AI Coding Is Massively Overhyped, Report Finds
AI Coding Is Massively Overhyped, Report Finds
"No Duh," say senior developers everywhere.
The article explains that vibe code often is close, but not quite, functional, requiring developers to go in and find where the problems are - resulting in a net slowdown of development rather than productivity gains.
About that "net slowdown". I think it's true, but only in specific cases. If the user already knows well how to write code, an LLM might be only marginally useful or even useless.
However, there are ways to make it useful, but it requires specific circumstances. For example, you can't be bothered to write a simple loop, you can use and LLM to do it. Give the boring routine to an LLM, and you can focus on naming the variables in a fitting way or adjusting the finer details to your liking.
Can't be bothered to look up the exact syntax for a function you use only twice a year? Let and LLM handle that, and tweak the details. Now, you didn't spend 15 minutes reading stack overflow posts that don't answer the exact question you had in mind. Instead, you spent 5 minutes on the whole thing, and that includes the tweaking and troubleshooting parts.
If you have zero programming experience, you can use an LLM to write some code for you, but prepare to spend the whole day troubleshooting something that is essentially a black box to you. Alternatively, you could ask a human to write the same thing in 5-15 minutes depending on the method they choose.
You mean relying blindly on a statistical prediction engine to attempt to produce sophisticated software without any understanding of the underlying principles or concepts doesn't magically replace years of actual study and real-world experience?
But trust me, bro, the singularity is imminent, LLMs are the future of human evolution, true AGI is nigh!
I can't wait for this idiotic "AI" bubble to burst.
so is the profit it was foretold to generate, but it actually costs money than its actually generating.
This article sums up a Stanford study of AI and developer productivity. TL;DR - net productivity boost is a modest 15-20%, or as low as negative to 10% in complex, brownfield codebases. This tracks with my own experience as a dev.
Senior Management in much of Corporate America is like a kind of modern Nobility in which looking and sounding the part is more important than strong competence in the field. It's why buzzwords catch like wildfire.
Lmao calling nobility would imply we can't vote our senior management and it often ends up being whoever "the king" wants or one of king's children.
Wait...
Are you trying to tell me that the people wanting to sell me their universal panacea for all human endeavours were... lying...? Say it ain't so.
I mean, originally they thought they had come upon a magic bullet. Turns out it wasn't the case, and now they're going to suffer for it.
You're assuming honesty and they've earned the opposite posture.
I'm not a programmer in any sense. Recently, I made a project where I used python and raspberry pi and had to train some small models on a KITTI data set. I used AI to write the broad structure of the code, but in the end, it took me a lot of time going through python documentation as well as the documentation of the specific tools/modules I used to actually get the code working. Would an experienced programmer get the same work done in an afternoon? Probably. But the code AI output still had a lot of flaws. Someone who knows more than me would probably input better prompts and better follow up requirements and probably get a better structure from the AI, but I doubt they'll get a complete code. In the end, even to use AI, you have to know what you're doing to use AI efficiently and you still have to polish the code into something that actually works.
From my experience, AI just seems to be a lesson in overfitment. You can't use it to do things nobody has done before. Furthermore, you only really get good responses from prompts related to Javascript
These types of articles always fail to mention how well trained the developers were on techniques and tools. In my experience that makes a big difference.
My employer mandates we use AI and provides us with any model, IDE, service we ask for. But where it falls short is providing training or direction on ways to use it. Most developers seem to go for results prompting and get a terrible experience.
I on the other hand provide a lot of context through documents and various mcp tooling, I talk about the existing patterns in the codebase and provide sources to other repositories as examples, then we come up with an implementation plan and execute on it with a task log to stay on track. I spend very little time fixing bad code because I spent the setup time nailing down context.
So if a developer is just prompting "Do XYZ". It's no wonder they're spending more time untangling a random mess.
Another aspect is that everyone seems to always be working under the gun and they just don't have the time to figure out all the best practices and techniques on their own.
I think this should be considered when we hear things like this.
I have 3 questions, and I'm coming from a heavily AI-skeptic position, but am open:
I see a place for AI-generated code, for instant functions that do something blending simple and complex. "Hey claude, write a function to take a string and split it at the end of every sentence containing an uppercase A". I had to write weird functions like that constantly as a sysadmin, and transforming data seems like a thing an AI could help me accelerate. I just don't see that working on a larger scale, though, or trusting an AI enough to allow it to integrate a new function like that into an existing codebase.
Thank you for reading my comment. I'm on the train headed to work and I'll try to answer completely. I love talking about this stuff.
1.
Do you believe that providing all that context, describing the existing patterns, creating an implementation plan, etc, allows the AI to both write better code and faster than if you just did it yourself?
For my work, absolutely. My work is a lot of tickets that were setup from multiple stories and multiple epics. It would be like asking me if I am really framing a house faster with a nail gun and compressor. If I were just hanging up a picture or a few pictures in the hallway, it's probably faster to use a hammer than to set up the compressor and nail gun, plus cleanup.
However, a lot of that documentation already exists by the time it gets to me. All of the Design Documents and Product Requirement Documents have already been formed, discussed, and approved by our architecture team and team leads. Imagine if you already had this documentation for the asteroid game; how much better do you think your LLM would do? Maybe this is the benefit of using LLMs for development at an established company. Btw, a lot of those Documents were also created with the assistance of AI by the Product Team, Architects, and Principle/Staff/Leads anyway.
how much twiddling of your existing codebase do you need to do before an AI can understand the business context of it?
With the help of our existing documents and codebase(s) I feel I dont have any issues with the model knowing what we're doing. I do have to set up my own context for how I want it to be done. To me this is like explaining to a Junior Engineer what I need them to help me with. If you're familiar with "Know when to Direct, when to Delegate, or when to Develop" I would say it lands in between Direct and Delegate. I have markdown files with my rules and guidelines and provide that as context. I use Augment Code which is pretty good with codebase context.
write an Asteroids clone using JavaScript and HTML 5 Canvas
I would try "Let's plan out the steps needed to write an Asteroids game using JavaScript and HTML 5. Identify and explain each step of the development plan. The game must build with no errors, be playable, and pass all tests. Do not wrote code at this time until our plan is approved" Then once it comes back with the initial steps, I would guide it further if needed. Finally I would approve the plan and tell it to execute while tracking it's steps (Augment Code uses a task log).
Are you running your models locally, or do you have some cloud service? If you give your entire codebase to a 3rd party as context, how much of your company's secret sauce have you disclosed?
We are required to use the frontier models that my employer has contracts with and are forbidden from using local models. In our enterprise contracts we have negotiated for no training on our data. I imagine we pay for that. I'm not involved in that level of interaction on the accounts.
How much pen-testing time are you spending on this code, error handling, edge cases, race conditions, data sanitation? An experienced dev understands these things innately, having fixed these kinds of issues in the past and knows the anti-patterns and how to avoid them. In all seriousness, I think this is going to be the thing that actually kills AI vibe coding, but it won't be fast enough. There will be tons of new exploits in what used to be solidly safe places.
We have other teams that handle a lot of these tasks. These teams are also using AI tools to get the job done. In addition, we have static testing tools on our repo like CodeRabbit and another one I can't remember the name of that looks specifically for security concerns. It will comment on the PR directly and our merge would be blocked until handled. Code coverage for testing is at 85% or it blocks the merge and we have a full QA department of Analysts and SDETs to QA. In addition to that we still have human approvals required (2 devs + Sr+). All of these people involved are still using AI tools to help them in each step.
I hope that answers your questions and gives you some insight into how I've found success in my experience with it. I will say that on my personal projects I don't go this far with process and I don't experience the same AI output that I do at work.
I work adjacent to software developers, and I have been hearing a lot of the same sentiments. What I don't understand, though, is the magnitude of this bubble then.
Typically, bubbles seem to form around some new market phenomenon or technology that threatens to upset the old paradigm and usher in a new boom. Those market phenomena then eventually take their place in the world based on their real value, which is nowhere near the level of the hype, but still substantial.
In this case, I am struggling to find examples of the real benefits of a lot of these AI assistant technologies. I know that there are a lot of successes in the AI realm, but not a single one I know of involves an LLM.
So, I guess my question is, "What specific LLM tools are generating profits or productivity at a substantial level well exceeding their operating costs?" If there really are none, or if the gains are only incremental, then my question becomes an incredulous, "Is this biggest in history tech bubble really composed entirely of unfounded hype?"
I think right now companies are competing until they're only 1 or 2 that clearly own the majority of the market.
Afterwards they will devolve back into the same thing search engines are now. A cesspool of sponsored ads and links to useless SEO blogs.
They'll just become gate keepers of information again and the only ones that will be heard are the ones who pay a fee or game the system.
Maybe not though, I'm usually pretty cynical when it comes to what the incentives of businesses are.
From what I've seen and heard, there are a few factors to this.
One is that the tech industry right now is built on venture capital. In order to survive, they need to act like they're at the forefront of the Next Big Thing in order to keep bringing investment money in.
Another is that LLMs are uniquely suited to extending the honeymoon period.
The initial impression you get from an LLM chatbot is significant. This is a chatbot that actually talks like a person. A VC mogul sitting down to have a conversation with ChatGPT, when it was new, was a mind-blowing experience. This is a computer program that, at first blush, appears to be able to do most things humans can do, as long as those things primarily consist of reading things and typing things out - which a VC, and mid/upper management, does a lot of. This gives the impression that AI is capable of automating a lot of things that previously needed a live, thinking person - which means a lot of savings for companies who can shed expensive knowledge workers.
The problem is that the limits of LLMs are STILL poorly understood by most people. Despite constructing huge data centers and gobbling up vast amounts of electricity, LLMs still are bad at actually being reliable. This makes LLMs worse at practically any knowledge work than the lowest, greenest intern - because at least the intern can be taught to say they don't know something instead of feeding you BS.
It was also assumed that bigger, hungrier LLMs would provide better results. Although they do, the gains are getting harder and harder to reach. There needs to be an efficiency breakthrough (and a training breakthrough) before the wonderful world of AI can actually come to pass because as it stands, prompts are still getting more expensive to run for higher-quality results. It took a while to make that discovery, so the hype train was able to continue to build steam for the last couple years.
Now, tech companies are doing their level best to hide these shortcomings from their customers (and possibly even themselves). The longer they keep the wool over everyone's eyes, the more money continues to roll in. So, the bubble keeps building.
The upshot of this and a lot of the other replies I see here and elsewhere seem to suggest that one big difference between this bubble and other past ones is that with this most recent one, there is so much of the global economy now tied to the fate of this bubble that the entire financial world is colluding to delay the inevitable due to the expected severity of the consequences.
This struck upon one of the greatest wishes of all corporations. A way to get work without having to pay people for it.
AI is a financial scam. Basically companies that are already mature promise great future profits thanks to this new technological miracle, which makes their stock more valuable than it otherwise would be. Cory Doctorow has written eloquently about this.
When the AI bubble bursts, even janitors and nurses will lose their jobs. Financial institutions will go bust.
I have never seen an AI generated code which is correct. Not once. I've certainly seen it broadly correct and used it for the gist of something. But normally it fucks something up - imports, dependencies, logic, API calls, or a combination of all them.
I sure as hell wouldn't trust to use it without reviewing it thoroughly. And anyone stupid enough to use it blindly through "vibe" programming deserves everything they get. And most likely that will be a massive bill and code which is horribly broken in some serious and subtle way.
for me it typically don't cause syntax errors, but the main thing it fucks up is what you specifically told them to do, where the output straight up does not perform the way your specification requires. If it's just some syntax errors at least the compiler can catch them, this you won't even know if you don't bother testing the output.
I've used Claude code to fix some bugs and add some new features to some of my old, small programs and websites. Not things I can't do myself, but things I can't be arsed to sit down and actually do.
It's actually gone really well, with clean and solid code. easily readable, correct, with error handling and even comments explaining things. It even took a gui stream processing program I had and wrote a server / webapp with the same functionality, and was able to extend it with a few new features I've been thinking to add.
These are not complex things, but a few of them were 20+ files big, and it manage to not only navigate the code, but understand it well enough to add features with the changes touching multiple files (model, logic, view layer for example, or refactor a too big class and update all references to use the new classes).
So it's absolutely useful and capable of writing good code.
This is the truth. It has tremendous value but it isn't a solution -- it's a tool. And if you don't know how to code or what good code looks like, then it is a tool you can't use!
would you deploy this server?
I've seen and used AI for snippets of code and it's pretty decent at that.
With my colleagues I always compare it to a battery powered drill. It's very powerful and can make shit a lot easier. But you'd not try to build furniture from scratch with only a battery powered drill.
You need the knowledge to use it - and also saws, screws, the proper bits for those screws and so on and so forth.
What bothers me the most is the amount of tech debt it adds by using outdated approaches.
For example, recently I used AI to create some python scripts that use polars and altair to parse some data and draw charts. It kept insisting to bring in pandas so it could convert the polars dataframes to pandas dataframes just for passing them to altair. When I told if that altair can use polars dataframes directly, that helped, but two or three prompts later it would try to solve problems by adding the conversion again.
This makes sense too, because the training material, on average, is probably older than the change that enabled altair to use polars dataframes directly. And a lot of code out there just only uses pandas in the first place.
The result is that in all these cases, someone who doesn’t know this would probably be impressed that the scripts worked, and just not notice the extra tech debt from that unnecessary dependency on pandas.
It sounds like it’s not a big deal, but these things add up and eventually, our AI enhanced code bases will be full of additional dependencies, deprecated APIs, unnecessarily verbose or complicated code, etc.
I feel like this is one aspect that gets overlooked a bit when we talk about productivity gains. We don’t necessarily immediately realize how much of that extra LoC/time goes into outdated code and old fashioned verbosity. But it will eventually come back to bite us.
How is it not correct if the code successfully does the very thing that was prompted?
F.ex. in my company we don’t have any real programmers but have built handful of useful tools (approx. 400-1600 LOC, mainly Python) to do some data analysis, regex stuff to cleanup some output files, index some files and analyze/check their contents for certain mistakes, dashboards to display certain data, etc.
Of course the apps may not have been perfect after the very first prompt, or even compiled, but after iterating an error or two, and explaining an edge case or two, they’ve started to perform flawlessly, saving tons of work hours per week. So how is this not useful? If the code creates results that are correct, doesn’t that make the app itself technically ”correct” too, albeit likely not nearly as optimized as equivalent human code would be.
To add on to what others have said, vibe coding is ushering in a new golden age for black hat hackers. If someone is rely entirely on AI to generate code they likely don't understand what the code they have is actually doing. This tends to lead to an app that works correctly for what the prompted specified but behaves badly the instant it has to handle anything outside of the prompt, like a malformed request or data outside the prompted parameters. As a result these apps tend to be easy to exploit by malicious actors, often in ways the original prompter never thought of.
If the code doesn't compile, or is badly mangled, or uses the wrong APIs / imports or forgets something really important then it's broken. I can use AI to inform my opinion and sometimes makes use of what it outputs but critically I know how to program and I know how to spot good and bad code.
I can't speak for how you use it, but if you don't have any real programmers and you're iterating until something works then you could be producing junk and not know it. Maybe it doesn't matter in your case if its a bunch for throwaway scripts and helpers but if you have actual code in production where money, lives, reputation, safety or security are at risk then it absolutely does.
It's not bad for your use case but going beyond that without issues and actual developpers to fix the vibe code is not yet possible for llms
Perhaps it should read "All AI is over hyped, over done and we should be over it"
I code with LLMs every day as a senior developer but agents are mostly a big lie. LLMs are great for information index and rubber duck chats which already is incredible feaute of the century but agents are fundamentally bad. Even for Python they are intern-level bad. I was just trying the new Claude and instead of using Python's pathlib.Path it reinvented its own file system path utils and pathlib is not even some new Python feature - it has been de facto way to manage paths for at least 3 years now.
That being said when prompted in great detail with exact instructions agents can be useful but thats not what being sold here.
After so many iterations it seems like agents need a fundamental breakthrough in AI tech is still needed as diminishing returns is going hard now.
Oh yes. The Great pathlib. The Blessed pathlib. Hallowed be it and all it does.
I'm a Ruby girl. A couple of years ago I was super worried about my decision to finally start learning Python seriously. But once I ran into pathlib, I knew for sure that everything will be fine. Take an everyday headache problem. Solve it forever. Boom. This is how standard libraries should be designed.
I disagree. Take a routine problem and invent a new language for it. Then split it into various incompatible dialects, and make sure in all cases it requires computing power that no one really has.
Pathlib is very nice indeed, but I can understand why a lot of languages don't do similar things. There are major challenges implementing something like that. Cross-platform functionality is a big one, for example. File permissions between Unix systems and Windows do not map perfectly from one system to another which can be a maintenance burden.
But I do agree. As a user, it feels great to have. And yes, also in general, the things Python does with its standard library are definitely the way things should be done, from a user's point of view at least.
If it wasn't for all the AI hype that it's going to do everyone's job, LLMs would be widely considered an amazing advancement in computer-human interaction and human assistance. They are so much better than using a search engine to parse web forums and stack overflow, but that's not going to pay for investing hundreds of billions into building them out. My experience is like yours - I use AI chat as a huge information index mainly, and helpful sounding board occasionally, but it isn't much good beyond that.
I will concur with the whole 'llm keeps suggesting to reinvent the wheel'
And poorly. Not only did it not use a pretty basic standard library to do something, it's implementation is generally crap. For example it offered up a solution that was hard coded to IPv4, and the context was very ipv6 heavy
Even though this shit was apparent from day fucking 1, at least the Tech Billionaires were able to cause mass layoffs, destroy an entire generation of new programmers' careers, introduce an endless amount of tech debt and security vulnerabilities, all while grifting investors/businesses and making billions off of all of it.
Sad excuses for sacks of shit, all of them.
Look on the bright side, in a couple of years they will come crawling back to us, desperate for new things to be built so their profit machines keep profiting.
Current ML techniques literally cannot replace developers for anything but the most rudimentary of tasks.
I wish we had true apprenticeships out there for development and other tech roles.
So when the AI bubble burst, will there be coding jobs available to clean up the mess?
There already are. People all over LinkedIn are changing their titles to "AI Code Cleanup Specialist".
I mean largely for most of us I hope. But I feel like the tech sector was oversatured because of all the hype of it being an easy get rich quick job. Which for some people it was.
AI coding is the stupidest thing I've seen since someone decided it was a good idea to measure the code by the amount of lines written.
It did solve my impostor syndrome though. Turns out a bunch of people I saw to be my betters were faking it all along.
More code is better, obviously! Why else would a website to see a restaurant menu be 80Mb? It's all that good, excellent code.
According to Deutsche Bank the AI bubble is a the pillar of our economy now.
So when it pops. I guess that's kinda apocalyptic.
Edit - strikethrough
Only for taxpayers ☝️
The people talking about AI coding the most at my job are architects and it drives me insane.
Software architects that don't write code are worse than useless
I am a software architect, an mainly usw it to refactor my own old code... But i am maybe not a typical architect...
I don't really care if people use it, it's more that it feels like a quarter of our architect meeting presentations are about something AI related. It's just exhausting.
I mean.. At best it's a stack overflow/google replacement.
There's some real perks to using AI to code - it helps a ton with templatable or repetitive code, and setting up tedious tasks. I hate doing that stuff by hand so being able to pass it off to copilot is great. But we already had tools that gave us 90% of the functionality copilot adds there, so it's not super novel, and I've never had it handle anything properly complicated at all successfully (asking GPT-5 to do your dynamic SQL calls is inviting disaster, for example. Requires hours of reworking just to get close.)
But we already had tools that gave us 90%
More reliable ones.
Fair, I've used it recently to translate a translations.ts file to Spanish.
But for repetitive code, I feel like it is kind of a slow down sometimes. I should have refactored instead.
Similarly I find it very useful for if I've written a tool script and really don't want to write the command line interface for it.
"Here's a well-documented function - write an argparser for it"
...then I fix its rubbish assumptions and mistakes. It's probably not drastically quicker but it doesn't require as much effort from me, meaning I can go harder on the actual function (rather than keeping some effort in reserve to get over the final hump).
I’ve had plenty of success using it to build things like docker compose yamls and the like, but for anything functional, it does often take a few tries to get it right. I never use its raw for anything in production. Only as a leaping off point to structure things.
For the missing 10% : the folder with copies of the code you have already wrote doing that.
(asking GPT-5 to do your dynamic SQL calls is inviting disaster, for example. Requires hours of reworking just to get close.)
Maybe it's the dynamic SQL calls themselves that are inviting disaster?
So much of the AI hype has been pointing to ten year old technology repackaged in a slick new interface.
AI is the iPod to the Zune of yesteryear.
At least when I'm cleaning up after shit devs who used Stack Overflow, I can usually search using a fragment of their code and find where they swiped it from and get some clue what the hell they were thinking. Now that they're all using AI chatbots, there's no trace.
I found that it only does well if the task is already well covered by the usual sources. Ask for anything novel and it shits the bed.
That's because it doesn't understand anything and is just vomiting forth output based on the code that was fed into it.
At absolute best.
My experience is it's the bottom stack overflow answers. Making up bullshit and nonexistent commands, etc.
They should make it more like SO and have it chastise you for asking a stupid question you should already know the answer to lol
If you know what you want, its automatic code completion can save you some typing in those cases where it gets it right (for repetitive or trivial code that doesn't require much thought). It's useful if you use it sparingly and can see through its bullshit.
For junior coders, though, it could be absolute poison.
In the beginning there were manufacturer's manuals, spec sheets, etc.
Then there were magazines, like Byte, InfoWorld, Compute! that showed you a bit more than just the specs
Then there were books, including the X for Dummies series that purported to teach you theory and practice
Then there was Google / Stack Overflow and friends
Somewhere along there, where depends a lot on your age, there were school / University courses
Now we have "AI mode"
Each step along that road has offered a significant speedup, connecting ideas to theory to practice.
I agree, all the "magic bullet" AI hype is far overblown. However, with AI something I new I can do is, interactively, develop a specification and a program. Throw out the code several times while the spec gets refined, re-implemented, tried in different languages with different libraries. It's still only good for "small" projects, but less than a year ago "small" meant less than 1000 lines of code. These days I'm seeing 300 lines of specification turn into 1500-3000 lines of code and have it running successfully within half a day.
I don't know if we're going to face a Kurzweilian singularity where these things start improving themselves at exponential rates, or if we'll hit another 30 year plateau like neural nets did back in the 1990s... As things are, Claude helps me make small projects several times faster than I could ever do with Google and Stack Overflow. And you can build significant systems out of cooperating small projects.
LLMs work great to ask about tons of documentation and learn more about high-level concepts. It's a good search engine.
The code they produce have basically always disappointed me.
On proprietary products, they are awful. So many hallucinations that waste hours. A manager used one on a code review of mine and only admitted it after I spent the afternoon chasing it.
Those happen so often. I've taken to stop calling them hallucinations anymore (that's anthropomorphising and over-selling what LLMs do imho). They are statistical prediction machines, and either they hit their practical limits of predicting useful output, or we just call it broken.
I think the next 10 years are going to be all about learning what LLMs are actually good for, and what they are fundamentally limited at no matter how much GPU ram we throw at it.
Not even proprietary, just niche things. In other words anything that's rarely used in open source code, because there's nothing to train the models on.
I sometimes get up to five lines of viable code. Then upon occasion what should have been a one liner tries to vomit all over my codebase. The best feature about AI enabled IDE is the button to decline the mess that was just inflicted.
In the past week I had two cases I thought would be "vibe coding" fodder, blazingly obvious just tedious. One time it just totally screwed up and I had to scrap it all. The other one generated about 4 functions in one go and was salvageable, though still off in weird ways. One of those was functional, just nonsensical. It had a function to check whether a certain condition was present or not, but instead of returning a boolean, it passed a pointer to a string and set the string to "" to indicate false... So damn bizarre, hard to follow and needlessly more lines of code, which is another theme of LLM generated code.
Imagine if we did "vibe city infrastructure". Just throw up a fucking suspension bridge and we'll hire some temps to come in later to find the bad welds and missing cables.
I'd much rather write my own bugs to have to waste hours fixing, thanks.
“No Duh,” say senior developers everywhere.
I'm so glad this was your first line in the post
it's slowing you down. The solution to that is to use it in even more places!
Wtf was up with that conclusion?
I don't think it's meant to be a conclusion. The article serves as a recap of several reports and studies about the effectivity of LLMs with coding, and the final quote from Bain & Company was a counterpoint to the previous ones asserting that productivity gains are minimal at best, but also that measuring productivity is a grey area.
I use AI as an entryway to learning or for finding the name or technique that I'm thinking of but can't remember or know it's name so then i can look elsewhere for proper documentation. I would never have it just blindly writing code.
Sadly search engines getting shitter has sort of made me have to use it to replace them.
Then it's also good to quickly parse an error for anything obviously wrong.
But will something be done about it?
NOooOoOoOoOoo. As long as it is still the new shiny toy for techbros and executive-bros to tinker with, it'll continue.
The most immediately understandable example I heard of this was from a senior developer who pointed out that LLM generated code will build a different code block every time it has to do the same thing. So if that function fails, you have to look at multiple incarnations of the same function, rather than saying “oh, let’s fix that function in the library we built.”
Yeah, code bloat with LLMs is fucking monstrous. If you use them, get used to immediately scouring your code for duplications.
Yeah if I use it and it generatse more than 5 lines of code, now I just immediately cancel it out because I know it's not worth even reading. So bad at repeating itself and falling to reasonably break things down in logical pieces..
With that I only have to read some of it's suggestions, still throw out probably 80% entirely, and fix up another 15%, and actually use 5% without modification.
I have been vibe coding a whole game in JavaScript to try it out. So far I have gotten a pretty ok game out of it. It's just a simple match three bubble pop type of thing so nothing crazy but I made a design and I am trying to implement it using mostly vibe coding.
That being said the code is awful. So many bad choices and spaghetti code. It also took longer than if I had written it myself.
So now I have a game that's kind of hard to modify haha. I may try to setup some unit tests and have it refactor using those.
Sounds like vibecoders will have to relearn the lessons of the past 40 years of software engineering.
As with every profession every generation... only this time on their own because every company forgot what employee training is and expects everyone to be born with 5 years of experience.
Wait, are you blaming AI for this, or yourself?
Blaming? I mean it wrote pretty much all of the code. I definitely wouldn't tell people I wrote it that way haha.
Glad someone paid a bunch of worthless McKinsey consultants what I could’ve told you myself
It is not worthless. My understanding is that management only trusts sources that are expensive.
Yep, going through that at work, they hired several consultant companies and near as I can tell, they just asked employees how the company was screwing up, we largely said the same things we always say to executives, they repeated them verbatim, and executives are now praising the insight on how to fix our business...
i use it for programming arduinos for my smarthome. its pretty nice but also aggravating.
I think maybe a good comparison is to written papers/ assignments. It can generate those just like it can generate code.
But it is not about the words themself, but about the content.
https://www.youtube.com/watch?v=VsE0BwQ3l8U&t=1492s
And the band plays on
Almost like its a desperate bid to blow another stock/asset bubble to keep 'the economy' going, from C suite, who all knew the housing bubble was going to pop when this all started, and now is.
Funniest thing in the world to me is high and mid level execs and managers who believe their own internal and external marketing.
The smarter people in the room realize their propoganda is in fact propogands, and are rolling their eyes internally that their henchmen are so stupid as to be true believers.
Might be there someday, but right now it’s basically a substitute for me googling some shit.
If I let it go ham, and code everything, it mutates into insanity in a very short period of time.
I'm honestly doubting it will get there someday, at least with the current use of LLMs. There just isn't true comprehension in them, no space for consideration in any novel dimension. If it takes incredible resources for companies to achieve sometimes-kinda-not-dogshit, I think we might need a new paradigm.
A crazy number of devs weren't even using EXISTING code assistant tooling.
Enterprise grade IDEs already had tons of tooling to generate classes and perform refactoring in a sane and algorithmic way. In a way that was deterministic.
So many use cases people have tried to sell me on (boilerplate handling) and im like "you have that now and don't even use it!".
I think there is probably a way to use llms to try and extract intention and then call real dependable tools to actually perform the actions. This cult of purity where the llm must actually be generating the tokens themselves... why?
I'm all for coding tools. I love them. They have to actually work though. Paradigm is completely wrong right now. I don't need it to "appear" good, i need it to BE good.
They are statistical prediction machines. The more they output, the larger the portion of their "context window" (statistical prior) becomes the very output they generated. It's a fundamental property of the current LLM design that the snake will eventually eat enough of it's tail to puke garbage code.
I think we've tapped most of the mileage we can get from the current science, the AI bros conveniently forget there have been multiple AI winters, I suspect we'll see at least one more before "AGI" (if we ever get there).
Writing apps with AI seems pretty cooked. But I've had some great successes using GitHub copilot for some annoying scripting work.
I think it's useful for writing mundane snippets I've written a million times or helping me with languages I'm less familiar with, but anything more compex becomes pretty spaghetti pretty quick.
AI is works well for mindless tasks. Data formatting, rough drafts, etc.
Once a task requires context and abstract thinking, AI can't handle it.
Trusting it to have not fucked with your data while formatting it is pretty bold.
Eh, I don't know. As long as you can break it down into smaller sub-tasks, AI can do some complex stuff. Just have to figure out where the line is. I've nudged it along into reading multiple LENGTHY API documentation pages and written some fairly complex scripting logic.
I miss the days when machine learning was fun. Poking together useless RNN models with a small dataset to make a digital Trump that talked about banging his daughter, end endless nipples flowing into America. Exploring the latent space between concepts.
No shit sherlock!
It remains to be seen whether the advent of “agentic AIs,” designed to autonomously execute a series of tasks, will change the situation.
“Agentic AI is already reshaping the enterprise, and only those that move decisively — redesigning their architecture, teams, and ways of working — will unlock its full value,” the report reads.
"Devs are slower with and don't trust LLM based tools. Surely, letting these tools off the leash will somehow manifest their value instead of exacerbating their problems."
Absolute madness.
It turns every prototyping exercise into a debugging exercise. Even talented coders often suck ass at debugging.
I am jack's complete lack of surprise.
The biggest value I get from AI in this space is when I get handed a pile of spagehtti and ask for an initial overview.
I thought that as well and got some code from someone that left the company and asked it to comment it.
It did the obvious "x= 5 // assign 5 to x" crap comments and then it got to the actually confusing part and just skipped that mess entirely....
Of course. Shareholders want results, and not just results for nVidia's bottom line.
I'd be inclined to try using it if it was smart enough to write my unit tests properly, but it's great at double inserting the same mock and have 0 working unit tests.
I might try using it to generate some javadoc though.. then when my org inevitably starts polling how much ai I use I won't be in the gutter lol
I personally think unit tests are the worst application of AI. Tests are there to ensure the code is correct, so ideally the dev would write the tests to verify that the AI-generated code is correct.
I personally don't use AI to write code, since writing code is the easiest and quickest part of my job. I instead use it to generate examples of using a new library, give me comparisons of different options, etc, and then I write the code after that. Basically, I use it as a replacement for a search engine/blog posts.
Ideally, there are requirements before anything, and some TDD types argue that the tests should come before the code as well.
Ideally, the customer is well represented during requirements development - ideally, not by the code developer.
Ideally, the code developer is not the same person that develops the unit tests.
Ideally, someone other than the test developer reviews the tests to assure that the tests do in-fact provide requirements coverage.
Ideally, the modules that come together to make the system function have similarly tight requirements and unit-tests and reviews, and the whole thing runs CI/CD to notify developers of any regressions/bugs within minutes of code check in.
In reality, some portion of that process (often, most of it) is short-cut for one or many reasons. Replacing the missing bits with AI is better than not having them at all.
To preface I don't actually use ai for anything at my job, which might be a bad metric but my workflow is 10x slower if i even try using ai
That said, I want AI to be able to do unit tests in the sense that I can write some starting ones, then it be able to infer what branches aren't covered and help me fill the rest.
Obviously it's not smart enough, and honestly I highly doubt it will ever be because that's the nature of llm, but my peeve with unit test is that testing branches usually entail just copying the exact same test but changing one field to be an invalid value, or a dependency to throw. It's not hard, just tedious. Branching coverage is already enforced, so you should know when you forgot to test a case.
Edit: my vision would be an interactive version rather than my company's current, where it just generates whatever it wants instantly. I'd want something to prompt me saying this branch is not covered, and then tell me how it will try to cover it. It eliminates the tedious work but still lets the dev know what they're doing.
I also think you should treat ai code as a pull request and actually review what it writes. My coworkers that do use it don't really proofread, so it ends up having some bad practices and code smells.
The reason tests are a good candidate is that there is a lot of boilerplate and no complicated business logic. It can be quite a time saver. You probably know some untested code in some project - you could get an llm to write some tests that would at least poke some key code paths, which is better than nothing. If the tests are wrong, it's barely worse than having no tests.
What model are you using? I've had such a radically different experience but I've only bothered with the latest models. The old ones weren't even worth trying with
I've seen it generate working unit tests plenty. In the sense that they pass.
...they do not actually test the functionality. Of course that function returns what you're asserting - you overwrote its actual output and checked against that!
One of the guys at my old job submitted a PR with tests that basically just mocked everything, tested nothing. Like,
with patch("something.whatever", return_value=True):
assert whatever(0) is True
assert whatever(1) is True
Except for a few dozen lines, with names that made it look like they were doing useful.
He used AI to generate them, of course. Pretty useless.
Oh wow. No shit. Anyway!
AI companies and investors are absolutely overhyping its capabilities, but if you haven't tried it before I'd strongly recommend doing so. For simple bash scripts and Python it almost always gets something workable first try, genuinely saving time.
AI LLMs are pretty terrible for nearly every other task I've tried. I suspect it's because the same amount of quality training data just doesn't exist for other fields.
Actually typing out code has literally never been the bottleneck. It's a vanishingly small amount of what we do. An experienced engineer can type out bash or Python scripts without so much as blinking. And better yet, they can do it without completely fabricating commands and library functions.
The hard part is truly understanding what it is you're trying to do in the first place, and that fundamentally requires a level of semantic comprehension that LLMs do not in any way possess.
It's very much like the "no code" solutions of yesteryear. They sound great on paper until you're faced with the reality of the buggy, unmaintainable nightmare pile of spaghetti code that they vomit into your repo.
LLMs are truly a complete joke for software development tasks. I remain among the top 3-4 developers in terms of speed and output at my workplace (and all of the fastest people refuse to use LLMs as well), and I don't create MRs chock full of bullshit that has to be ripped out (fucking sick of telling people to delete absolutely useless tests that do nothing but slow down our CI pipeline). The slowest people are those that keep banging their head against the LLM for "efficiency" when it's anything but.
It's the fucking stupidest trend I've seen in my career and I can't wait until people finally wake up and realize it's both incredibly inefficient and incredibly wasteful.
Oh god, please don't use it for Bash. LLM-generated Bash is such a fucking pot of horse shit bad practices. Regular people have a hard enough time writing good Bash, and something trained on all the fucking crap on StackOverflow and GitHub is inevitably going to be so bad...
Signed, a senior dev who is the "Bash guy" for a very large team.
The problem isn't the tool, it's the user: they don't know if they're getting good code or not, therefore they cannot make the prompt to improve it.
In my view the problems occur when using AI to do something you don't already know how to do.
I've found it's pretty good at refactoring existing code to use a different but well-supported and well documented library. It's absolutely terrible for a new and poorly documented library.
I recently tried using Copilot with Claude to implement something in a fairly young library, and did get the basics working, including a long repetitive string of "that doesn't work, I'm getting error msg [error]". Seven times of that, and suddenly it worked! I was quite amazed, though it failed me in many other ways with that library (imagining functions and options that don't exist). But then redoing the same thing in the older, better supported library, it got it right on the first try.
But maybe the biggest advantage of AI coding is that it allows me to code when my brain isn't fully engaged. Of course the risk there is that my brain might not fully engage because of the AI.
I always need to laugh when I read "Agentic AI"
I would say absolutely in the general sense nost people, and the salesmen, frame them in.
When I was invited to assist with the GDC development, I got a chance to partner with a few AI developers and see the development process firsthand, try my hand at it myself, and get my hands on a few low parameter models for my own personal use. It's really interesting just how capable some models are in their specific use-cases. However, even high param. models easily become useless at the drop of a hat.
I found the best case, one that's rarely done mind you, is integrate the model into a program that has the ability to call a known database. With a properly trained model to format output in both natural language and use a given database for context calls, and concrete information, the qualitative performance leaps ahead by bounds. Problem is, that requires so much customization it pretty much ends up being something a capable hobbyist would do, it's just not economically sound for a business to adopt.
Wait, it was hyped? Not just ridiculed?
A VP and his taking head sycophants at my work has not shut up about it. They went through some trouble to automatically measure employee use of the AI and made it a performance measure. So now we have it generate code so we don't get fired and mostly throw it away.
Oh, I did not know hell was hiring.
Im not super surprised, but AI has been really useful when it comes to learning or giving me a direction to look into something more directly.
Im not really an advocate for AI, but there are some really nice things AI can do. And i like to test the code quality of the models i have access to.
I always ask for a ftp server and dns server, to check what it can do and they work surprisingly well most of the time.
The good news is: AI is a lot less impressive than it seemed at first.
The bad news is: so are a lot of jobs.
LLMs/"Vibe Coding" is probably a little bit more useful than the average intern with some tasks bumping up to an early career hire (what would historically be a Junior Engineer before title inflation/stagnation).
As in: it can generate code that might do what you want. But you need (actual) senior engineers to review the code thoroughly. And... how do people get the experience they need to do that?
Which basically results in turning everyone into a manager. Except your reports aren't humans and you don't get more pay. Instead your reports are vscode plugins. Which... sounds like absolute hell but I can get why the (wannabe) management class loves that.
Even that description is vastly overselling it's usefulness. Every time someone says it's like a junior dev I just sigh, because literally the only reason I like junior devs is because they turn into not junior devs. Never once has assigning something to a junior dev made my job easier. The entire goal is to train them to the point they make PRs that I don't have to walk them through reworking.
Nowhere close to any junior ime. Grads learn very quickly. Interns only job is to understand. Code academy career switchers understand requirements and will ask questions. Subservient AI does fuck all of any of those things
They are more akin to yet another Rapid Application Development wave imo. Go see how the previous iterations have done. Lots are still with us (rails ftw!). I'll bet most will outlive LLMs
I can't even get Copilot to write Vitest files for React without making a mountain of junk code that describes drivel.
Were you trying to say "drivel?"
Yes. I'll go fix that.
I've found success using more powerful LLMs to help me create applications using the Rust programming language. If you use a weak LLM and ask it to do something very difficult you'll get bad results. You still need to have a fundamental understanding of good coding practices. Using an LLM to code doesn't replace the decision making.
Based on my experience with claude sonnet and gpt4/5... It's a little useful but generally annoying and fails more often than works.
I do think moderate use still comes out ahead, as it saves a bunch of typing when it does work, but I still get annoyed at the blatantly stupid suggestions I keep having to decline.
I remember GPT 4 being useless and constantly giving wrong information. Now with newer models they've become significantly more useful, especially when prompted to be extremely careful and to always double check to ensure the best response.
Usefulness really comes down to which model is being used. I've noticed most developers choose GPT for Copilot because that's what they are familiar with (or they often don't have a choice due to company policy). I recommend to try Claude Sonnet. How it works is true magic.
But I agree, repetitive tasks is what it should be used for. Planning is still programmer's job
GPT has been quite hit and miss for me, but Claude is usually quite solid.
It needs micromanaging, otherwise it will do bad design decisions and go off on unrelated side quests. When micromanaged it'll get you to that MVP very fast.
The trap is that you need to be able to find the errors it makes, or at least call them out immediately. Trying to have co-pilot fix it's own mistakes is usually a neverending prompt-cycle.
It can summarise big code bases fast, and find how things fit together a lot faster than me. It's been very useful when being thrown in head first into a new project.
Claude Sonnet? so you enjoy being lied to on a daily basis because that's all you're getting from Claude. Claude Code is alright but using Claude.ai you might as well just throw a dart at wall to get a solution on something.
Sounds exactly like my experience with Vibe Coding.
Industry? Yes, industry hires people who know how to do things needed by industry and who do nothing besides those things.
Programmers outside "industry" more often find themselves writing using the libraries they see for the first time and using languages they never thought to use. AI helps a lot here.
Except LLMs are absolutely terrible at working with a new, poorly documented library. Commonly-used, well-defined libraries? Sure! Working in an obscure language or an obscure framework? Good luck.
LLMs can surface information. It's perhaps the one place they're actually useful. They cannot reason in the same way a human programmer can, and all the big tech companies are trying to sell them on that basis.
Well, don't use it with new, poorly documented libraries. That is a common sense rule: use the tool where it is useful.
Somehow many LLM criticizers just claim that LLMs are shit because they can't autonomously write code. Yes, they can't. But they can do many other useful things.
shocked_pikachu_face.jpg
For most large projects, writing the code is the easy part anyway.
Writing new code is easier than editing someone else's code but editing a portion is still better than writing the entire program again from start to end.
Then there is LLMs which force you to edit the entire thing from start to end.