10 things I learned from burning myself out with AI coding agents

Vastin

Smack-Fu Master, in training
15
Unfortunately, your last point is likely the most important - you can make things too fast.

The fact is, the coding industry is already suffering for that, as is the media landscape. It's already very inexpensive to make things quickly, and as any good economist will tell you, value is a matter of supply vs demand, and AI will vastly increase supply without doing anything to increase demand - the value of all the products so created will drop, precipitously.

If software was very expensive, this would be a relief - but it isn't. It's already cheap, particularly in the entertainment field, and taking the cost trend too far is also bad for industries. It devalues everything they do, potentially to the point of collapse.

Right now in the game industry, there are a lot more games than anyone can ever play. Even a gamer that wants to indulge entirely in a highly specific niche will likely find more than they could ever play. AAA development is being shredded by forces that it cannot manage, as they struggle to make the best possible game to get players highly divided attention - but most of them fail, burning hundreds of millions of dollars in the process. At the other end Indy development is booming and creating a renaissance in gaming, but even in this idyllic consumer case most indy developers will certainly fail even if they make a very good game, because only a small number can ever bubble up into public view, even with a surprisingly egalitarian platform like Steam aiding discoverability.

This is all before AI has had a real impact on development, and unfortunately the results will likely be disastrous for the industry, with an appalling flood of new content that no-one will ever see. Amazon's self-publishing book section has already gone down in flames under an onslaught of low-effort low-quality AI books flooding out the efforts of any human author - but they aren't making money either, no-one is, and a sector that was growing and thriving is suddenly basically dead.

Like cheap-ass plastic toys and kitchen utensils, there can be such a thing as too cheap, and it can badly damage entire market sectors, destroying the ability of people to engage with that market as they are flooded with mass produced spam that makes discoverability of quality releases functionally impossible, and drives the value of even high quality releases close to zero - eventually consumers give up and go find someplace to spend their money where they can still find some tangible value for their money.
 
Upvote
16 (17 / -1)

nray

Smack-Fu Master, in training
70
Two points:

1. Like most articles on LLMs, there isn't a discussion of the underlying real economics of developing, training and running these models. I understand why, because cost is a complex multi-faceted topic which would be complex and time-consuming to explore. There are the practical costs (currently decoupled from any pricing scheme), the environmental costs (both to people's local environments in proximity to datacenters and larger global ones due to increased consumption of carbon-rich fuels), and societal costs (degredation in the reliability of information due to ease of production of non-factual but convincing media, psychosis induced in otherwise healthy individuals from prolonged interaction with LLMs, accelerated social isolation created by LLMs attracting already isolated people into net-negative pseudo-social interactions, and general confusion people due to assumptions that confident-sounding responses represent some sort of knowledge which LLMs are not actually capable of but people mistakenly think is there).

2. Anecdotal results are entertaining, but that is mostly what I'm seeing across the board. I yearn for articles that would at least have two people of equal skill tackle similar problems over the same timeframe, one leveraging LLMs, the other not, and have the results judged by an independent third party. In an ideal world we'd have several groups of people at different skill levels all taking part for a better data set. This would help alleviate intrinsic biases humans have especially when interacting with LLMs that have been primed for sycophantic responses.
 
Upvote
13 (13 / 0)

Jensen404

Ars Scholae Palatinae
1,092
I wonder whether there is any chance that ChatGPT is reporting my results. Sort of freaking me out.
Sounds like there’s a good chance of that.
I had a conversation with someone on Reddit about some minutia of how the displays on the Apple Vision Pro work. I had trouble finding information online.
The next day, I tried asking ChatGPT. In its response, it referenced the conversation I had the day before.

I wish I knew.

If the response you received didn’t have any links, it was probably part of its training, while in my case it did a web search. If it’s part of the model itself, ChatGPT itself probably doesn’t know where it got that info. If you asked it, it would just do a web search to find a reference, but not necessarily the original source of its training.
 
Upvote
1 (1 / 0)
“Meanwhile, the poor Python, by effectively removing all barriers to communication between LLM companies and hobby code scraped from Stack Overflow, has caused more and sloppier programming than anything else in the history of creation.”
I read that in my head in a David Attenborough voice...
 
Upvote
2 (2 / 0)
Unfortunately, your last point is likely the most important - you can make things too fast.

The fact is, the coding industry is already suffering for that, as is the media landscape. It's already very inexpensive to make things quickly, and as any good economist will tell you, value is a matter of supply vs demand, and AI will vastly increase supply without doing anything to increase demand - the value of all the products so created will drop, precipitously.

If software was very expensive, this would be a relief - but it isn't. It's already cheap, particularly in the entertainment field, and taking the cost trend too far is also bad for industries. It devalues everything they do, potentially to the point of collapse.

Right now in the game industry, there are a lot more games than anyone can ever play. Even a gamer that wants to indulge entirely in a highly specific niche will likely find more than they could ever play. AAA development is being shredded by forces that it cannot manage, as they struggle to make the best possible game to get players highly divided attention - but most of them fail, burning hundreds of millions of dollars in the process. At the other end Indy development is booming and creating a renaissance in gaming, but even in this idyllic consumer case most indy developers will certainly fail even if they make a very good game, because only a small number can ever bubble up into public view, even with a surprisingly egalitarian platform like Steam aiding discoverability.

This is all before AI has had a real impact on development, and unfortunately the results will likely be disastrous for the industry, with an appalling flood of new content that no-one will ever see. Amazon's self-publishing book section has already gone down in flames under an onslaught of low-effort low-quality AI books flooding out the efforts of any human author - but they aren't making money either, no-one is, and a sector that was growing and thriving is suddenly basically dead.

Like cheap-ass plastic toys and kitchen utensils, there can be such a thing as too cheap, and it can badly damage entire market sectors, destroying the ability of people to engage with that market as they are flooded with mass produced spam that makes discoverability of quality releases functionally impossible, and drives the value of even high quality releases close to zero - eventually consumers give up and go find someplace to spend their money where they can still find some tangible value for their money.
In short, curation and exclusivity are necessary which means when demand is low there are more losers.

If you want fewer losers, quality going up has a certain ceiling but you have to find a way to drive demand and AI doesn't help with this at all no matter how you use it to complete a project.

AI is marketed as if it can increase productivity AND demand and the former has major qualifiers, the latter is never in evidence no matter how much shill propaganda is written otherwise.
 
Upvote
6 (7 / -1)

Aurich

Director of Many Things
41,280
Ars Staff
Anecdotal results are entertaining, but that is mostly what I'm seeing across the board. I yearn for articles that would at least have two people of equal skill tackle similar problems over the same timeframe, one leveraging LLMs, the other not, and have the results judged by an independent third party. In an ideal world we'd have several groups of people at different skill levels all taking part for a better data set. This would help alleviate intrinsic biases humans have especially when interacting with LLMs that have been primed for sycophantic responses.
There have been attempts at this, with experienced programmers, not a whole spectrum of levels. They didn't show that the LLMs actually helped.

But it's a difficult thing to really study. There isn't a true way to A/B someone doing the same task with and without a tool.
 
Upvote
15 (15 / 0)

clewis

Ars Tribunus Militum
1,837
Subscriptor++
More and more as time goes on, I feel like I am the odd duck out here. I actually like to learn stuff, I want to spend the time and effort to grasp things myself instead of relying on LLMs, I'm not afraid of having to put some work into it.

I randomly decided to start learning Rust like two weeks ago or something and I've been programming something akin to Autohotkey as my learning project and I have used Gemini while at it, but I've only used it as a learning aid, not to design or write the code for me -- I pop in to ask if I can't quite figure out how to do something or if I am hitting an error I can't quite grasp and since I am not familiar with the win32 API, it has been useful in figuring out some of the functions I should use to achieve something, but it's very much just an aid for me to find the information I need faster. I could still get all this done even without it, it'd just take some more time for me to e.g. browse through the win32 API documentation.

And then I see the constant stream of Reddit-posts, YouTube-videos, various blog-posts and so on all just nilly-willy using LLMs without trying to learn even the tiniest bit themselves or putting in any more effort than just writing prompts. It's an fscking constant stream of shite and it annoys me. It makes me worried: I feel like people like me, people willing to learn, are a dying breed.
A buddy of mine had one of the agents build him a Syllabus for a class to teach him Rust. I looked over it. It wasn't great, but I'd say it wouldn't've been out of place in a high school class room.

I've learned so many languages, I rarely bother to learn up front. My current employer uses Ruby and Go, so I just open up the text editor, and read the existing code. Linting and unit tests are enough guard rails.

That said, I did pick up a book when I learned Haskel and Erlang. Those were sufficiently foreign that I couldn't just jump in. I'd probably need a book to re-learn Lisp and Prolog.
 
Upvote
3 (3 / 0)

clewis

Ars Tribunus Militum
1,837
Subscriptor++
My biggest issue is that it has sucked the fun out of programming for me. I used to enjoy the problem solving and satisfaction from creating things myself.

The productivity boost is too large to ignore with AI, but the end result is I feel most of my work is just code review now.
I found the opposite. I use it to do the easy 90% of the work, leaving the challenging 10% for me to finish.

But I do the same at work too. I only take the bugs and features that no one else is willing to touch. Or I assist when somebody else gets stuck. So I'm used to somebody else doing the easy part.
 
Upvote
2 (2 / 0)
High level languages are all very well, but the amount of bloat is somewhat maddening for a microcontroller hardware guy like me. [...] The project manager was touting "time to market" and "I'll take that risk", and pushed us to use Windows CE instead of other options.
I hear you, I feel the same way about people using Python for anything serious. BUT, using CE would have a distinct advantage (theoretically): hardware independence. Not that your boss would know hardware independence from a hole in the ground, of course.
Kids who are interested in learning programming at this moment in time, will have a huge advantage compared to the nerds from decades past. They will learn quicker. They will have so many more reps under their belt.
Absolute hogwash, they won't learn any of the fundamentals anymore! We're already seeing CS grads who don't know a pointer from Shinola, or don't know anything about data structures. It's a far cry from my days where a Data Structures course involved a lab coding everything from a linked list to a balanced B-Tree from scratch, and yes, with pointers; or a programming course where the introductory lab was "here's a recursive QuickSort, make it iterative". Those two courses caused about half of the students to leave; by year 2, EVERYONE could program. Now most CS graduates we see take a week to write 5 lines of Python. It's pathetic.

StackOverflow was bad enough... but LLM usage will give you graduates that can generate code, not graduates that can actually program. I think the problem a lot of people have is not knowing the difference.

No, it creates instant, infinite technical debt and legacy code. There is no one to explain why certain choices were made which is a definition of legacy code:
This. A zillion times this. It's the definition of being in "maintenance mode" from day one, which is clinically insane, and paralyzing.

I suppose you could argue that if you have very strong confidence in tests then maybe it doesn't matter so much what the actual code is as long as the tests pass
"Nobody understands the code that the AI wrote, but at least the tests pass and it conforms to all the code standards!" "Cool, who wrote the tests?" "The same AI,." "Oh, who is checking coding standards?" "The same AI." Nobody is going to fully code-review the fourth blanket code review the new intern told AI to do, not if it's a 20,000 line change, and there are only 1 or 2 people left who actually understand what the code does.

That's kind of the crux here, AI can help, and AI can work, but only if you understand what it's cranking out. "It would be clinically insane to have people use AI to generate code if they don't even understand what it is producing!" Yes, yes it would, but it would also be very, very cheap, so don't you DARE tell me that there will not be many a C-suite that will happily do it. The prototypes work, right? What's the problem? Do you even KNOW how expensive these snooty full-time senior devs are?

The "extra level of abstraction" analogy doesn't work here, you're not farming out how things are done, you're farming out WHAT is done (i.e. business logic). Mark my words, over the next decade or two there are going to be many people to die from this kind of application making life-critical decisions because of this; and if there is any justice, it will take the corporations that did it with them -- haha, as if.

This is exactly why these tools will not result in significant loss of work for software developers.
I don't think any of the senior professional devs here are saying that. There will still be plenty of work -- most likely, there will be even more work.

The consensus is that the work will be shittier. Instead of writing code, or reviewing code from juniors and raising them up to become seniors... we're now just reviewing garbage, and tuning prompts, lather, rinse, repeat. It is going to be soul-crushing, repetitive, thankless, and yet extremely demanding work.

We're not worried about a significant loss of WORK, we are worried about a significant loss of JOBS. Haha, what's the difference? Well, our WORK will get worse, and worse, and worse.
 
Upvote
14 (16 / -2)
There have been attempts at this, with experienced programmers, not a whole spectrum of levels. They didn't show that the LLMs actually helped.

But it's a difficult thing to really study. There isn't a true way to A/B someone doing the same task with and without a tool.
Which makes it kind of hilarious to have the near-universal judgement of experienced programmers, essentially being "yeah, it's neat for hobby shit, but it is utterly useless for the things I do" be dismissed by people who -- at best -- have programming as a long-term hobby.

Yes, we're old crusty bastards, but gosh darn dagnabbit, maybe we've seen a fad or two before.
 
Upvote
12 (13 / -1)
There are the practical costs (currently decoupled from any pricing scheme),
Oh, don't worry, that correction will come. The entire idea for programming LLMs is to get enough actual developers fired to have crucial businesses wholly dependent on the AI teat for developing new features, writing tests for it, doing the testing, and doing the deployment.

If you no longer have those pesky senior devs, senior DBAs, or any devops people, imagine how cheap you can run your R&D! We were wasting money all along on all those crusty bastards that kept asking uncomfortable questions in meetings anyway.
the environmental costs (both to people's local environments in proximity to datacenters and larger global ones due to increased consumption of carbon-rich fuels),
And why would they give a fuck about any of that?

and societal costs [all the horrific shit we already know, it's just too depressing to re-read]
And why would they give a fuck about any of that?

I used to think that AI would just get rid of junior positions and leave us senior devs to clean up all the mess, which is bad enough -- there is no developer alive that enjoys doing code reviews.

But the more I think about it, no, that is not what is going to happen. Once the C-suites sees ANYTHING AI-generated going into production and not immediately catch on fire, that's now the golden baseline. And suddenly, the old fogies are waahahaaaay too expensive, why are those assholes still around? I can hire 7 interns for the same price to generate that code,

So that means a very fast and massive exodus of the senior devs that can afford to do so, and remaining workforce of patent moron vibe coders and a handful of senior devs who cannot afford to retire yet. It'll go swimmingly.
 
Upvote
9 (9 / 0)

clewis

Ars Tribunus Militum
1,837
Subscriptor++
Earlier I decided there were two things I wanted to experiment with and learn more about and I combined them: learn Rust, and see what AI can do for me with code generation.

I had a test program I wanted that was kind of complicated: SFTP client and server, multiple worker threads passing jobs and synchronizing between each other, lots of configurability, a GUI interface, ... Definitely not the best choice for "my first Rust program", but I figured Claude could help me make it work. A test tool to automate tons of mundane repetitive tasks on the servers in our lab.

<snip>
Every time I've automated/controlled something with FTP, I've eventually replaced it with rsync -e ssh.

Although at some scale, stuff I've managed with rsync eventually got replaced with Configuration Management. I preferred Chef at the time, but I've heard good things about modern Ansible. Although I run everything as kubenetes pods now, so I just ship new container images and let AWS manage the nodes.
 
Upvote
0 (1 / -1)

clewis

Ars Tribunus Militum
1,837
Subscriptor++
As a soon to be software engineer graduate turn wannabe junior developer, hearing established developers say "It helps me a lot but I feel bad for junior developers" increasingly sounds similar to, "Fuck you, got mine".
We're not currently hiring, but I will be hiring juniors in the future. But we're in growth mode. We want a bunch of people being productive, not the fewest number of bodies managing a fixed quantity of work. The more productive they are, the better.
 
Upvote
-2 (0 / -2)
Every time I've automated/controlled something with FTP, I've eventually replaced it with rsync -e ssh.

Although at some scale, stuff I've managed with rsync eventually got replaced with Configuration Management. I preferred Chef at the time, but I've heard good things about modern Ansible. Although I run everything as kubenetes pods now, so I just ship new container images and let AWS manage the nodes.
Umm, neat?
 
Upvote
0 (0 / 0)
It's not like the "many" you are referring to have not seen attempts at major technological change that were smoke and mirrors. On the contrary, senior programmers have seen so many Next Big Things. We saw ActiveX, the Semantic Web, All Apps Will Be Java, All Apps Will Be Flash, All Apps Will Be Silverlight, blockchain, microservices, wearables, Segways, 3d TV, Web 2.0, IoT, Web3 (are we at the Web4 buzzword yet?), Big Data, NFTs... we saw them all.

The reality is that far, far more "new shiny bronze metals" are claimed to be the new disruptor but turn out to be a flash in the pan. We usually don't see the most disruptive tech trends until after it has already made the biggest changes.
I guess the next question is how will we have next new big things? These tools can only regurgitate what they have been trained on. Therefore, how do you ever get new languages? If developers are no longer the providers of problem solutions, but only debuggers of LLM output, you don't have people working on the next new thing. Even if some passionate savant creates a next new thing, how would you ever promulgate it in a world where all content must come from the pre-existing slop machine?
 
Upvote
3 (6 / -3)
In defense of Python, as an early adopter of it, the program logic and underlying math I use programming in it is the same as I’d use in C, Java, or any other language. The main advantage is that I have a bunch of data structures and libraries built-in rather than having to allocate the memory myself or use Boost libraries. It abstracts, but those abstractions are fully and clearly defined and I can delve into the details whenever necessary for debugging or improving performance. Or write the bottleneck code in something faster if necessary.

(There’s also a security advantage, in that not having to allocate memory myself stops me from screwing it up and giving someone access to the stack.)

The comparison with LLMs doesn’t hold, because the grammar of asking LLMs to write code is not only non-deterministic but fundamentally unexplainable. It’s a neural net with millions upon millions of weights. The designers of LLMs can’t even explain how a particular output comes about; they were swearing blind that they couldn’t regurgitate full pictures and books until a bunch of security researchers proved them wrong.
 
Last edited:
Upvote
8 (8 / 0)

overtoad

Seniorius Lurkius
41
The comment you're replying to is concerned less with AI coding as "additional abstraction" and referring to how untrained individuals can make apparent progress with tasking they're generally incompetent with.

It is not accurate to compare AI prompting to another layer of abstraction because unlike high level programming abstraction, AI prompting is neither a well documented or reliably interpreted methodology. Even programming languages that do not cleanly map to processor instructions, such as functional programming languages and weak-typing like Python has, have hard rules that form a specific, learnable system with predictable behavior.

It's something that a trained software engineer can sign-off on because they can directly study the known limitations of the abstract layer and how it works. This isn't true for AI generated code. You can't sign-off on the capabilities of AI generated code by looking at the prompt and performing simple tests on the compiled code.* A skilled software engineer who understands how a requirement needs to be met can prompt an AI, look at the output and verify the AI took the right approach without producing side effects; but an unskilled 'prompt-engineer' that's only familiar with coaxing answers out an AI can't because they won't know how a problem should be solved. For normal layers of abstraction that a software-engineer doesn't fully understand the implementation of, there is documentation to be read regarding the scalability, effects, and resultant mechanics of the high level layer. The 'prompt-engineer' does not have reliable writings and documentation to study, only syntax that accomplishes a task they don't actually understand.

* In theory, you can come up with and run a set of tests to guarantee that the code works. In practice, this requires an incredibly thorough understanding of your requirements, something the 'prompt-engineer' doesn't have.

** Throughout all of this, my definition of 'prompt-engineer' excludes everyone who actually has the expertise to perform the tasking at hand without the use of generative AI. For a skilled software engineer who knows what they're doing, an AI can just be another way of arriving at the syntax they need. The original post is referring to the hiring of unskilled AI-prompters as a cheap alternative to proper engineers.
say you weren't around at the beginning of the web without saying...
 
Upvote
-7 (0 / -7)

overtoad

Seniorius Lurkius
41
In defense of Python, as an early adopter of it, the program logic and underlying math I use programming in it is the same as I’d use in C, Java, or any other language. The main advantage is that I have a bunch of data structures and libraries built-in rather than having to allocate the memory myself or use Boost libraries. It abstracts, but those abstractions are fully and clearly defined and I can delve into the details whenever necessary for debugging or improving performance. Or write the bottleneck code in something faster if necessary.

(There’s also a security advantage, in that not having to allocate memory myself stops me from screwing it up and giving someone access to the stack.)

The comparison with LLMs doesn’t hold, because the grammar of asking LLMs to write code is not only non-deterministic but fundamentally unexplainable. It’s a neutral net with millions upon millions of weights. The designers of LLMs can’t even explain how a particular output comes about; they were swearing blind that they couldn’t regurgitate full pictures and books until a bunch of security researchers proved them wrong.
1. grammar asking for code to be written is nondeterministic
2 grammar to convey intent is fundamentally unexplainable
3. intent being conveyed to a neural network (non-artificial) with millions upon (some very large number of) millions of weights.
4. large numbers of neuroscience research careers into (biological) neural networks can't explain how a particular output comes about
-and a bonus-
5. asking the (biological) neural network why it produces an output generates, at best, a post-hoc, self serving narrative created out of whole cloth.

yep! sounds exactly like the process of getting squishy humans to write code
 
Upvote
-5 (1 / -6)

the cave troll

Ars Scholae Palatinae
1,270
Subscriptor++
Nice article. This really matched my experience.

I have almost zero formal coding background, but I’ve used AI to hack together a few small apps purely as a learning exercise. I’m not trying to ship anything or sell it, which probably changes the equation quite a bit.

Reading both the article and the comments, it’s clear how much background matters. I can see why experienced engineers find AI coding frustrating. Brittle output, strange mistakes, and the last 10% still being very real work. If you care about clean, maintainable code, that would get old fast.

But for me, coming at this with a "beginner’s mind," though, it’s been a lot of fun. I expect things to break. When they do, I adjust the prompt, learn a little more, and keep going. No pressure, no deadlines, no illusion that the AI actually “knows” anything.

For learning and exploration, it’s been a surprisingly enjoyable ride. Bugs, hallucinations, and all.

So, this is not a criticism of you using this as a fun hobby, but I feel that there is an important misconception that needs to be cleared up. It might seem that the difference between a novice programmer and an expert programmer is just that the latter can generate better code more quickly than the former--that is, that the latter is doing essentially the same general kind of thing as the former, but with the benefit of more experience. However, what really sets the expert apart from the novice is their mindset. Whereas the novice thinks of programming as a way of telling the computer what to do, the expert thinks of programming as the art of defining the right vocabulary with which to describe the problem that needs to be solved, followed by writing the solution using that vocabulary and/or making that vocabulary available to others to save them time. Put another way, programming fundamentally involves taking complicated ideas and binding them into words that can be used by speaking them, with programming languages supplying the syntax.

Anyway, I am not telling you this to convince you to stop what you are doing, but just because I see a lot of misconceptions like these seem to spring up so I felt the need to say something to you and the wider audience.
 
Upvote
7 (7 / 0)
"This opinion comes not as an endorsement but as personal experience"

Yeah, but this isn't your personal blog, though. This is just like those youtubers that go "This is not advertisement because nobody's paying me!" Ah so it's advertising you do for free, cool, cool.

Also: Not ONE mention of ethical concerns. Not ONE. Very, very interesting.
I guess some people are always on the cross.

Have you never heard of an opinion piece in news media? That said, I strongly suspect Benj is being directed by Conde Nast(or, well, someone) to write these articles, and they do really seem to be more of an advertisement more often than not. However, in an opinion piece, I'm fine accepting large amounts of bias.

In a standard news article, I am not OK with that same level of bias. And this is not a standard news article. So to me, it is fine, as I know what I'm getting into by reading it.

And why would ethics (a valid concern) even come into play here, considering this is an opinion piece?
 
Upvote
-9 (0 / -9)
Quote
pokrface
pokrface
That said, I strongly suspect Benj is being directed by Conde Nast(or, well, someone) to write these articles...

Hey: FYI, you're absolutely dead-ass wrong.

—Benj's manager + one of the people who edited this piece
Upvote
-9 (0 / -9)

hillspuck

Ars Scholae Palatinae
2,179
That reminds me of my boss ~10 years ago who wanted to know how we could add HTML5 to our product. Our 100% locally-installed WinForms/WIN32 product.
This is the problem at the root of so many industries. Bosses that got to be boss by talent playing the boss game, not by talent doing the actual thing the company does. They are constantly looking for New Initiatives, because that's how that class keeps score. You get enough initiatives completed and you level up. Doesn't matter if those initiatives were really needed or are successful in the long term. In fact, it's fine if they're not, because the next boss or two that comes along will need to start an initiative to rip out the old thing the previous bosses added and replace it with their own great new initiative. It's a cycle that just repeats in so many businesses.

I found the opposite. I use it to do the easy 90% of the work, leaving the challenging 10% for me to finish.

But I do the same at work too. I only take the bugs and features that no one else is willing to touch. Or I assist when somebody else gets stuck. So I'm used to somebody else doing the easy part.
While the first 90% is easy in many ways, in others it's actually quite hard. It's laying the foundations for that last 10%, and if the foundations are built poorly that last 10% is going to be much worse. It's an easy job for those that have the expertise to do it. I think AI has not been proven to have that expertise, just as many junior coders have not. That's why even though it's the easy part, it's important to have experience and skill involved at that stage.
 
Upvote
10 (10 / 0)

dtich

Ars Scholae Palatinae
1,401
Subscriptor
Literally everyone saying what you are all say the same thing:

"I'm a hobbyist programmer and it's great for me, so that means it's going to be great for everyone"

Very narrow-minded approach.

Also, why the heck do you AIbros all talk the same? Do you also use the LLM to construct your replies?
Literally? Everyone?? Hm.

I never said "it's going to be great for everyone", nor do I think that. I said more nuanced and thoughtful things, though. But I guess you think I sound like an 'AIbro'... which I am not. Nor do I think my thoughts were 'narrow-minded'; while they may have indeed been from my perspective, I never claimed to speak for everyone, and I can't honestly do much about that--my opinion is from my perspective, it's tautological, for any one of us. But, you don't know me, clearly. Your post really added nothing to the conversation and I can't say much more for this response. And, no, I didn't use 'the LLM', I used my brain to construct this reply, because it still works very well. Can't say the same for all of us.
 
Upvote
0 (1 / -1)

nhck

Seniorius Lurkius
16
I think Python is a good analogy. The language came along and allowed all kinds of people without a software engineering background to write code that solved a problem for them, significantly lowering the barrier to entry.

The thing about python that lowered the barrier to entry was the intense rush of community support it got when it first came on the scene, and the modules that were bundled with it. Not necessarily the ease of the language itself. Comparatively speaking, Python is like a lot of other languages.. you dont really need a degree or engineering background to write apps with any language. It's just about the amount of time and effort you want to invest in doing so. Having all these modules available right away (in any language) saves you from so much of that time and effort. Bottom line, it's the support, imo.
 
Upvote
6 (6 / 0)

graylshaped

Ars Legatus Legionis
68,279
Subscriptor++
Upvote
5 (5 / 0)

John Abbe

Smack-Fu Master, in training
53
I think Python is a good analogy.

Also, since you didn't mention it, a good language for working with LLMs, especially as Mojo project progresses. It's a superset of Python do compatible, but it compiles and is built from the ground up to work with different kinds of processors - CPUs, GPUs, TPUs, etc. https://www.modular.com/mojo
 
Upvote
0 (0 / 0)

Aurich

Director of Many Things
41,280
Ars Staff
I guess some people are always on the cross.

Have you never heard of an opinion piece in news media? That said, I strongly suspect Benj is being directed by Conde Nast(or, well, someone) to write these articles, and they do really seem to be more of an advertisement more often than not. However, in an opinion piece, I'm fine accepting large amounts of bias.

In a standard news article, I am not OK with that same level of bias. And this is not a standard news article. So to me, it is fine, as I know what I'm getting into by reading it.

And why would ethics (a valid concern) even come into play here, considering this is an opinion piece?
Ah yes, the conspiracy theory of (checks notes) the Ars Technica AI reporter being told he should write some stories about using AI. 😂

How do people think this all works? Like the daily flow of news you read? Is it a grand conspiracy every time a writer is assigned a story to you? Or when a writer here states an opinion on something?

This is the gig. Benj was hired to write about AI, that's what he does. Sometimes he has his own story ideas based on the daily flow of news, sometimes an editor says "hey you should hit the news about Apple going with Gemini" or whatever, and sometimes he plans out a more long form feature like "what if I did a piece on trying AI code for a month?"

The features are pitched internally, they're edited, it's a normal thing all the writers do. Writing about personal experiences is a great way to round out hitting daily news and providing depth and interest you can't get from another outlet.

The only thing that's weird here is that people are even trying to make a thing out of the most basic daily operations.
 
Upvote
20 (20 / 0)

hillspuck

Ars Scholae Palatinae
2,179
The features are pitched internally, they're edited, it's a normal thing all the writers do. Writing about personal experiences is a great way to round out hitting daily news and providing depth and interest you can't get from another outlet.

The only thing that's weird here is that people are even trying to make a thing out of the most basic daily operations.
Personally, my "conspiracy theory" is that Conde Nast often forgets Ars even exists. This is in no way to diminish Ars' excellent work (especially since it's my #1 go to website for the things it covers), but they seem like one of the smaller fishes in CN's pond. They have like a half billion customers worldwide. Ars seems like a few drops in the bucket on that scale, and one that is not recognizable in mainstream culture.
 
Upvote
7 (7 / 0)

faux eel

Smack-Fu Master, in training
5
Subscriptor
I really hope you're correct, but my thought process has been that if it allows people to be more productive, there will typically be layoffs. If a company has 1000 developers and they find the tool allows each of them to even be 15% more productive, they can easily justify 100+ people being let go. They might find more work for them, but in my experience when they can save money, they will.
Those people will hopefully work on niches and nooks that were previously deemed too labor intensive for custom software. Ideally we are moving to more fit for purpose (small scale) software, since (arguably) programming just got cheaper.
 
Upvote
0 (1 / -1)

Idiotzoo

Smack-Fu Master, in training
78
I keep hearing “these tools are not going away” but is that true? Certainly in their current form companies are burning billions to provide these tools. what happens when that inevitably comes to an end? Perhaps locally run, lower capability LLMs could be useful for programmers. Maybe the real cost of the tools is worth paying, but I somehow doubt that. Ultimately, as I understand things today, the only inevitability is the money is going to run out and the affordable availability of these ai tools will end.
 
Upvote
10 (11 / -1)

pokrface

Senior Technology Editor
21,543
Ars Staff
Personally, my "conspiracy theory" is that Conde Nast often forgets Ars even exists.
Out of all the conde nast-related conspiracy theories I have seen posted to the Ars comment section since the acquisition in 2008, this one is closest to reality, lol. Generally the only time anyone at Ars who isn't a senior manager communicates with anyone at conde is when they need to file an expense report or during benefits open enrollment in November.
 
Upvote
8 (8 / 0)
I spent some fun time with my youngest kid playing Violent Checkers. This also reminds me I need to polish my Tron game with super abilities. I guess it's a good opportunity to give another chance to these coding agents.
Really glad you enjoyed playing it! Of all the game ideas I cooked up, I find myself playing that one the most.

One thing I forgot to mention in the article is I have a spiral-bound notebook of game design ideas going back about 20 years, so I had a lot of concepts sitting around to try out with these coding agents. It's really fun seeing those ideas finally come to life and then refining them to make them fun to play.
 
Upvote
3 (3 / 0)
Ah yes, the conspiracy theory of (checks notes) the Ars Technica AI reporter being told he should write some stories about using AI. 😂

How do people think this all works? Like the daily flow of news you read? Is it a grand conspiracy every time a writer is assigned a story to you? Or when a writer here states an opinion on something?

This is the gig. Benj was hired to write about AI, that's what he does. Sometimes he has his own story ideas based on the daily flow of news, sometimes an editor says "hey you should hit the news about Apple going with Gemini" or whatever, and sometimes he plans out a more long form feature like "what if I did a piece on trying AI code for a month?"

The features are pitched internally, they're edited, it's a normal thing all the writers do. Writing about personal experiences is a great way to round out hitting daily news and providing depth and interest you can't get from another outlet.

The only thing that's weird here is that people are even trying to make a thing out of the most basic daily operations.
Yes indeed! Believe me, nobody told me to get sick with COVID for a month and go nutty making non-stop game demos using AI tools for 6-8 hours a day during my winter vacation. That was entirely my own madness. :) It was fun but also mentally exhausting, as documented in the article. The tools aren't perfect, and they perform exactly as stated, including all the drawbacks and limitations.

To everyone who has reached out and enjoyed the piece, I appreciate the kind words and positive feedback.
 
Upvote
11 (11 / 0)

RammyBodger

Ars Centurion
340
Subscriptor
However, I don't think it's unreasonable to argue that models trained on GPLv3 code must be GPLv3 themselves.
Thank you for saying this. It has never made sense to me either that an LLM trained on GPLv3 code and which can be tricked into producing GPLv3 output fairly simply can produce code which is anything other than GPLv3 licensed. It is very much a derivative work.
 
Upvote
8 (8 / 0)

Pooga

Ars Scholae Palatinae
1,349
Subscriptor++
1. grammar asking for code to be written is nondeterministic
2 grammar to convey intent is fundamentally unexplainable
3. intent being conveyed to a neural network (non-artificial) with millions upon (some very large number of) millions of weights.
4. large numbers of neuroscience research careers into (biological) neural networks can't explain how a particular output comes about
-and a bonus-
5. asking the (biological) neural network why it produces an output generates, at best, a post-hoc, self serving narrative created out of whole cloth.

yep! sounds exactly like the process of getting squishy humans to write code
You're not wrong.

It's exactly like the process of getting that one guy who skated through his classes by copy his classmates' work without ever really comprehending the fundamental concepts of what he was copying to write code. That guy has seen enough examples that for most relatively simple assignments he can put together something that compiles and does something like what it is supposed to do. He struggles much more when he tries to synthesize something that requires combining multiple examples into a hybrid solution with elements of each.

He's liable to write something with fundamental syntax errors he doesn't recognize as errors because he never really went beyond pattern recognition in his understanding of what he was "learning".

If asked for something that was never covered in any of the classes he took, he has no ability to extrapolate from core principles or a general understanding of basic elements of programming that apply across languages because he lacks any ability to understand or expand his mental model beyond what he got in his initial training.

I've worked with devs that aren't 100% dissimilar. They learned just enough to get hired, then are quickly in over their heads and leaning heavily on the senior devs to basically do their work for them. Sometimes it's just a learning curve regarding the differences between classroom programming and applying that in real world scenarios. Those guys will usually get their feet under them and go one to become better over time. Some just never really internalized the basics and no amount of support will get them beyond doing simple, basic tasks to let the real devs tackle the core issues - assuming they can do that well enough to be worth keeping on. If the actual devs have to spend too much time correcting the work of that guy, he won't be in that position for long.

I guess one saving grace of having an AI agent in this position instead of a bad junior dev is that's it's unlikely the AI can schmooze their way into a management track where they can call the shots on tasks they have no understanding about...
 
Upvote
2 (2 / 0)

DCStone

Ars Tribunus Militum
2,822
That reminds me of my boss ~10 years ago who wanted to know how we could add HTML5 to our product. Our 100% locally-installed WinForms/WIN32 product.
I hope you told them "Sure thing!" and then slapped the HTML5 logo in a corner of an obscure splash-screen somewhere. I mean, technically that would be adding HTML5 to your product!
 
Upvote
1 (1 / 0)
I hear and understand all the good things this has brought to the author's experience coding. The speed, the filling in of his own knowlege gaps in the course of building new software.

For me, the problem is that this is a "tool" that you "speak" into, look at the results, then speak into again with slightly different words, then examine the results to see if the problem you saw was fixed. Rinse and repeat.

It's a black box.

You don't know the inner workings and you don't have direct control of the tool. You're essentially doing "trial and error until success," even if it's only on that "last 10%" part.

In another field with which I'm involved, there is a saying: "The first 90% of the job consumes the first 90% of the time involved, while the last 10% of the job consumes the second 90% of the time involved."

That seems like it may well apply to AI-assisted programing, too.

I, personally, want more deterministic control of my tools. And I want to know how they work, so that I can better understand how to use them. This is an "IMHO," not a damnation of the use of AI in coding.
 
Upvote
1 (1 / 0)