I also have found that AI has not lived up to many of its promises and have dialed back what AI gets control of. My projects were turning into unmaintainable messes. The people who say coding is solved aren't paying attention.
Are you in Python by chance? Python has a lot of crazy hidden/inexplicit/spooky action at a distance stuff (especially in the frameworks) that can make LLMs gunk up code by defensively programming or just burn context chasing data provenance
A model can reproduce large swaths of its training data exactly. It’s a different algorithm that powers its learning process (it’s why it needs trillions of tokens to even learn basics of language).
If there was a spectrum from copying on one end to creative production inspired from something else on the other end, the human generally lies heavily on the right end, while the model is much more on the left, that gap is large enough, that yes the model is in some sense “copying”.
Just because a model can reproduce parts of its training set doesn't mean that's what it's doing when it solves a programming problem. It can also reason about the problem, draw on its knowledge of algorithms and data structures, write tests targeting APIs it's never seen before, generate synthetic data and run experiments, etc. etc. Also, the claim that it can reproduce large swaths of its training data verbatim is an empirical one. I would be surprised if it could even reproduce 0.1% of the books it's ingested, for example.
Saying that AI can't do anything but copy or steal from humans seems to be a rhetorical technique used by people who are still unaware or in denial about the capabilities of the agentic systems released in the past few months. They can now one-shot theorems and programming problems in a few minutes that would be difficult and time-intensive for even the 99.9th percentile human expert.
No human on earth can copy at this scale. Yes AI is beyond a database lookup, it does have reasoning on top of this knowledge, but my original claim that if there is a spectrum between copying with minimal changes and creative inspiration with minimal copying, AI is one the copying end while humans are on the creative end. Humans are very bad at reproducing anything verbatim, even if they wrote it like a week ago.
I'm skeptical that there is a single spectrum like you're describing. It's not well defined. Say a human and an LLM prove a new theorem independently (without external help, i.e. from their own neural weights and reasoning). How do we measure how much each of them copied from previous work, as opposed to having learned from or been influenced by it?
A good workman shuts up and finds better tools without complaining.
Save your "you're holding it wrong" if you're not going to suggest how to hold it.
Cult speak escape hatches are intellectually lazy.
Edit: My latest project is all GPT-6 Astra High. It takes a lot of steering to keep it from adding a bunch of, while useful, features that are not strictly enough to the point. That main issue is it’ll use a lot of extra tokens in the process!
What was your process?
In case it is unclear, I am genuinely curious. I have great success with chatbots, but vibing coding has never gotten me further than a proof-of-concept.
https://williamcotton.github.io/datafarm-studio
Some demos of the above charting language:
https://williamcotton.github.io/algraf/demos
WASM, in browser editor, LSP, and more.
https://github.com/NousResearch/hermes-agent is 99% (just a guess) LLM generated. 1140 closed pull requests this week. 1.5k closed issues. The github insights page for commits doesn't load for me presumably because it can't handle this scale of commits. But I estimate ~1K commits per day on average.
There's a blog entry https://nousresearch.com/refactoring-hermes-with-1393-agents that details some work that was done by LLMs to refactor and improve the code.
I guess they know how to hold it?
I had a look at the kind of issues that are reported at that project (there's 15k of them, so I can at best assess a couple). It looks like a complete mess: A lot of concurrency and resource mismanagement issues and edge cases that in a better-managed project would have been avoided by construction. They will now will likely be solved by more defensive programming, driving overall complexity ever upwards.
If you really want to check some quantity metrics to try to reason about code quality, look at whether "fix" PRs are overall LOC neutral or negative (not counting tests). In this project, almost every "fix" is an addition. Worse, almost every fix is more branching.
If almost every PR is some sort of fix, and most of them add branching, and there's thousands of them weekly... That leads to only one place and I want to be nowhere near it.
Show me an AI that adds features by deleting code (https://www.folklore.org/Negative_2000_Lines_Of_Code.html) and I'll pay attention.
Your response was: "Well you're not doing it right, but these hermes devs know what they're doing".
But the blog post you linked to shows their prompt, which is:
> I want god files broken up. I want simplification across the board. I want unification of helpers and methods that can be reused. I want less if-if-if-if-if-if-else routing. I want code legibility up. I want interpretability of the codebase and how things connect to each other up.
So it sounds like AI made their code a mess too. They then tried to make the point of how much money they saved cleaning up the code with AI, that AI made a mess of to begin with.
And if you look at the merged PRs on that project, a ton of them are bug fixes... to the code the AI wrote. And that's been my personal experience too: AI creates a huge amount of churn in a codebase. Just vast amounts of PRs fixing code that the AI itself wrote.
I’m saying it’s probably multiple factors and both you and GP are right.
Seriously, I find I need to slow down the rate of change. I don't move forward until I understand the change proposed and have updated the docs. At the same time, I find that keeping up with the LLM/agent is exhausting. 3 hours with an LLM leaves me as tired as 6 hours with a keyboard had previously. I find that coding when tired or fuzzy yields code that shouldn't have been written in the first place. Sadly, once it's been written and debugged, the temporary fix becomes permanent.
I'm using local models, and they go slow enough that I have no trouble following along with what they're doing; but visually, both go and typescript, along with react native, make me puke. So I wouldn't be able to do this without AI.
I describe how to do it in my comment history, but it's basically a Super-TDD along with some custom engineering harness.
I don't want to say skill issue, but the same way you can give a chain saw to a teenager and one to a skill craftsman, well, AI can obviously create whatever you want it to do.
I think some of the variety is simply how fast SOTA models pump out garbage that you simply have to close your eyes because it's not sensible to just watch characters flow across the screen.
Almost all the coding I'm doing via AI is just faster than readable. But I can see the thinking traces and I stop to model when it's obvious it doesn't understand my intent, etc.
So I'm not doubting you created garbage. I'm just doubting that it's a product of soley AI use.
The models aren't good at architecture and design. But they take direction on architecture and design and design well and can refactor code quite effectively. AI agents can absolutely be used to clean up vibe coded code bases once you figure out if the investment is worth it. The mess can be avoided if you give them sufficient guidance on architecture and design upfront.
That said, doing so purely in text form doesn't feel great right now. I've been thinking about UML lately. The problem with that was the roundtrip after the code was generated and then the implenetation happened. I don't necessarily think UML is the solution, but neither is walls of dense text.
Can you elaborate to back up this claim? WHat exactly is your yardstick for "being good at SW design and architecture"?
Because I found the current SOTA AI models being great at architecture and design, much better in fact than most average real-world devs. Is your yardstick just the John Carmacks of the world by any chance? Because most devs are not John Carmack. They are also not Linus Torvalds, they are not Stallmann, etc.
Maybe your LLM experience is still stuck in the 2023 era of ChatGPT?
And do you consider yourself to be representative of the average developer, above them, or below them?
LLMs don't even need to be better than the average dev, let alone the top performing ones, like you. If they can just be better than the bottom 20% of devs and white collar workers in general(easily achievable when you've been around the block and saw how many useless people just keep warm chairs for high wages in large companies), that's already a huge win for those products.
What I mean, at a previous job I had ran into a memory leak issue in our backend and discovered a colleague pushed a library into prod which came with comments in the source code saying "DO NOT USE IN PROD, IT CAUSES A MEMORY LEAK!". There's cases where human stupidity and carelessness far surpasses whatever issues LLMs cause so maybe the average dev isn't really that much better than the SOTA LLMs.
1) duplication - LLMs are great at generating lots of text, so its faster and easier for them to generate entirely new facilities that overlap heavily with existing ones then it is for them find existing facilities that should be expanded and refactored (note I just said 'find'; actually editing raises the time and difficulty even more). This is fine for a while as the duplicate facilities usually work just fine, up until something needs to be changed across all of them and they miss changing one or more of them, things break, and a bunch of tokens have to be burned tracking down the issue.
2) ever increasing surface area - even when making changes that do expand a facility without much duplication they frequently only add without removing much of anything or changing the overall design of the facility to reduce the amount of state its tracking and the number of branches it has based on that state. I've never seen one decide to split up something large or with too many responsibilities on their own. They will happily create a god class or function and just keep making it bigger.
Late 2025 also had a step change when agents could largely code autonomously without handholding like previously, and to be honest it's not worth hearing opinions about AI from before that time, that's how significant the change was.
Where AI did make a lot of impact is triage. I can throw a messy bug report of an intermittent issue at claude, and tell I need issue reproduced and fix developed, and there is a well above 50% chance it'll deliver. Never commit that fix to the codebase as-is, of course.
Unlike a compiler it won't give up at the first sign of trouble but that just means it left alone it will dig bigger and bigger holes.
Treat prompt engineering as a discipline and refine your technique. When it produces garbage throw out the work and start over until you figure it out.
`total = dev + review`
If dev approaches zero, but you review at the same pace as you always have, are you in a better position? Yes.
Will you potentially have a backlog of code waiting for review? Also yes.
Would you prefer to be waiting for the dev team for all of the time instead, then still have the same amount of reviewing to do at the end of it? Absolutely not.
On the other hand asking these clankers "review the feature branch I wrote" and "review my entire codebase for bugs" or "help me debug this" has saved me months of prospective work.
And more recently most major models have been getting _really_ good at RE, for example you can have OAI models (and maybe A/'s if they don't refuse) use idalib MCP and reverse-engineer stuff from start to finish, then follow up with GLM 5.3 for vuln assessment and exploit PoC.
Stuff that used to take weeks or months now just takes a few hours, or less.
You say you used AI and your projects turned into unmaintainable messes, so your conclusion is that it means AI is not living up to its promises.
I guess if the argument is “AI makes it so you always get a great result no matter how you use it”, then your argument is sound. Your projects not working out proves that AI doesn’t always work no matter what.
However, that doesn’t mean you can’t use AI to create sustainable and well organized code. Failing to do something doesn’t mean it’s impossible and anyone who thinks they can is not paying attention.
I can’t run a marathon. If I went out and tried to run one, I would get a few miles and collapse, failing completely.
I don’t think it would be reasonable, though, at that point to say “running a marathon is impossible, anyone who says they can do it clearly lying. I tried and didn’t even make it 5 miles!”
I wish people would stop assuming their experience with something is the only possible truth.
As ai becomes better these people will begin changing their story because it’s utterly obvious what’s happening.