This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
I join the chorus of people who don't quite understand what your problem is with LLMs. What kind of code do you write? The tools are at the point where I can give them a picture of a screen I want along with some API endpoints and it reliably spits out immediately functioning react code. I can then ask it to write the middleware code for those endpoints and finally ask it to create a sproc for the database component. It's possible you hover high above us react monkeys and barely even consider that programming but surely you understand that's the level like at least half of all programmers operate on? I had copilot do all these things today, I know that it can do these things. So where is the disconnect? It's truly possible there is some higher plane of coding us uninspired 9-5 paycheck Andy's can only obliquely perceive and this is your standard for being able to program but it'd be nice if you could just say that to resolve the confusion.
I’ll give a description of what I do.
I manage servers. Or rather, I write code to do this, in accordance with some rather specific customer contracts. The times we take action, and the actions we take, are highly constrained. Even the basic concept of updates is not especially simple. I’m sure you remember Crowdstrike taking most of the Windows world down in a day. What I do is not so apocalyptic on the world scale, but our customers would find a similar event devastating. So most of my time is spent figuring out every possible path through server states and ensuring that they all lead back to places where faults can be cheaply recovered. These properties lie above the code. You can’t understand them, for the most part, just by reading the code. But they are incredibly important and must be thoroughly safeguarded, and even highly intelligent humans who just happen to be ignorant of the problem space or are a little careless have made really, really bad mistakes here. The code compiled, the tests passed, and it even seemed to work for a little in our integration environments - but it was horrifically flawed and came within an ace of causing material customer damage. So I don’t much trust an LLM which has a much more constrained sort of awareness, and in practice, they don’t much deliver.
I realize that’s a little vague, but I hope it explains a little about a more backend perspective on these problems. If I were more clever I’d give a clear example which was not real, but barring that, I hope a characterization helps.
More options
Context Copy link
As somone who's been working in the field of machine learning since 2012 and generally agrees with @SubstantialFrivolity's assesment, I think that what we are looking here is a bifurcation in opinion between people looking for "bouba" solutions and those looking for "kiki" solutions.
If you're a high-school student or literature major with zero background in computer science looking to build a website or develop baby's first mobile app LLM generated code is a complete game changer. Literally the best thing since sliced bread. (The OP, and @self_made_human's comments reflect this)
If you're a decently competent programmer at a big tech firm, LLMs are at best a mild productivity booster. (See @kky's comments below)
If you are decently competent programmer working in an industry where things like accuracy, precision, and security are core concerns, LLMs start to look anti-productive as in the time you spent messing around with prompts, checking the LLM's work, and correcting it's errors, you could've easily done the work yourself.
Finally if you're one of those dark wizards working in FORTRAN or some proprietary machine language because this is
SpartaIBM/Nvidia/TMSC and the compute must flow, you're skeptical of the claim that an LLM can write code that would compile at all.I think this fairly nicely summarizes how I feel. Not that I do work in one of those industries to be fair, but it's part of my personal work ethic I guess you might say. I want computers (and programs) to be correct first and foremost. Speed or ease of development don't mean much to me if the result can't be relied upon. Not only that, I want my tools to be correct first and foremost. I wouldn't accept a hammer where the head randomly fell off the handle 10% of the time or even 1% of the time. So I similarly have very little patience for an LLM which is inherently going to make mistakes in non-deterministic ways.
Preach, brother. Software is made to be clear and predictable. Learning to make it that way, one line at a time, is our craft. You can always tell the brilliant programmer apart because 99% of that code is simple as can be and 1% is commented like a formal proof. Worse than LLMs, reliance on LLMs risks undermining this skill. Who can say if something is correct if the justification is just that it came from the machine? There needs to be an external standard by which code is validated, and it must be internalized by humans so they can judge.
More options
Context Copy link
More options
Context Copy link
You have to contend with the fact that like 95+% of employed programmers are at this level for this whole thing to click into place. It can write full stack CRUD code easily and consistently. five years ago you could have walked into any bank in any of the top 20 major cities in the united states with the coding ability of o3 and some basic soft skills and be earning six figures within 5 years. I know this to be the case, I've trained and hired these people.
I did allude that there might be a level of programming where one needs to see through the matrix to do but in SF's post and in most situations I've heard the critique in it's not really the case. They're just using it for writing config files that are annoying because they pull together a bunch of confusing contexts and interface with proprietary systems that you need to basically learn from institutional knowledge. The thing LLMs are worst at. Infrastructure and configuration are the two things most programmers hate the most because it's not really the more fulfilling code parts. But AI is good at the fulfilling code parts for the same reason people like doing them.
In time LLMs will be baked into the infrastructure parts too because it really is just a matter of context and standardization. It's not a capabilities problem, just a situation where context is splined between different systems.
If anything this is reversed, it can write FORTRAN fine, it probably can't do it in the proprietary hacked together nonsense installations put together in the 80s by people working in a time where patterns came on printed paper and might collaborate on standards once a year at a conference if they were all stars. but that's not the bot's fault. This is the kind of thinking that is impressed by calculators because it doesn't properly understand what's hard about some things.
I feel like I'm taking crazy pills here. No one's examples about how it can't write code are about it writing code. It's all config files and vague evals. No one is talking about it's ability to write code. It's all devops stuff.
Ironically I considered saying almost this exact thing in my above comment, but scratched it out as too antagonistic.
The high-school students and literature majors are impressed by LLMs ability to write code because they do not know enough about coding to know what parts are easy and what parts are hard.
Writing something that looks like netcode and maybe even compiles/runs is easy. (All you need is a socket, a for loop, a few if statements, a return case, and you're done) Writing netcode that is stable, functional, and secure enough to pass muster in the banking industry is hard. This is what i was gesturing towards with "Bouba" vs "Kiki" distinction. Banks are notoriously "prickly" about thier code because banking (unlike most of what Facebook, Amazon, and Google do) is one of those industries where the accuracy and security of information are core concerns.
Finally which LLM are you using to write FORTRAN? because after some brief experimentation niether Gemini nor Claude are anywhere close.
What do you imagine is the ratio just at banks between people writing performant net code and people writing crud apps? If you want to be an elitist about it then be my guest, but it's a completely insane standard. Honestly the people rolling out the internal llm tooling almost certainly outnumber the people doing the work you're describing.
I do not think that expecting basic competency is an "insane standard" or even that elitist. Stop making excuses for sub-par work and answer the question.
Which LLM are you using to write FORTRAN?
What sort of problem did you ask it to solve?
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
I mean, my full opinion and experience with LLMs is much harsher than my comment suggested, but I don’t want to start fights with enjoyers on the net. (At least, not this time.) Chances are their circumstances are different. But I would be seriously offended if someone sent me AI-generated code in my main area of expertise because it would be subtly or blatantly wrong and be a serious waste of my time trying to figure out all the errors of logic which only become apparent if you understand the implicit contracts involved in the domain. Goodness knows it’s bad enough when merely inexperienced programmers ask for review without first asking advice on how to approach the problem, or even without serious testing…
I know that pain.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Oh for heaven's sake, dude. When did I ever say I consider myself better than anyone else, that I would deserve such a litany of sarcasm directed at me? I don't think that and certainly haven't said it. I am just an ordinary programmer - I doubt very much that I'm better at programming than anyone here except the non-programmers, and I'm sure I'm worse than more than a few. Not only did I say "hey I'm not trying to litigate this right now" and that got ignored, now I get people dogpiling me saying I'm a troll or think I'm better than everyone else or whatever.
But fine, since you and @SnapDragon are insistent on pressing me on the topic (and since I apparently didn't say to him what my experience was, my bad on that, but I know I have posted this in a previous thread before), I will reiterate the things that I personally have seen LLMs fall flat on their face with. This is of course in addition to the various embarrassments that are public, like Microsoft's ill-conceived attempt to let Copilot loose on PRs.
These were all within the last year, though I couldn't tell you exactly when or what model or anything. And I've been honest that sometimes it has done good work for me, namely in generating short snippets of code in a language (or using an API) that I know well enough to recognize as correct when I see it, but not well enough to produce without laborious reading of docs. I've never claimed that LLMs work 0% of the time (if people have taken that away, I've done a poor job communicating), but the failure rate is much too high for them to be considered viable tools in my book. Most frustratingly, the things that I actually need help on, the ones where I don't know really anything about the topic and a workable AI assistant would actually save me a ton of time, are precisely the cases where it fails hard (as in my examples where stuff doesn't even work at all).
So those are again my experiences with LLMs that have caused me to conclude that they are hype without substance. Disagree if you like, I don't mind if you find it useful and like I have tried to say I'm not actually trying to convince people of my views on this topic any more. Like I tried to say earlier, the only reason I posted in this thread was to push back on the idea that one simply must be ignorant if they don't think LLMs are good at coding (and other things). That idea is neither true, necessary, or kind (as the rules allude to) and I felt that it deserved some sort of rebuttal. Though heaven knows I wish I had just left it alone and had peace and quiet rather than multiple people jumping down my throat.
Apologies if I came on too hard, it's just you've been expressing this opinion for a while and had gone down several reply chains without bringing the thing to the object level. It's emblematic of the whole question, AI is "spikey", as in it's very good at some things and inexplicably bad at some other things. I don't think a lot of people would take so much offense if you just said it still seems bad at some tasks, that's broadly a consensus. But when you just say it "sucks at code" it's perplexing to the people watching it effortlessly do wide swaths of what used to be core programming work.
I could definitely see it struggle with highly context dependent config files but something seems strange about it not producing at least a valid file, did you try different prompts and giving it different contexts? I find giving it an example of valid output helps but I'm not familiar with fluentd and it's possible giving it enough context is unreasonable.
I have not tried that, but it also seems like kind of a failure of the tool if I have to, you know? The whole point of a tool that can understand natural language is that you can just talk to it normally. If one has to figure out how to word the incantations just right to get a useful result... I'm not sure how that's better than just figuring out the code myself at that point.
Prompting is a skill like any other. Sending it off without context is like telling an underling to fix your config file without explaining or letting them look at the system they're writing it for. It's often a mistake to assume the prompt needs to be something a human would understand. You can and should just dump unformatted logs, barely related examples of working config files, anything you can imagine an underline with infinite time in a locked room might find useful in solving your problem.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
FWIW, I appreciate this reply, and I'm sorry for persistently dogpiling you. We disagree (and I wrongly thought you weren't arguing in good faith), but I definitely could have done a better job of keeping it friendly. Thank you for your perspective.
That does sound like a real Catch-22. My queries are typically in C++/Rust/Python, which the models know backwards, forwards, and sideways. I can believe that there's still a real limit to how much an LLM can "learn" a new language/schema/API just by dumping docs into the prompt. (And I don't know anything about OpenAI's custom models, but I suspect they're just manipulating the prompt, not using RL.) And when an LLM doesn't know how to do something, there's a risk it will fake it (hallucinate). We're agreed there.
Maybe using the best models would help. Or maybe, given the speed things are improving, just try again next year. :)
Thanks. And for my part I'm sorry that I blew you off unjustly; I really thought I had explained myself in detail but I was wrong.
And yeah, the tech might improve. I imagine you can see why I'm skeptical of the strong predictions that it'll do so (given that I don't agree it's as good as people say it is today), but I try to keep an open mind. It is possible, so we'll see.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link