive biases.
See: Gell-Mann amnesia effect.
The miracle slop machine miraculously produces garbage.
Shocker.
Usage is rising because corporate executives started getting kickbacks and thinking they could cut staff by implementing it. But developers who have actually had to use it have realized it can be useful in a few scenarios, but requires a ton of review of anything it writes because it rarely understands context and often makes mistakes that are really hard to debug because they are subtle. So anyone trying to use it for a language or system they don’t understand well is going to have a hard time.
because it rarely understands context
It never understands context.
And it cannot understand context because it does not think, it’s just an expensive prediction tool
It’s just Markov chains running on a heap of poorly governed sources. Expensive because Markov chains are brutish to process. The only intelligence is in the source they calculate on. Literally a basic “what goes in comes out” math process.
Counterpoint: they want number go up.
Pro Tip: it doesn’t even matter if number go up, when they know how to suck up to even higher-ups.
That’s not true. If you give it context, it understands and retains context quite well. The thing is that you can’t just say “write code for me” and expect it to work.
Also, certain models are better than certain tasks than others.
Executives are getting kickbacks? I’ve gotta do some research here.
This is a true statement.
There’s relatively little debate among developers that the tools are or ought to be useful,
Yes there is. No one wants to listen to us. I’ve had 3 levels of people above me ask me how I’ve incorporated AI into my workflow. I don’t get any pushback because my effectiveness is well known, yet the top down edict that everyone else use these shitty tools continues unabated.
Where I work, my skip-levels have started debating on whether they want to consider if an engineer uses AI as a factor for reviews, pay raises and incentives, and are tracking who uses it by way of licenses.
It’s a bit ridiculous IMO because they’re essentially saying “we don’t care if slop makes it into the code base, so long as you are using AI, you will remain gainfully employed.”
I’ve seen a lot of stupid shit over my career but this AI zealotry just takes the cake.
Everyone is so convinced these tools will make software get made faster, but I’m not even convinced that it gives even a modest benefit. For me personally they definitely don’t, and it seems to lead junior devs horribly astray as often as it helps speed them up.
It feels like I’m not even looking at the same reality as everyone else at this point.
I wish I had one like you at where I work, just to keep sane for a bit longer. Everyone else is just way too into this shit even though for most it seems like a gigantic time sink.
I mean “ought to be useful,” sure that would be nice. They ain’t, but perhaps “ought to be.”
It’s useful for things I’d otherwise Google. It makes a great ORM, when you know exactly what you want to do with a lot of mundane code. And it’s so much better than adding a framework for an ORM.
I don’t like ORMs, but I’d rather use a battle tested ORM than some vibe coded data layer.
You know, you can just look at it. It’s pretty simple, easy to look at, pretty repetitive code where it’s generally pretty easy to spot if something’s wrong.
Vibe coding is more hitting “accept all” and not looking at it at all (or not knowing how to look at it).
Or, I could just write it myself, instead of ending up like these guys https://sketch.dev/blog/our-first-outage-from-llm-written-code
Last job, year ago as of today, worked at small software dev. We were all talking about AI’s usefulness.
No one, not a soul from the CEO down misunderstood the applications. Yeah, it’s great for getting over a hump. Stuck? Meh, maybe the LLM will kick out a useful path I hadn’t known or considered. Got around a problem with PowerShell and Google Calendar, made a neat integration, far faster than I could have figured it myself. All I got was a couple of lines of code, all I needed, wrote the rest myself. Kinda like stealing code off any given site, but faster. Wonderful tool, really!
End of story. AI isn’t writing end-to-end working code, as some leaders think. It’s a damned useful tool for getting around roadblocks, so yeah, you can code faster. We were all amazed. But no one thought it was intelligent or would replace us. But in the end, we’ll need fewer dev hours. Sorry. That’s simply true for competent firms.
And all the boilerplate. And the test cases. And the CLI one liners. Maybe even small refactorings.
AI is amazing when used correctly.
Yep, and the general public is too stubborn to accept a little thing like nuance. Neither are the CEO assholes that can’t stop talking about layoffs and replacing jobs, out in the open.
Geez louise, maybe it’s not intelligence after all. I think you’d need sentience to apply the word intelligence. Those wacko marketing people.
LLMs will always fail to help developpers because reviewing is harder than writting. To review code effectivly you must know the how and why of the implementation in front of you and LLMs will fail to provide you with the necessary context. On top of that a good review evaluate the code in relation to the project and other goal the LLM will not be able to see.
The only use for LLM in coding is as an alternative search bar for stackoverflow
The only use for LLM in coding is as an alternative search bar for stackoverflow
I’d argue it can also be useful as a form of autocomplete, or writing whatever boilerplate code; that still isn’t outsourcing your thinking to the text predictor.
When I tried the autocomplete in IntelliJ it kept trying to guess what I wanted to do instead of autocompleting what I was typing so I don’t know about that part.
Still millions of ton of CO2 for a search bar and autocomplete doesn’t seems like a good idea.
Probably depends on what you do. I haven’t used AI autocomplete myself, so I can’t talk from experience, but what I had in mind was the somewhat repetitive work I’ve been doing recently with gui widgets. I expect an LLM to get that mostly right.
AI coding tools are a great way to generate boilerplate, blat out repetitive structures and help with blank page syndrome. None of those mean you can ignore checking what they generate. They fit into the same niche as Stackoverflow - you can yoink all the 3rd party code snippets you want, but it’s going to be some work to get them running well even if you understand what they’re doing, and if you neglect this step hoo boy can it come back to bite you!
Great, a ridiculously expensive lorum ipsum generator.
Sometimes lorum ipsum is useful.
Sure, but you don’t need an LLM for that. That’s like using a bazooka to kill a housefly.
This has been my argument for a while. If you’re doing boilerplate once in a while, it’s a good way to keep even the boring part of your skills sharp.
If you’re doing it regularly, just make a fucking template you can copy paste, or set it up in your IDE’s code snippet functionality.
I just see this as future job security.
Oh, AI fucked your codebase? Well, it’ll take twice as much time to undo it all and fix it, my rate is $150/hr. Thanks.
That’s and awful long time to do
git checkout HEAD~10Shhhh, they don’t need to know this.
Hey look, handmade NFT!
This is a really old jpeg. If I remember correctly it was meant as a protest joke against NFTs. NFTs claim was that each was unique and couldn’t be copied or something. Someone replied with this quick doodle on paper saying that this was original, unique, and truly could not be copied and they’d part with it for a measly $5,000,000.
Paraphrasing a lot, but that was the gist of it. I cropped it and added the text for the memes.
There was trust?
But now you can spend 4 hours trying to get it to say the right thing and do it all in one output. Where as before it might take you on your own 1 hour if you were having a bad day.
