The AI coding trap
chrisloy.dev
chrisloy.dev
The plan-build-test-reflect loop is equally important when using an LLM to generate code, as anyone who's seriously used the tech knows: if you yolo your way through a build without thought, it will collapse in on itself quickly. But if you DO apply that loop, you get to spend much more time on the part I personally enjoy, architecting the build and testing the resultant experience.
> While the LLMs get to blast through all the fun, easy work at lightning speed, we are then left with all the thankless tasks
This is, to me, the root of one disagreement I see playing out in every industry where AI has achieved any level of mastery. There's a divide between people who enjoy the physical experience of the work and people who enjoy the mental experience of the work. If the thinking bit is your favorite part, AI allows you to spend nearly all of your time there if you wish, from concept through troubleshooting. But if you like the doing, the typing, fiddling with knobs and configs, etc etc, all AI does is take the good part away.
1) Bad actors using AI at scale to do bad things
2) AI just commodifying everything and making humans into zoo animals
I'm on a small personal project with it intentionally off, and I honestly feel I'm moving through it faster and certainly having a better time. I also have a much better feel for the code.
These are all just vibes, in the parlance of our times, but it's making me question why I'm bothering with LLM assisted coding.
Velocity is rarely the thing in my niche, and I'm not convinced babysitting an agent is all in all faster. It's certainly a lot less enjoyable, and that matters, right?
Using "AI" is just like speed reading a math book without ever doing single exercise. The proponents rarely have any serious public code bases.
That's not to say that LLMs as good as some of the more outrageous claims. You do still need to do a lot of work to implement code. But if you're not finding value at all it honestly reflects badly on you and your ability to use tools.
The craziest thing is i see the above type of comment on linked in regularly. Which is jaw dropping. Prospective hiring managers will read it and think "Wow you think advertising a lack of knowledge is helpful to your career?" Big tech co's are literally firing people with attitudes like the above. There's no room for people who refuse to adapt.
I put absolute LLM negativity right up there with comments like "i never use a debugger and just use printf statements". To me it just screams you never learnt the tool.
To me it just feels different. Learning to use a debugger made me feel more powerful and "in control" (even though I still use a lot of print debugging; every tool has its place). Using AI assisted coding makes me feel like a manager who has to micro-manage a noob - it's exhausting.
To many of us coding us simply more fun. At the same time, many of us could benefit from that exercise with or without the LLM.
The open source code of these companies is also not that great and definitely not bug free. Perhaps these companies should do more thinking and less tooling politics.
You are in a forum full of people that routinely claim that vibe coding is the future, that LLMs already can fully replace engineers, and if you don't think so you are just a naysayer that is doing it wrong.
Rephrasing your claim, LLMs are just moderately useful, far from being the future-defining technology people invested in it wants it to be. But you choose to rally against people not interested in marketing it further.
Given the credentials you decided to share, I find it unsurprising.
Violent insecurity and authoritarianism... definitely not compensating for anything there.
As LLM get better, more and more people will be able to create projects with only rudimentary language understanding. I don't think LLMS can ever be as good as some of the outrageous claims; it's a lot like that 3rd grade project kids do on writing instruction on making a PB&J. LLM's cannot read minds and will only follow the prompt given to them. What I'm trying to say is that eventually there will be a time where being able to effectively manage coding agents efficiently will be more externally valuable than knowing how to write code.
This isn't to say that engineering experience is not valuable. Having a deep understanding of how to design and build secure and efficient software is a huge moat between experienced engineers and vibecoders like me, and not learning how to best use the tools that are quickly changing how the world operates will leave them behind.
You described the current AI Bubble.
I mean, I guess when I was really early in my career I'd get a kick out of writing a clever loop or whatever, and I drank deep from all the low level coding wisdom that was available, but the scope of what I care about these days has long since expanded outward.
And this should not be a surprise at all. Humans are optimisers of truth NOT maximisers. There is a subtle and nuanced difference. Very few actually spend their entire existence being maximizers, its pretty exhausting to be of that kind.
Optimising = we look for what feels right or surpassses some threshold of "looks about right". Maximizing = we think deeply and logically reason to what is right and conduct tests to ensure it is so.
Now if you have the discipline to choose when to shift between the two modes this can work. Most people do not though. And therein lies the danger.
AI can only recycle the past.
Since we don't know what else might already exist in the world without digging very deep, we fool ourselves into thinking that we do something very original and unique.
This AI will be built from it's own excretions, recursively.
That might be a hilarious trap.
The giant's shoulder just give us access to useful an lucrative spaces in which to build that novel stuff.
The article sort of goes sideways with this idea but pointing out that AI coding robs you a deep understanding of the code it produces is a valid and important criticism of AI coding.
A software engineer's primary job isn't producing code, but producing a functional software system. Most important to that is the extremely hard to convey "mental model" of how the code works and an expertise of the domain it works in. Code is a derived asset of this mental model. And you will never know code as well as a reader and you would have as the author for anything larger than a very small project.
There are other consequences of not building this mental model of a piece of software. Reasoning at the level of syntax is proving to have limits that LLM-based coding agents are having trouble scaling beyond.
This feels very true - but also consider how much code exists for which many of the current maintainers were not involved in the original writing.
There are many anecdotal rules out there about how much time is spent reading code vs writing. If you consider the industry as a whole, it seems to me that the introduction of generative code-writing tools is actually not moving the needle as far as people are claiming.
We _already_ live in a world where most of us spend much of our time reading and trying to comprehend code written by others from the past.
What's the difference between a messy codebase created by a genAI, and a messy codebase where all the original authors of the code have moved on and aren't available to ask questions?
The difference is the hope of getting out of that situation. If you've inherited a messy and incoherent code base, you recognize that as a problem and work on fixing it. You can build an understanding of the code through first reading and then probably rewriting some of it. This over time improves your ability to reason about that code.
If you're constantly putting yourself back into that situation through relegating the reasoning about code to coding agent, then you won't develop a mental model. You're constantly back at Day 1 of having to "own" someone else's code.
If only there were some people in software engineering in this situation before AI… oh wait.
In the current times you’re either an agent manager or you’re in for a surprise.
This opinion seems to be popular, if only in this forum and not in general.
What I do not understand is this;
In order to use LLM's to generate code, the engineer
has to understand the problem sufficient enough to
formulate prompt(s) to use in order to get usable
output (code). Assuming the engineer has this level
of understanding along with knowledge of the target
programming language and libraries used, how is using
LLM code generation anything more than a typing saver?Where this gets exhausting is when you assume certain things that you know are necessary but don't want to verify - maybe it let's you submit an email form with no email, or validates password as an email field for some reason, etc. But as LLMs improve their assumptions or you manage context correctly, the scale tips towards this being a useful engineering tool, especially when what you are doing is a well-trodden path.
This is why I’ve constrained my use of AI agents to mostly “read-only and explain” use cases, but I have very strict conditions for letting it write. In any case, whatever productivity gains you supposedly “get” for its write scenarios, you should be subtracting your expenses to fix its output later and/or payments made for a larger context window or better reasoning. It’s usually not worth the trouble to me when I have plenty of experience and knowledge to draw from and can write the code as it should be myself.
Want to make a quick change or fix? The agent will likely figure out a way to do it in minutes rather the than hours it would take me to do so.
Want to get a good understanding of the architecture and code layout? Working with an agent for search and summary cuts my time down by an order of magnitude.
So while agree there’s a lot more “what the heck is this ugly pile of if else statements doing?” And “why are there three modules handling transforms?”, there is a corresponding drop in cost to adding features and paying down tech debt. Finding the right balance is a bit different in the agentic coding world, but it’s a different mindset and set of practices to develop.
When the agent can't kick the can any more who is going to be held responsible? If it is going to be me then I'd prefer to have spent the hours understanding the code.
This is actually a pretty huge question about AI in general
When AI is running autonomously, where is the accountability when it goes off the rails?
I'm against AI for a number of reasons, but this is one of the biggest. A computer cannot be held accountable therefore a computer must never make executive decisions
While an engineer is "it" they just have to cross their fingers and hope no job ending skeletons are resurrected until they can tag some other poor sod.
Messy codebases made by humans are known to be a bad thing that causes big problems for software that needs to be maintained and changed. Much effort goes into preventing them and cleaning them up.
> We _already_ live in a world where most of us spend much of our time reading and trying to comprehend code written by others from the past.
I bring this up because it's a solution to what you're pointing out as a problem and yet the status quo is to write even messier and harder to understand code (even before AI code). So I'm just saying, humans are really good at shooting themselves in the foot and blaming it on someone else or acting like the bullet came out of nowhere.
> What's the difference between
If it's the same level of messiness, then sure, it's equal. But in a real world setting there's a continuous transition of people. One doesn't work on code in isolation, quit, and then a new person works on that code also in isolation. So maybe it's not the original authors but rather the original authors are a Ship of Theseus. Your premise isn't entirely accurate and I think the difference matters
In 1969 as a newly hired graduate working for the largest construction company in the country, one of my first assignments was to read through a badly formatted COBOL source code listing on paper, line by line, with 2 others. Each of us had a printout of a different version of the software, trying to discover where exactly the three versions were giving different outputs. Plus ça change, plus c'est la même chose
In my experience, the type of messes created by humans and the type of messes created by genAI are immensely different, and at times require different skill sets to dissect.
Very unskilled programmers can use generative AI to create complex code that is hard to understand.
Unskilled programmers on their own write simpler code with more obvious easy to solve mistakes.
On the other hand, all companies I have worked for tried to avoid have unmaintained code (with different levels of success). AI tech debt seems to be added on purpose pushed by upper management.
There is a technical and cultural difference.
Regardless how horrible somebody else's code is, there is some underlying method, or logic reflecting how given person forms mental model of the problem and breaks it down to little manageable pieces. You can learn that style, over time even ignoring it and seeing their code in same ways you see yours. llm code has none of that, if yes its by pure chance that won't repeat.
All code is temporary and should be treated as ephemeral. Even if it lives for a long time, at the end of the day what really matters is data. Data is what helps you develop the type of deep understanding and expertise of the domain that is needed to produce high quality software.
In most problem domains, if you understand the data and how it is modeled, the need to be on top of how every single line of code works and the nitty-gritty of how things are wired together largely disappears. This is the thought behind the idiom “Don’t tell me what the code says—show me the data, and I’ll tell you what the code does.”
It is therefore crucial to start every AI-driven development effort with data modeling, and have lots of long conversations with AI to make sure you learn the domain well and have all your questions answered. In most cases, the rest is mostly just busywork, and handing it off to AI is how people achieve the type of productivity gains you read about.
Of course, that's not to say you should blindly accept everything the AI generates. Reading the code and asking the AI questions is still important. But the idea that the only way to develop an understanding of the problem is to write the code yourself is no longer true. In fact, it was never true to begin with.
You need to understand both the data, and the transformations that are being made to the data.
Why? Code has always been the artifact. Thinking about and understanding the domain clearly and solving problems is where the intrinsic value is at (but I'd suspect that in the future this, too, will go away).
Artifacts are snapshots of system knowledge (code, builds, docs, configs, etc.).
The product is the living whole that emerges from these artifacts working together and delivering value.
When I see someone dismissing the code as a small irrelevant part of the task of writing software, it's like hearing that the low-level design and physical construction of a bridge is an irrelevant side-effect of my desire to cross a body of water. Like, maybe that's true in a philosophical sense, but at the end of the day we are building a real-world bridge that needs to conform to real-world constraints, and every little detail is going to be important. I wouldn't want to cross a bridge built by someone who thinks otherwise.
This is why we have the idiom: “Don’t tell me what the code says—show me the data, and I’ll tell you what the code does.”
It might be true if you just look at development cost, but if you look at value as a whole it isn't. And even just development cost its often not true, since time spent waiting around by the developer for tests to run and things to start also slows things down, taking a bit of time there to reduce cpu time is well worth it just to get things done faster.
Also, why
That is an extreme case though, I didn't mean that all optimizations are always worth it, but if we look at marginal value gained from optimizations today the payback is usually massive.
It isn't done enough since managers tend to undervalue user and developer time. But users don't undervalue user time, if your program wastes their time many users will stop using it, users are pretty rational about that aspect and prefer faster products or sites unless they are very lacking. If a website is slow a few times in a row I start looking for alternatives, and data says most users do that.
I even stopped my JetBrains subscription since the editor got so much slower in an update, so I just use the one I can keep forever as I don't want their patched editor. If it didn't get slower I'd gladly keep it as I liked some of the new features, but it being slower was enough to make me go back.
Also, while managers can obvious agree that making developer spend less time waiting is a good thing, it is very rare for managers to tell you to optimize compilation times or such, and pretty simple optimizations there can often make that part of the work massively faster. Like, if you profile your C++ compiler and look what files it spends time compiling, then look at those files to figure out why its so slow there, you can find these weird things and fixing those speeds it up 10x, so what took 30 seconds now takes 3 seconds, that is obviously very helpful and if you are used to that sort of thing you could do it in a couple of hours.
Many of us do frequently argue for something similar. Take a look at Casey Muratori’s performance aware programming series if you care about the arguments.
Maybe we should, all it took was Figma taking it seriously and working at a lower level to make every other competitor feel awful and clunky next to it then it went on to dominate the market.
The difference is that there is no point (I know or would encounter) in which a compiler would not actually be able to do the job, and I would need to write manual assembly to fix some parts that the compiler could not compile. Yes a proficient programmer could probably do that to optimise the code, but the code would run and do the job regardless. That is not the case for LLMs, there is a non-zero changeyou get to the point of LLM agents getting stuck and it makes more sense to get hands dirty than iterating with agents.
At that point yeah, a project manager would be able to build everything.
We are not there.
Letting a Python interpreter manage your memory is one thing because it's usually irrelevant, but you can't say the same thing about business logic. Encoding those precise rules and considering all of the gnarly real-world edge cases is what defines your software.
There are no "higher level details" in software development, those are in the domain of different jobs like project managers or analysts. Once AI can reliably translate fuzzy natural language into precise and accurate code, software development will simply die as a profession. Our jobs won't morph into something different - this is our job.
I'm the non-software type of Engineer. I've always kind of viewed code as a way to bridge mathematics and control logic.
When I was at university I was required to take a first year course called "Introduction to Programming and Algorithms". It essentially taught us how to think about problem solving from a computer programming perspective. One example I still remember from the course was learning how you can use a computer solve something like Newton's Method.
I don't really hear a lot of software people talk about Algorithms but for me that is where the real power of programming lives. I can see some idealized future where you write programs just by mix and matching algorithms and almost every problem becomes essentially a state machine. To move from state A to State B I apply these transformations which map to these well known algorithms. I could see an AI being capable of that sort of pattern matching.
One-shotting anything like this is a non-starter for any remotely complex task. The reason is that fuzzy language is ambiguous and poorly defined. So even in this scenario you enter into a domain where it's going to require iterative cycling and refinement. And I'm not even considering the endless meta-factors that further complicate this, like performance considerations depending on how you plan to deploy.
And even if language were perfectly well defined, you'd end up with 'prompts' that would essentially be source codes in their own right. I have a friend who is rather smart, but not a tech type - and he's currently working on developing a very simple project using LLMs, but it's still a "real" project in that there are certain edge cases you need to consider, various cross-functionality in the UI that needs to be carried out, interactions with some underlying systems, and so on.
His 'prompt' is gradually turning into just a natural language program, of comparable length and complexity. And with the amount of credits he's churning through making it, in the end he may well have been much better off just hiring some programmers on one of those 'gig programming' sites.
------
And beyond all of this, even if you can surmount these issues - which I think may be inherently impossible - you have another one. The reason people hire software devs is not because they can't do it themselves, but because they want to devote their attention to other things. E.g. - most of everybody could do janitorial work, yet companies still hire millions of janitors. So the 'worst case' scenario would be that you dramatically lower the barriers to entry to software development, and wages plummet accordingly.
The best humans on the planet constantly fail at this so the assumption is that some post LLM AGI will do this?
Although I do agree that there is a possibility that we'll build a relatively reliable abstraction using LLMs at some point, so this issue will go away. There probably be some restrictions, but I think it is possible.
In any of my teams with moderate to significant code bases, we've always had to lean very hard into code comments and documentation, because a developer will forget in a few months the fine details of what they've previously built. And further, any org with turnover needs to have someone new come in and be able to understand what's there.
I don't think I've met a developer that keeps all of the architecture and design deeply in their mind at all times. We all often enough need to go walk back through and rediscover what we have.
Which is to say... if the LLM generator was instead a colleague or neighboring team, you'd still need to keep up with them. If you can adapt those habits to the generative code then it doesn't seem to be a bit leap.
No it isn't. There's literally nothing about the process that forces you to skip understanding. Any such skips are purely due to the lack of will on the developer's side. This lack of will to learn will not change the outcomes for you regardless of whether you're using an LLM. You can spend as much time as you want asking the LLM for in-depth explanations and examples to test your understanding.
So many of the criticisms of coding with LLMs I've seen really do sound like they're coming from people who already started with a pre-existing bias, fiddled with with for a short bit (or worse, never actually tried it at all) and assumed their limited experience is the be-all end-all of the subject. Either that, or they're typical skill issues.
This is the whole point. The marginal dev will go to the path of least resistance, which is to skip the understanding and churn out a bunch of code. That is why it's a problem.
You are effectively saying "just be a good dev, there's literally nothing about AI which is stopping you from being a good dev" which is completely correct and also missing the point.
The marginal developer is not going to put in the effort to wield AI in a skillful way. They're going to slop their way through. It is a concern for widespread AI coding, even if it's not a concern for you or your skill peers in particular.
Purely vibe-coded projects will soon break in unexplainable ways as they grow beyond trivial levels. Once that happens their devs will either need to adapt and learn coding for real or be PIP'd. I can't imagine any such devs lasting long in the current layoff-happy environment. So it seems like a self-correcting problem no?
(Maybe AGI, whatever that is, will change things, but I'm not holding my breath.)
The real problem we should be discussing is, how do we convince students and apprentices to abstain from AI until they learn the ropes for real.
I wouldn't expect things to calm down for a while, even if real-life results are worse. You can make excuses for underperformance of these things for a very long time, especially if the CEO or other executives are invested.
> The real problem we should be discussing is, how do we convince students and apprentices to abstain from AI until they learn the ropes for real
I hate to say it but that's never going to happen :/
At the same time, I'll also admit that AI resistance is real; we see it in the comments here for various reasons -- job displacement fears, valid complaints about AI reliability, ethical opposition, etc. So there could be a valid need for strong incentives to adopt it.
Unfortunately, AI is also deceptively hard to use effectively (a common refrain of mine.) Ideally AI mandates would come with some structured training tailored for each role, but the fact that this is not happening makes me wonder about either the execs' competency or their motives.
Learning the ropes looks different now. You used to learn by doing, now you need to learn by directing. In order to know how to direct well, you have to first be knowledgeable. So, if you're starting work in an unfamiliar technology, then a good starting point is read whatever O'Reilly book gives a good overview, so that you understand the landscape of what's possible with the tool and can spot when the LLM is doing (now) obvious bullshit.
You can't just Yolo it for shit you don't know and get good results, but if you build a foundation first through reading, you will do a lot better.
My current project is in a technical domain I had very little prior background in, but I've been getting actual, visible results since day one because of AI. The amazing thing is that for any task I give it, the AI provides me a very useful overview of the thing it produces, and I have conversations with it if I have further questions. So I'm building domain knowledge incrementally even as I'm making progress on the project!
But I also know that this is only possible because of the pre-existing foundation of my experience as a software engineer. This lets me understand the language the AI uses to explain things, and I can dive deeper if I have questions. It also lets me understand what the code is doing, which lets me catch subtle issues before they compound.
I suppose it's the same with reading books, but books being static tend to give a much broader overview upfront, whereas interacting with LLMs results in a much more focused learning path.
So a foundation is essential, but it can now be much more general -- such as generic coding ability -- but that only comes with extensive hands-on experience. There is at least one preliminary study showing that students who rely on AI do not develop the critical problem solving, coding and debugging skills necessary to be good programmers:
That's just it. You can only use AI usefully for coding* once you've spent years beating your head against code "the hard way". I'm not sure what that looks like for the next cohort, since they have AI on day 1.
* That is, assuming it's nontrivial.
My mental model of it is that coding with LLMs amplified both what you know and what you don't.
When you know something, you can direct it productively much faster to a desirable outcome than you could on your own.
When you don't know something, the time you normally would have spent researching to build a sufficient understanding to start working on it can be replaced with evaluating the random stuff the LLM comes up with which oftentimes works but not in the way it ought to, though since you can get to some result quickly, the trade-off to do the research feels somehow less worth it.
Probably if you don't have any idea how to accomplish the task you need to cultivate the habit of still doing the research first. Wielding it skillfully is now the task of our industry, so we ought to be developing that skill and cultivating it in our team members.
I do think when AI writes comprehensible code you can spend as much time as necessary asking questions to better understand it. You can ask about tradeoffs and alternatives without offending anybody and actually get to a better place in your own understanding than would be possible alone.
The comparison is oniy reasonable if most of your job is spent trying to understand their code, and make sure it did what you wanted. And with them standing next to you, ready to answer questons, explain anything I don't understand and pull in any external, relevant parts of the codebase.
The industry has institutionalized this by making code reviews a very standard best practice. People think of code reviews mainly as a mechanism to reduce bugs, but turns out the biggest benefits (born out by studies) actually are better context-sharing amongst the team, mentoring junior engineers, and onboarding of new team-mates. It ensures that everyone has the same mental model of the system despite working on different parts of it (c.f. the story of the blind men and the elephant.) This results in better ownership and fewer defects per line of code.
Note, this also doesn't mean everybody reviews each and every PR. But any non-trivial PR should be reviewed by team-mates with appropriate context.
People tend to take the path of least resistance, maybe not everyone, maybe not right away, but if you create opportunities to write poor code then people will take them - more than ever it becomes important to have strong CI, review and testing practices.
Edit: okay, maybe I am feeling a little pessimistic this morning :)
There's nothing about C that "forces" people to write buffer overflows. But, when writing C, the path of least resistance is to produce memory-unsafe code. Your position reminds me of C advocates who say that "good developers possess the expertise and put in the effort to write safe code without safeguards," which is a bad argument because we know memory errors do show up in critical code regardless of what a hypothetical "good C dev" does.
If the path of least resistance for a given tool involve using that tool dangerously, then it's a dangerous tool. We say chefs should work with sharp knives, but with good knife technique (claw grip, for instance) safety is the path of least resistance. I have yet to hear of an LLM workflow where skimming the generated code is made harder than comprehensively auditing it, and I'm not sure that such a workflow would feel good or be productive.
You need to care if (a) your business logic requirements are super annoyingly complex, (b) you have hard performance requirements, or (c) both. (c) is the most rare, (a) is the most common of those three conditions; much of the programmer pay disparity between the top and the middle or bottom is due to this, but even the jobs where the complexity is "only" business requirements tend to be quite a bit better compensated than the "simple requirements, simple needs" ones.
I think there's a case to be made that LLM tools will likely make it harder for people to make that jump, if they want to. (Alternately they could advance to the point where the distinction changes a bit, and is more purely architectural; or they could advance to the point where anyone can use an LLM to do anything - but there are so many conditional nuances to what the "right decision" is in any given scenario there that I'm skeptical.)
A lot of times floor-raising things don't remove the levels, they just push everything higher. Like a cheap crap movie today will visually look "better" from a technology POV (sharpness, special effects, noise, etc) than Jurassic Park from the 90s, but the craft parts won't (shot framing, deliberate shifts of focus, selection of the best takes). So everyone will just get more efficient and more will be expected, but still stratified.
And so some people will still want to figure out how to go from a lower-paying job to a higher-paying one. And hopefully there are still opportunities, and we don't just turn into other fields, picking by university reputations and connections.
But one of the most fun things you can do is C: creative game development coding. Like coding world simulations etc, you want to be both very fast but the rules and interactions etc is very coupled and complex compared to most regular enterprise logic that is more decoupled.
So while most work programmers do fits A, the work people dream about doing is C, and that means LLM doesn't help you make fun things, it just removes the boring jobs.
And as we mapped put this landscape, hadn't there been countless situations where things felt dumb and annoying, and then situation in sometimes they became useful, and sometimes they remained dumb? Something you thought is making you actively loosing brain cells as you're doing them, because you're doing them wrong?
Or are you to claim that every hurdle you cross, every roadblock you encounter, every annoyance you overcome has pedagogical value to your career? There are so many dumb things out there. And what's more, there's so many things that appear dumb at first and then, when used right, become very powerful. AI is that: Something that you can use to shoot yourself in the foot, if used wrong, but if used right, it can be incredibly powerful. Just like C++, Linux, CORS, npm, tcp, whatever, everything basically.
You can describe what the code should do with natural language.
I've found that using literate programming with agent calls to write the tests first, then code, then the human refining the description of the code, and going back to 1 is surprisingly good at this. One of these days I'll get around to writing an emacs mode to automate it because right now it's yanking and killing between nearly a dozen windows.
Of course this is much slower than regular development but you end up with world class documentation and understanding of the code base.
The role of the programmer would then be to test if the rules are being applied correctly. If not, there are no bugs to fix, you simply clarify the business rules and ask for a new program.
I like to imagine what it must be like for a non technical business owner who employees programmers today. There is a meeting where a process or outcome is described, and a few weeks / months / years a program is delivered. The only way to know if it does what was requested is to poke it a bit and see if it works. The business owner has no metal modal of the code and can't go in and fix bugs.
update: I'm not suggesting I believe AI is anywhere near being this capable.
The programming building blocks popular today (lines of code in sub-routines and modules) do not support such a jump?
This same error in thinking happens in relation to AI agents too. Even if the agent is perfect (not really possible) but other links in the chain are slower, the overall speed of the loop still does not increase. To increase productivity with AI you need to think of the complete loop, reorganize and optimize every link in the chain. In other words a business has to redesign itself for AI, not just apply AI on top.
Same is true for coding with AI, you can't just do your old style manual coding but with AI, you need a new style of work. Maybe you start with constraint design, requirements, tests, and then you let the agent loose and not check the code, you need to automate that part, it needs comprehensive automated testing. The LLM is like a blind force, you need to channel it to make it useful. LLM+Constraints == accountable LLM, but LLM without constraints == unaccountable.
Oh, that's a great idea. Then you get a constraint-based language and write your constraints there!
I’ve switched to using Rust because of the rich type system and pedantic yet helpful compiler errors. I focus on high level design, traits, important types - then I write integration tests and let Claude go to town. I’ve been experimenting with this approach on my side project (backend web services related to GIS - nothing terribly low level) for about 4 months now and I honestly don’t know if it’s any faster than just writing the code myself. I suspect it’s not or only marginally faster at best.
I often find that I end up in a place where the ai generated code just has too many issues collected over iterations and needs serious refactoring that the agent is incapable of performing satisfactorily. So I must do it myself and that work is substantially harder than it would have been had I just written everything myself in the first place.
At work - I find that I have a deep enough understanding of our codebase that the agents are mostly a net-loss outside of boilerplate.
Perhaps I’m holding it wrong but I’ve been doing this for a while now. I am extremely motivated to build a successful side project and try to bootstrap myself out of the corporate world. I read blogs and watch vlogs on how others build their workflows and I just cannot replicate these claims of huge productivity gains.
Here's mine, I use Cline occasionally to help me code but more and more I find myself just coding by hand. The reason is pretty simple which is with these AI tools you for the most part replace writing code with writing a prompt.
I look at it like this, if writing the prompt, and the inference time is less than what it would take me to write the code by hand I usually go the AI route. But this is usually for refactoring tasks where I consider the main bottleneck to be the speed at which my fingers can type.
For virtually all other problems it goes something like this: I can do X task in 10 minutes if i code it manually or I can prompt AI to do it and by the time I finish crafting the prompt and execute, it takes me about 8 minutes. Yes that's a savings of 2 minutes on that task and that's all fine and good assuming that the AI didn't make a mistake, if I have to go back and re-prompt or manually fix something, then all of a sudden the time it took me to complete that task is now 10-12 minutes with AI. Here the best case scenario is I just spent some AI credits for zero time savings and worse case is I spent AI credits AND the task was slower in the end.
With all sorts of tasks I now find myself making this calculation and for the most part, I find that doing it by hand is just the "safer" option, both in terms of code output but also in terms of time spent on the task.
I'm convinced I spend more time typing and end up typing more letters and words when AI coding than when not.
My hands are hurting me more from the extra typing I have to do now lol.
I'm actually annoyed they haven't integrated their voice to text models inside their coding agents yet.
I find the generic ones like the ones I can use anywhere on Mac to be crap.
If you've used the ChatGPT voice to text model you know what I mean.
There are also third parties like Wispr that I haven't tried, but might do a better job? No idea.
You can use it with Spokenly (free app, bring your own Soniox API key) on macOS and iOS (virtual voice keyboard)
Disclaimer: I've worked for Soniox
That being said, these agents may still just YOLO and ignore your instructions on occasion, which can be a time suck, so sometimes I still get my hands dirty too :)
I think this depends. I prefer the thinking bit, but it's quite difficult to think without the act of coding.
It's how white boarding or writing can help you think. Being in the code helps me think, allows me to experiment, uncover new learnings, and evolve my thinking in the process.
Though maybe we're talking about thinking of different things? Are you thinking in the sense of what a PM thinks about ? User features, user behavior, user edge cases, user metrics? Or do you mean thinking about what a developer thinks about, code clarity, code performance, code security, code modularization and ability to evolve, code testability, innovative algorithms, innovative data-structure, etc. ?
But the result of that thinking would hardly ever align neatly with whatever an LLM is doing. The only time it wouldn’t be working against me would be drafting boilerplate and scaffolding project repos, which I could already automate with more prosaic (and infinitely more efficient) solutions.
Even if it gets most of what I had in mind correct, the context switching between “creative thinking” and “corrective thinking” would be ruinous to my workflow.
I think the best case scenario in this industry will be workers getting empowered to use the tools that they feel work best for their approach, but the current mindset that AI is going to replace entire positions, and that individual devs should be 10x-ing their productivity is both short-sighted and counterproductive in my opinion.
Tools, libraries and platforms are accidental complexities. If you have already learned how to use them, you can avoid the pitfalls and go straight to the solution, which is why the common advice is to use boring technologies, as the solutions are widely documented and there are a lot of case studies.
If it's something new, then you can learn as you go by starting small and refactor as you're gaining more confidence. Copy-pasta or code generation is usually bad in that case. You don't know enough to judge the long-term costs.
Code is tech debt. When people talk about software engineering, it's to make sure that this debt doesn't outweigh the benefits of using the software.
What makes you regard this as an anti-AI take? To my mind, this is a very pro-AI take
Are you genuinely saying you never saw a critique of AI on environmental impact, or how it amplifies biases, or how it widens the economic gap, or how it further concentrates power in the hands of a few, or how it facilitates the dispersion of misinformation and surveillance, directly helping despots erode civil liberties? Or, or, or…
You don’t have to agree with any of those. You don’t even have to understand them. But to imply anti-AI arguments “hinge on the idea that technology forces people to be lazy/careless/thoughtless” is at best misinformed.
Go grab whatever your favourite LLM is and type “critiques of AI”. You’ll get your takes.
The energy cost is nonsensical unless you pin down a value out vs value in ratio and some would argue the output is highly valuable and the input cost is priced in.
I don't know if it will end up being a concentrated power. It seems like local/open LLMs will still be in the same ballpark. Despite the absurd amounts of money spent so far the moats don't seem that deep.
Baking in bias is a huge problem.
The genie is out of the bottle as far as people using it for bad. Your own usage won't change that.
> You don’t have to agree with any of those. You don’t even have to understand them. But to imply anti-AI arguments “hinge on the idea that technology forces people to be lazy/careless/thoughtless” is at best misinformed.
We can certainly discuss some of those points, but that’s not what is in question here. The OP is suggesting there is only one type of anti-AI argument they are familiar with and that they’d “love” to see something different. But I have to question how true that is considering the myriad of different arguments that exist and how easy they are to find.
At least when doing stuff the old way you learn something if you waste time.
That said AI is useful enough and some poker games are +EV.
So this is more caution-AI than anti-AI take. It is more an anti-vibe-koolaid take.
The thing is that, once you're experienced enough, it's faster to just glance at the code and have the answer right, instead of playing the guessing game with AI.
> and once it has the context you ask to think about implementing feature X
I'm always amazed at someone using that methodology. When I think about a feature, first is to understand the domain, second is which state I'm like to start from and where all the data are. If you don't get these two steps right, what you'll have is a buggy/incomplete implementation. And if you do get these steps right, the implementation is likely trivial.
Also trivial work can benefit the coder. Like a light jog between full sprints for your brain. Reviewing code can be more taxing than writing it as you need to retieve the full context at once instead of incremental steps.
I don't know... that seems like a false dichotomy to me. I think I could enjoy both but it depends on what kind of work. I did start using AI for one project recently: I do most of the thinking and planning, and for things that are enjoyable to implement I still write the majority of the code.
But for tests, build system integration, ...? Well that's usually very repetitive, low-entropy code that we've all seen a thousand times before. Usually not intellectually interesting, so why not outsource that to the AI.
And even for the planning part of a project there can be a lot of grunt work too. Haven't you had the frustrating experience of attempting a re-factoring and finding out midway it doesn't work because of some edge case. Sometimes the edge case is interesting and points to some deeper issue in the design, but sometimes not. Either way it sure would be nice to get a hint beforehand. Although in my experience AIs aren't at a stage to reason about such issues upfront --- no surprise since it's difficult for humans too --- of course it helps if your software has an oracle for if the attempted changes are correct, i.e. it is statically-typed and/or has thorough tests.
Because it still needs to be correct, and AI still is not producing correct code
My strong belief after almost twenty years of professional software development is that both us and LLMs should be following the order: build, test, reflect, plan, build.
Writing out the implementation is the process of materializing the requirements, and learning the domain. Once the first version is out, you can understand the limits and boundaries of the problem and then you can plan the production system.
This is very much in line with Fred Brooks' "build one to throw away" (written ~40 years ago in the "The Mythical Man-Month". While often quoted, if you never read his book, I urge you to do so, it's both entertaining, and enlightening on our software industry), startup culture (if you remove the "move fast break things" mantra), and governmental pilot programs (the original "minimum viable").
This argument is wearing a little thin at this point. I see it multiples times a day, rephrased a little bit.
The response, "How well do you think your thinking will go if you had not spent years doing the 'practice' part?", is always followed by either silence or a non-sequitor.
So, sure, keep focusing on the 'thinking' part, but your thinking will get more and more shallow without sufficient 'doing'
The important part that everyone glosses over is the "gaining experience" part.
The experience you gained writing code lead to you being tech lead / team lead /architect.
The experience you get from those roles, including "helping people get unstuck", makes you valuable because there are people involved, not just technology. IOW, that is different to the experience you get from prompting.
We have yet to see how valuable the experience from prompting will be. At this point the prompters are just guessing that their skills won't atrophy, and that their new experience won't be at the same level as vibe-coders who can't spell "Python".
As a fairly senior person myself, and an occasional user of LLMs, and someone who has tried CC in recent months, the experience I got from LLMs, while not nothing, was not recognised by me as valuable in any way - it basically put me at the same skill level as a vibe-coder.
OTOH, the experience I got mentoring very junior engineers the month before that I recognised as instantly valuable; at the end of it I had learned new strategies for dealing with people, growing them, etc.
The only "experience" you get with LLM is "put another coin into the slot and pull the lever again".
I relate it to directors on a production. It's certainly very valuable to know how to operate a camera, and especially to understand lighting, storytelling, etc. It gives you insight in how to work with the people who are actually doing these tasks. It helps you to know when someone is gaslighting you, etc.
That being said, it's kind of an insane statement to say that all a director does is pull a lever. I'm sure there are a ton of wannabe directors who try to do exactly that and proceed to fail miserably if they don't adapt quickly to reality. But having a great director is obviously a huge differentiator in output.
Do I think we'll have as many programmers in the future as we do today? Probably not. I think we're going to see a real decimation of coders, but at the same time we might (I say "might") see much greater overall production that may not otherwise exist from the less talented vibers or w/e ridiculously critical name you want. Some of that is certainly going to be interesting and maybe even radically game changing.
IMO our feelings about this are about as relevant as shaking our fist at the cosmos.
In my experience no. The agents get trapped by the exact programming pitfalls a junior would. The LLM is basically a 16 year old who read a given languages For Dummies book cover to cover 3-4x and has the syntax down but understands little about actually programming especially once you run into any real complexity. However 100% of those limitations can be overcome by proper architecture, testing, specification / requirements analysis (which is a lost art in the time of Agile but which I am a master of), and a sprinkle of technical strategic guidance. Especially the agent doesn’t understand its limitations so you need to have an eye for when it’s working on a problem that’s outside the competency its token window can produce. I could go on for 2 hours but bottom line is IMHO there’s more to it than this simple claim.
True. But the roles as such require you to do a lot of thinking by helping a LOT of people. You end up shuffling between multiple projects/deliverables. Here we are talking about probably a developer working on a single project/deliverable and then equating it to AI. Not to mention the easy to forget part is that by the time you are a tech lead / team lead / architect you have so many hours that you know some stuff like back of your hand.
My assembly has definitely rotted and I doubt I could do it again without some refreshing but it's been replaced with other higher-level skills, some which are general like using correct data structures and algorithms, and others that are more specific like knowing some pandas magic and React Flow basics.
I expect this iteration I'll get a lot better at systems design, UML, algorithm development, and other things that are slightly higher level. And probably reverse-engineering as well :) The computer engineering space is still vast IMHO....
Or perhaps you’re asking how people will become good at delegation without doing? I don’t know — have you been “doing” multiple years of assembly? If not, how are you any good at Python (or whatever language you currently use?). Probably you’d say you don’t need to think about assembly because it has been abstracted away from you. I think AI operates similarly by changing the level of abstraction you can think at.
People have argued for years that software architects must write code.
Regarding your second paragraph: When you write python you then debug it at the level of the abstraction. You never debug the python interpreter. You can try to treat AI like an abstraction but it immediately breaks down as soon as you go to debug. It would only be a complete abstraction if you never had to deal with the generated code.
Tech leads can exist in many variants, but usually they spend the majority of time in code, so they don't lose it. If they become too good at managing and change their priorities, they _will_ gradually drift away too.
Yes? If not 100% then a number pretty close to that. Definitely 100% of all the managers/non-coding leads I’ve worked with
Also I find the idea that most managers or technical leads are doing any kind of "deep thought" hilarious, but that's just maybe my apathy towards management speaking.
Yes, obviously this happens.
Do you seriously think that skills don't rust when you stop using them daily?
Here's a couple points which are related to each other:
1) LLMs are statistical models of text (code being text). They can only exist because huge for-profit companies ingested a lot of code under proprietary, permissive and copyleft licenses, most of which at the very least require attribution, some reserve rights of the authors, some give extra rights to users.
LLM training mixes and repurposes the work of human authors in a way which gives them plausible deniability against any single author, yet the output is clearly only possible because of the input. If you trained an LLM on only google's source code, you'd be sued by google and it would almost certainly reproduce snippets which can be tracked down to google's code. But by taking way, way more input data, the blender cuts them into such fine pieces that the source is undetectable, yet the output is clearly still based on the labor of other people who have not been paid.
Hell, GPT3 still produced verbatim snippets of inverse square root and probably other well known but licensed code. And github has a checkbox which scans for verbatim matches so you don't accidentally infringe copyright by using copilot in a way which is provable. Which means they take extra care to make it unprovable.
If I "write a book" by taking an existing book but replacing every word with a synonym, it's still plagiarism and copyright infringement. It doesn't matter if the mechanical transformation is way more sophisticated, the same rules should apply.
2) There's no opt out. I stopped writing open source over a year ago when it became clear all my code is unpaid labor for people who are much richer than me and are becoming richer at a pace I can't match through productive work because they own assets which give them passive income. And there's no license I can apply which will stop this. I am not alone. As someone said, "Open-Source has turned into a form of unpaid internship"[0]. It might lead to a complete death of open source because nobody will want to see their work fed into a money printing machine (subscription based LLM services) and get nothing in return for their work.
> But if you like the doing, the typing, fiddling with knobs and configs, etc etc, all AI does is take the good part away.
I see quite the opposite. For me, what makes programming fun is deeply understanding a problem and coming up with a correct, clear to understand, elegant solution. But most problems a working programmer has are just variations of what other programmers had. The remaining work is prompting the LLMs in the right way that they produce this (describing the problem instead of thinking about its solutions) and debugging bugs LLMs generated.
A colleague vibe coded a small utility. It's useful but it's broken is so many ways, the UI falls apart when some text gets too long, labels are slightly incorrect and misleading, some text handle decimal numbers in weird ways, etc. With manually written code, a programmer would get these right the right time. Potential bugs become obvious as you're writing the code because you are thinking about it. But they do not occur to someone prompting an LLM. Now I can either fix them manually which is time consuming and boring, or I can try prompting an LLM about every single one which is less time consuming but more boring and likely to break something else.
Most importantly, using an LLM does not give me deeper understanding of the problem or the solution, it keeps knowledge locked in a black box.
I honestly don't know how one can use Claude Code (or other AI agents) in a 'coding first thinking later' manner.
Imagine an AI as smart as some of the smartest humans, able to do everything they intellectually do but much faster, cheaper, 24/7 and in parallel.
Why would you spend any time thinking? All you'll be doing it is the things an AI can't do - 1) feeding it input from the real world and 2) trying out its output in the real world.
1) Could be finding customers, asking them to describe their problem, arranging meetings, driving to the customer's factory to measure stuff and take photos for the AI, etc.
2) Could be assembling the prototype, soldering, driving it to the customer's factory, signing off the invoice, etc.
None of that is what I as a programmer / engineer enjoy.
If actual human-level AI arrives, it'll do everything from concept to troubleshooting, except the parts where it needs presence in the physical world and human dexterity.
If actual human-level AI arrives, we'll become interfaces.
Consider leaving a disclaimer next time. Seems like you have a vested interest in the current half-baked generation of AI products succeeding
As someone writing lots of research code, I do get caught being careless on occasion since none of it needs to work beyond a proof of concept, but overall being able to just write out a spec and test an idea out in minutes instead of hours or days has probably made a lot of things exist that I'd otherwise never be arsed to bother with. LLMs have improved enough in the past year that I can easily 0-shot lots of ad-hoc visualization stuff or adapters or simple simulations, filters, etc. that work on the first try and with probably fewer bugs than I'd include in the first version myself. Saves me actual days and probably a carpal tunnel operation in the future.
I can guarantee this is happening in a lot of companies right now. Any company where there is pressure to use AI tools + pressure to get faster results it's inevitable.
— OSS exploded on the promise that software you voluntarily contributed to remains to benefit the public, and that a large corporation cannot tomorrow simply take your work and make it part of their product, never contributing anything back. Commercially operated LLMs threaten OSS both by laundering code and by overwhelming maintainers with massive, automatically produced and sometimes never read by a human patches and merge requests.
— Being able to claim that any creative work is merely a product of an LLM (which is a reality now for any new artist, copywriter, etc.) removes a large motivator for humans to do fully original creative work and is detrimental to creativity and innovation.
— The ends don’t justify the means, as a general philosophical argument. Large-scale IP theft had been instrumental at the beginning of this new wave of applied ML—and it is essentially piracy, except done by the powerful and wealthy against the rest of us, and for profit rather than entertainment. (They certainly had the money to license swaths of original works for training, yet they chose to scrape and abuse the legal ambiguity due to requisite laws not yet existing.)
— The plain old practical “it will drive more and more people out of jobs”.
— Getting everybody used to the idea that LLMs now mediate access to information increases inequality (making those in control of this tech and their investors richer and more influential, while pushing the rest—most of whom are victims of the aforementioned reverse piracy—down the wealth scale and often out of jobs) more than it levels the playing field.
— Diluting what humanity is. Behaving like a human is how we manifest our humanness to others, and how we deserve humane treatment from them; after entities that walk and talk exactly like a human would, yet which we can be completely inhumane to, become commonplace, I expect over time this treatment will carry over to how humans treat each other—the differentiator has been eliminated.
— It is becoming infeasible to operate open online communities due to bot traffic that now dwarves human traffic. (Like much of the above, this is not a point against LLMs as technology, but rather the way they have been trained and operated by large corporate/national entities—if an ordinary person wanted to self-host their own, they would simply not have the technical capability to cause disruption at this scale.)
This is just what I could recall off the top of my head.
Isn't "AI coding" trained almost entirely on open source code and published documentation?
… so I hope you can see why I don’t actually agree with your comment about who’s allowed to be a artist, and not just dismiss me as a glazer
You wanna be an artist? Put in the work.
I doubt anyone would consider it a problem if large-scale commercial LLM operators were required to respect licenses and negotiate appropriate usage terms. Okay, maybe with one exception: their investors and shareholders.
It is not! It's a very recent invention. Especially its application to creative works contradicts thousands of years of the development of human culture. Consider folk songs.
> I doubt anyone would consider it a problem if large-scale commercial LLM operators were required to respect licenses and negotiate appropriate usage terms. Okay, maybe with one exception: their investors and shareholders.
And the issue I'm gesturing at is that you run into different contradicting conclusions about how LLMs should interact with copyright depending on exactly what line of logic you follow, so the courts will never be able to resolve how it should work. These are issues can only be conclusively resolved with writing new laws to decide it's going to work, but that will eventually only make the contradictions worse and complicate the hoops that people will have to jump through as the technology evolves in new ways.
First, let’s note that creative work includes a lot more than just arts (crucially, invention).
In music, by your logic you may disagree with recognising song composition as IP, but you have to agree that being able to earn royalties from businesses playing your performance (even if it is a cover) serves as a proxy to people coming to listen and express their appreciation to a performer back when audio recording didn’t exist.
Also, let’s distinguish IP in general and its current legal implementation, such as protections lasting longer than author’s life. It should be noted that complexity in art did also grow since then, but it may or may not (I have no strong opinion here) make sense to grant the author post-humous protections.
> you run into different contradicting conclusions about how LLMs should interact with copyright depending on exactly what line of logic you follow, so the courts will never be able to resolve how it should work.
The courts can identify which ways of LLM use follow the spirit of the IP framework, encouraging innovation and creativity. As it is, current commercial LLMs slowly erode it, creating a feeling of “nothing belongs to anyone in particular, so why bother putting in the hard work”, profiting a minority of individuals while harming society over longer term. It is not difficult to see how applying the copyright as is could put an end to this, ensuring authors have control over their work, with the only consequence being slightly worse bottom lines at a handful of corporations with market caps the size of countries.
But I think programming is much more "how to use the building blocks" and mathematics than ingesting narratives and themes. More like ingesting a dictionary and thesaurus and then writing a book about vampires.
In my experience AI coding is not going to spew out a derivative of another project unless your objective is actually to build a derivative of that software. If your code doesn't do the same or look the same it doesn't really meet the criteria to be a derivative of someone else's.
I mostly use Cursor for writing test suites in Jest with TypeScript, these are so specific to my work I don't think it's possible they've infringed someone else's.
The objective of a for-profit corporation may well be to build a derivative of free software, benefitting from the work volunteer engineers put into it. Previously, if that software is GPL, it would imply a clean-room reimplementation of a massive codebase. With LLM laundering, relevant companies could as well simply claim “we got this from copilot” and they would be right (note that they don’t need to have used an LLM—the mere legality of these license-laundering LLMs means you can simply copy this from a GPL codebase and claim that an LLM output it due to its non-deterministic nature). This goes contrary to the promise of copyleft licenses that volunteer contributor work will remain to benefit the public and could not be expropriated this way, which led to OSS explosion in the first place.
I'm curious for more thoughts on "will drive more and more people out of jobs”. Isn't this the same for most advances in technology (e.g., steam engine, computers s, automated toll plazas, etc.). In some ways, it's motivation for making progress; you get rid of mundane jobs. The dream is that you free those people to do something more meaningful, but I'm not going to be that blindly optimistic :) still, I feel like "it's going to take jobs" is the weakest of arguments here.
Mundane job may be mundane (though note that it is sometimes subjective), but it earns someone bread and butter and it is always economic stress when the job is gone and many people have to retrain.
If we were to believe those of us who paint this technology as mind-bogglingly world-changing, that someone is now nearly everyone and unlike the previous time there is no list of jobs you could choose from (that would last longer than the time it takes to train).
If we were not to believe the hype, still: when those jobs got automated back then, people moved to jobs that are liable to be obsolete this time, except there is also just more people overall, so even purely in terms of numbers this seems to be a bigger event.
I actually end up using LLMs in the planning phase more often than the writing phase. Cursor is super good at finding relevant bits of code in unfamiliar projects, showing me what kind of conventions and libraries are being used, etc.
It does not! If you're using interactive IDE AI, you spend your time keeping the AI on the rails, and reminding it what the original task is. If you're using agents, then you're delegating all the the mid-level/tactical thinking, and perhaps even the planning, and you're left with the task of writing requirements granular enough for an intern to tackle, but this hews closer to "Business Analyst" than "Software Engineer"
New-fangled compiled languages...
Or who use modern, strictly-typed languages.
New-fangled type-safe languages...
As someone that has been coding since it was wiring up NAND gates on a circuit board, I'm all for the new ways, but there will definitely be a lot of mistakes, jargon, and blind alleys; just like every other big advancement.
You can do this in Python, or you can do this in English. But at the end of the day the engineer must input the same information to get the same behavior. Maybe LLMs make this a bit more efficient but even in English it is extremely hard to give exact specification without ambiguity (maybe even harder than Python in some cases).
I'm not impressed by AI because it generates slop. Copilot can't write a thorough working test suite to save it's life. I think we need a design and test paradigm to properly communicate with AI for it to build great software.
My experience with low level systems programming is that it’s like working with a developer who is tremendously enthusiastic but has little skill and little understanding of what they do or don’t understand. Time I would have spent writing code is replaced by time spent picking through code that looks superficially good but is often missing key concepts. That may count as “thinking” but I wouldn’t categorize it as the good kind.
Where it excels for me is as a superpowered search (asking it to find places where we play a particular bit-packing game with a particular type of pointer works great and saves a lot of time) and for writing one-off helper scripts. I haven’t found it useful for writing code I’m going to ship, but for stuff that won’t ship it can be a big help.
It’s kind of like an excavator. If you need to move a bunch of dirt from A to B then it’s great. If you need to move a small amount of dirt around buried power lines and water mains, it’s going to cause more trouble than it’s worth.
It's also been my experience that AI will speed up the easy / menial stuff. But that's just not the stuff that takes up most of my time in the first place.
Yet every time that someone here earnestly testifies to whatever slight but real use they’ve found of AI, an army of commentators appears ready to gaslight them into doubting themselves, always citing that study meant to have proven that any apparent usefulness of AI is an illusion.
At this point, even just considering the domain of programming, there’s more than enough testimony to the contrary. This doesn’t say anything about whether there’s an AI bubble or overhype or anything about its social function or future. But, as you note, it means these cardboard cutout critiques of AI need to at least start from where we are.
I don't think anyone's saying that about technology in general. Many safety-oriented technologies force people to be more careful, not less. The argument is that this technology leads people to be careless.
Personally, my concerns don't have much to do with "the part of coding I enjoy." I enjoy architecture more than rote typing, and if I had a direct way to impose my intent upon code, I'd use it. The trouble is that chatbot interfaces are an indirect and imperfect vector for intent, and when I've used them for high-level code construction, I find my line-by-line understanding of the code quickly slips away from the mental model I'm working with, leaving me with unstable foundations.
I could slow down and review it line-by-line, picking all the nits, but that moves against the grain of the tool. The giddy "10x" feeling of AI-assisted coding encourages slippage between granular implementation and high-level understanding. In fact, thinking less about the concrete elements of your implementation is the whole advantage touted by advocates of chatbot coding workflows. But this gap in understanding causes problems down the line.
Good automation behaves in extremely consistent and predictable ways, such that we only need to understand the high-level invariants before focusing our attention elsewhere. With good automation, safety and correctness are the path of least resistance.
Chatbot codegen draws your attention away without providing those guarantees, demanding best practices that encourage manually checking everything. Safety and correctness are the path of most resistance.
> The argument is that this technology leads people to be careless.
We can't optimize for an objective truth when that objective truth doesn't exist. So while doing our best to align our models they must simultaneously optimize they ability to deceive us. There's little to no training in that loop where outputs are deeply scrutinized, because we can't scale that type of evaluation. We end up rewarding models that are incorrect in their output.
We don't optimize for correctness, we optimize for the appearance of correctness. We can't confuse the two.
The result is: when LLMs make errors, those errors are difficult for humans you detect.
This results in a fundamentally dangerous tool, does it not? Tools that when they error or fail they do so safely and loudly. Instead this one fails silently. That doesn't mean you shouldn't use the tool but that you need to do so with an abundance of caution.
> I could slow down and review it line-by-line, picking all the nits, but that moves against the grain of the tool.
That's the opposite of lazy, so hopefully answers OP.
As an engineer and researcher, I prefer lies (models, simplifications), that are known to me, rather than unknown unknowns.
I don't need to know exact implementation details, knowledge of aggregate benchmarks, fault rates and tolerances is enough. A model is a nice to have.
This approach works, in science (physics, chemistry, biology, ...) and in engineering (including engineering agentic and social sustems- social engineering).
> As an engineer and researcher, I prefer lies (models, simplifications), that are known to me, rather than unknown unknowns.
I'll make a corollary to help:
~> There's a simple fact: humans prefer lies that they believe are truths over lies that they do know are lies.
As a researcher myself, yes, I do also prefer known falsehoods over unknown falsehoods but we can only do this from a metaphysical perspective. If I'm aware of an unknown then it is, by definition, not an unknown unknown.
How do you preference a falsehood which you cannot identify as a falsehood?
How do you preference an unknown which you do not know is unknown?
We have strategies like skepticism to deal with this help with this but this doesn't make the problem go away. It ends up with "everything looks right, but I'm suspicious". Digging in can be very fruitful but is more frequently a waste of time for the same reason: if a mistake exists we have not identified the mistake as a mistake!
> I don't need to know exact implementation details, knowledge of aggregate benchmarks, fault rates and tolerances is enough.
Evidence can only bound your confidence in the truth of some claim but it cannot prove it. The dual to this is a much simpler problem, as disproving a claim can be done with a singular example. This distinction often isn't as consequential in engineering as there are usually other sources of error that are much larger.
As an example, we all (hopefully) know that you can't prove the correctness of a program through testing. It's a non-exhaustive process. BUT we test because it bounds our confidence about its correctness and we usually write cases to disprove certain unintended behaviors. You could go through the effort to prove correctness but this is a monumental task and usually not worth the effort.
But right now we're talking about a foundational problem and such a distinction matters here. We can't resolve the limits of methods like RLHF without considering this problem. It's quite possible that there's no way around this limitation since there are no objective truths the majority of tasks we give LLMs. If that's true then the consequence is that a known unknown is "there are unknown unknowns". And like you, I'm not a fan of unknown unknowns.
We don't actually know the fault rates nor tolerances. Benchmarks do not give that to us in the general setting (where we apply our tools). This is a very different case than, say, understanding the performance metrics and tolerances of an o-ring. That part is highly constrained and you're not going to have a good idea of how well it'll perform as a spring, despite those tests having a lot of related information.
Additionally, why is it that whenever I mention success stories accomplished with the help of LLMs, people rush to say "does not count because it is not >100k LOC". Why does it not count, why should it not count? I would have written it by hand, but I finished much faster with the help of an LLM. These are genuine projects that solve real problems. Not every significant project has to have >100k LOC. I think we have a misunderstanding of the term "significant".
> nobody in this or any meaningful software engineering discussion is talking about software projects that are 1000, or even 10000, SLoC.
Why?
> these are trivial and uninteresting sizes.
In terms of what exactly?
Because small programs are really quick and easy to write, there was never a bottleneck making them and the demand for people to write small programs is very small.
The difficulty of writing a program scales super linearly with size, an experienced programmer in his current environment easily writes a 500 line program in a day, but writing 500 meaningful lines to an existing 100k line codebase in a day is not easy at all. So almost all developer time in the world is spent making large programs, small programs is a drop in an ocean and automating that doesn't make a big difference overall.
Small programs can help you a lot, but that doesn't replace programmers since almost no programmers are hired to write small programs, instead automatically making such small programs mostly helps replace other tasks like regular white collar workers etc whose jobs are now easier to automate.
It's entirely possible to have a 100k LOC system be made up of effective a couple hundred 500 line programs that are composed together to great effect.
That's incredibly rare but I did once work for a company who had such a system and it was a dream to work in. I have to think AIs are making a massive impact there.
> It's entirely possible to have a 100k LOC system be made up of effective a couple hundred 500 line programs that are composed together to great effect.
You may also wish to look at UNIX Philosophy. The idea that programs should be small and focused. A program should do one thing and do it well. But there's a generalization to this philosophy when you realize a function is a program.
I do agree there's a lot of issues with design these days but I think you've vastly oversimplified the problem.
To me, this sounds like an nightmare—I'm sure anyone who's worked at a shop with way too many microservices would agree. It's trivial to right-click a function call and jump to its definition; much harder to trace through your service mesh and find out what, exactly, is running at `load-balancer.kube.internal:8080/api`.
> but writing 500 meaningful lines to an existing 100k line codebase in a day is not easy at all.
Not because the single line was hard to write but because the context in which it needed to be written.
Typing was never the bottleneck and I'm not sure why this is the main argument for LLMs (e.g. "LLMs save me from the boilerplate). When typing is a bottleneck it seems like it's more likely that the procedure is wrong. Things like libraries, scripts, and skeletons tend to be far better solutions for those problems. In tough cases abstraction can be extremely powerful, but abstraction is a difficult tool to wield.
The bottleneck is the thinking and analyzing.
My feelings exactly.
LLM code generation (at least, the sort where people claim they're being 10X-ed) feels like it competes with frameworks. "An agent built this generic CRUD webapp on its own with only 30 minutes of input from me!"—well, I built an equivalent webapp in 30 minutes with Django. These are off-the-shelf solutions to solved problems. Yes, a framework like Django requires up-front learning, but in the end it leaves you with fewer lines of code to maintain, as opposed to custom-generated LLM code.
> I have written a 1000 LOC program in C using an LLM.
> I have reviewed it "line by line" though, I do not know why I would not do this.
But it's pretty rare you're going to be able to review every line in a mature project, even if you're developing that project. Those can contain hundreds or even thousands of files with hundreds (hopefully not thousands) of LOC. While it's possible to review every line it's pretty costly in time and it's harder since the code is changing as you're doing this...
Think of it this way, did you also review all the lines of code in all the libraries you used? Why not? The reasoning will be pretty similar. This isn't to say we shouldn't spend more time exploring the code we work with nor that we likely wouldn't benefit from this, but that time is a scarce resource. So the problem is when the LLM is churning out code faster than you can review.
While coding you are hopefully also debugging and thinking. By handing coding over to the LLM you decouple this. So you reduce your time writing lines of code but increase time spent debugging and analyzing. There will be times where this provides gains but IME this doesn't happen in serious code. But yeah, my quick and dirty scripts can be churned out a lot faster. That saves time, but not 10x. At least not for me
I am just asking. Everyone says 1k LOC is nothing, yet they want to replace 1k LOC in C with 1k LOC in Rust. You can do it in day. You are a professional!
Or what is your point? That 1k LOC projects are useless or pointless? Because if so, I seriously beg to differ.
> So the problem is when the LLM is churning out code faster than you can review.
I start small, I can review just fine.
> when people talk about safety, it does matter in Rust, right?
> Everyone says 1k LOC is nothing
> Or what is your point?
I said in another comment[1] that I've spent hours or days to write /one line/ of code, or even partial. Does that mean I was doing a bad job? Was I just slacking off? I think this is something many developers have experienced. Were we all lazy? Dumb?
I'd argue that you can't answer that question from the information alone. Sometimes a single line of code is crazy hard to figure out. If you haven't seen this before, allow me to introduce you to some old coding lore[2]
//When I wrote this, only God and I understood what I was doing
//Now, God only knows
The point here is that it is really hard to measure the quality and effectiveness of a programmer. The context of the problem is not something that can be abstracted away when evaluating them. Unfortunately, this means to evaluate them you also need to be an expert programmer AND have meaningful context to understand the specific problems they are working on. There's not a thing you can do from a spreadsheet. The truth is that if you optimize from the spreadsheet you'll only introduce more Jira tickets. There's a joke that there's 2 types of 10x programmers. The other one is the programmer that introduces 100x the jira tickets while completing them 10x as fast. The problem is that that programmer doesn't see the bigger scope and makes mistakes that leads to new tickets. This might be like your new rockstar junior dev. They fill out tickets but are solving the problems in isolation, not in context of the codebase. This leads to more complexity and bugs later on, but that lag in effect is hard to measure/identify so it is easy to think they are a rockstar but actually a problem.
> I start small, I can review just fine.
So the point is that everything is highly context driven. That's what matters. You need nuance and care. It is not easy to say what makes good code or even identify it. So... LGTM
[0] https://github.com/uutils/coreutils/issues/7092
[1] https://news.ycombinator.com/item?id=45409430
[2] https://stackoverflow.com/questions/184618/what-is-the-best-...
[3] https://betterexplained.com/articles/understanding-quakes-fa...
[4] https://news.ycombinator.com/item?id=45060059
https://news.ycombinator.com/item?id=44416817
And I am expert programmer (I would like to believe) and I use LLMs just to get some refresher of my options and whatnot, and I choose where the project goes, with my knowledge. All my prompts are very specific, which requires knowledge.
Pretty clearly that’s not the divide anyone’s talking about, right?
Your argument should maybe be something about thinking about the details vs thinking about the higher level. (If you were to make that argument, my response would be: both are valuable and important. You can only go so far working at one level. There are certainly problems that can be solved at one level, but also ones that can’t.)
> Just as tech leads don't just write code but set practices for the team, engineers now need to set practices for AI agents. That means bringing AI into every stage of the lifecycle
The technology doesn't force people to be careless, but it does make it very easy to be careless, without having to pay the costs of that carelessness until later.
I think this might simply be how the human brain works. Take autonomous driving as an example: while the car drives on its own the human driver is supposed to be alert and step in if needed. But does that work? Or will the driver's mind wander off because the car has been driving properly for the last half hour? My gut feeling is that it's inevitable that we'll eventually just shut out everything that goes smoothly and by the time it doesn't it might be too late.
We are not that different from our ancestors who used to roam the forests, trying to eat before they get eaten. In such an environment there is constantly something going on, some critters crawling, some leaves rustling, some water flowing. It would drive us crazy if we could not shut out all this regular noise. It's only when an irregularity appears that our attention must spring into action. When the leaves rustle differently than they are supposed to there is a good chance that there is some prey or a predator to be found. This mechanism only works if we are alert. The sounds of the forest are never exactly the same, so there is constant stimulation to keep up on our toes. But if you are relaxing in your shelter the tension is gone.
My fear is that AI is too good, to the point where it makes us feel like being in our shelter rather than in the forest.
Yes. Productivity accelerates at an exponential rate, right up until it drives off a cliff (figuratively or literally).
I view the story of LLMs akin to the Concorde. Something catastrophic will happen that will be too big to ignore and all trust will implode.
What about if the "knowing/understanding" bit is your favorite part?
It’s not force but simply human nature. We invent tools to do less. That’s the whole point of tools.
AI isn't a technology. (No more than asking your classmate to do your homework for you is a "technology".)
Please don't conflate between AI and programming tools. AI isn't a tool, it is an oracle. There's a huge fundamental gap here that cannot be bridged.
Which industries are those? What does that mastery look like?
> There's a divide between people ...
No, there is not. If one is not willing to figure out a couple of ffmpeg flags, comb through k8s controller code to see what is possible and fix that booting error in their VMs then failure in "mental experiences" is certain.
The most successful people I have met in this profession are the ones who absolutely do not tolerate magic and need to know what happens from the moment they press the ON on their machine, till the moment they turn is OFF again.
Though I will also dogpile on the "thankless tasks" remark and say that the stuff that I have AI blast through is very thankless. I do not enjoy changing 20 different files to account for a change in struct definition.
Why would I want a fuzzy, vague, imprecise, up-to-interpretation programming language? I already have to struggle with that in documentation, specifications, peers and – of course – myself. Why would I take the one precise component and make it suffer from the same?
This contrasts of course with tasks such as search, where I'm not quite able to precisely express what I want. Here I find LLMs to be a fantastic advance. Same for e.g. operations between imprecise domains, like between natural languages.
Does this divide between "physical" and "mental" exist? Programming languages are formal languages that allow you to precisely and unambiguously express your ideas. I would say that "fiddling" with the code (as you say) is a kind of mental activity.
If there is actually someone out there that only dislikes AI coding assistants because they enjoy the physical act of typing and now have to do less of it (I have not seen this blog post yet), then I might understand your point.
Eh, physical and mental isn't the divide — it's more like people who enjoy code itself as a craft and people who simply see it as a means to an end (the application). Much like a writer might labor over their prose (the code) while telling a story (the application). Writing code is far more than the physical act of typing to those people.
Not forces, encourages.
A dev that spends an undue amount of time fiddling with knobs and configs probably sucks. Their mind isn't on the problem that needs to be solved.
Yeah I'm actually quite surprised that so many people are just telling AI to do X without actually developing a maintainable plan to do so first. It's no wonder that so many people are anti-vibe-coding — it's because their exposure to vibe coding is just telling Replit or Claude Code to do X.
I still do most of my development in my head, but I have a go-to prompt I ask Claude code when I'm stuck: "without writing any code, and maintaining existing patterns, tell me how to do X." it'd spit out some stuff, I'd converse with it to make sure it is a feasible solution that would work long term, then I tell it to execute the plan. But the process still starts in my head, not with a prompt.
First, skilled engineers using LLMs to code also think and discuss and stare off into space before the source code starts getting laid down. In fact: I do a lot, lot more thinking and balancing different designs and getting a macro sense of where I'm going, because that's usually what it takes to get an LLM agent to build something decent. But now that pondering and planning gets recorded and distilled into a design document, something I definitely didn't have the discipline to deliver dependably before LLM agents.
Most of my initial prompts to agents start with "DO NOT WRITE ANY CODE YET."
Second, this idea that LLMs are like junior developers that can't learn anything. First, no they're not. Early-career developers are human beings. LLMs are tools. But the more general argument here is that there's compounding value to working with an early-career developer and there isn't with an LLM. That seems false: the LLM may not be learning anything, but I am. I use these tools much more effectively now than I did 3 months ago. I think we're in the very early stages of figuring how to get good product out of them. That's obvious compounding value.
I like asking for the plan of action first, what does it think to do before actually do any edits/file touching.
But there’s a guiding light that both the LLM and I can reference.
Regardless of that, personally i'd really like it if they could actually learn from interacting with them. From a user's perspective what i'd like to do is to be able to "save" the discussion/session/chat/whatever, with everything the LLM learned so far, to a file. Then later be able to restore it and have the LLM "relearn" whatever is in it. Now, you can already do this with various frontend UIs, but the important part in what i'd want is that a) this "relearn" should not affect the current context window (TBH i'd like that entire concept to be gone but that is another aspect) and b) it should not be some sort of lossy relearning that loses information.
There are some solutions but there are all band-aids to fundamental issues. For example you can occasionally summarize whatever discussed so far and restart the discussion. But obviously that is just some sort of lossy memory compression (i do not care that humans can do the same, LLMs are software running on computers, not humans). Or you could use some sort of RAG but AFAIK this works via "prompt triggering" - i.e. only via your "current" interaction, so even if the knowledge is in there but whatever you are doing now wouldn't trigger its index the LLM will be oblivious to it.
What i want is, e.g., if i tell to the LLM that there is some function `foo` used to barfize moo objects, then go on and tell it other stuff way beyond whatever context length it has, save the discussion or whatever, restore it next day, go on and tell it other stuff, then ask it about joining splarfers, it should be able to tell me that i can join splarfers by converting them to barfized moo objects even if i haven't mentioned anything about moo objects or barfization since my previous session yesterday.
(also as a sidenote, this sort of memory save/load should be explicit since i'd want to be able to start from clean slate - but this sort of clean slate should be because i want to, not as a workaround to the technology's limitations)
Models don't have memory, and they don't have understanding or intelligence beyond what they learned in training.
You give them some text (as context), and they predict what should come after (as the answer).
They’re trained to predict over some context size, and what makes them good is that they learn to model relationships across that context in many dimensions. A word in the middle can affect the probability of a word at the end.
If you insanely scale the training and inference to handle massive contexts, which is currently far too expensive, you run into another problem: the model can’t reliably tell which parts of that huge context are relevant. Irrelevant or weakly related tokens dilute the signal and bias it in the wrong direction, the distribution flatten or just ends up in the wrong place.
That's why you have to make sure you give it relevant well attended context, aka, context engineering.
It won't be able to look at a 100kloc code base and figure out what's relevant to the problem at hand, and what is irrelevant. You have to do that part yourself.
Or what some people do, is you can try to automate that part a little as well by using another model to go research and build that context. That's where people say the research->plan->build loop. And it's best to keep to small tasks, otherwise the context needing for a big task will be too big.
Basically, yes. I know the way LLMs currently work wouldn't be able to provide what i want, but what i want is a different way that does :-P (perhaps not even using LLMs).
It seems to work quite well and I'll often be pleasantly surprised when Claude retrieves some useful background I've stored, and seems to magically "know what I'm talking about".
Not perfect by any means and I think what you're describing is maybe a little more fundamental than bolting on a janky database to the model - but it does seem better than nothing.
TBH i'm not even sure if that is possible with LLMs, especially in a way that does not rely on using the context.
haha I always do that. I think it's a good way to have some control and understand what it is doing before the regurgitation. I don't like to write code but I love the problem solving/logic/integrations part.
Before I switched to a different agent, I routinely ran into situations where I would say "write a plan to do x", it would start planning, and I would steer it by saying something like "update such and such a file, instead of other-file" and it would immediately update it, even though it was in plan mode.
Then I would point out "you're in plan mode, don't update files", and it would go absolutely ham undoing changes and furiously apologizing "You're right! I'm in plan mode, let me undo those changes I wasn't supposed to make!" - meaning that now it's broken the rules twice.
Plan mode does not disable any writing tools, it just changes the system prompt, judging by my experience anyway.
What have you figured out so far, apart from explicit up-front design?
Yes, and the thinking time is a significant part of overall software delivery, which is why accelerating the coding part doesn't dramatically change overall productivity or labor requirements.
In practice this can’t happen because 30 minutes into coding you will find something that nobody thought about.
Sure you can. Mapping out the unknowns (and then having a plan to make each one knowable) is the single most important function of whoever you have designing your architecture.
Up-front architecture isn't about some all-knowing deity proclaiming the perfect architecture from on high. It's an exercise in risk management, just like any other engineering task.
and
> isn't about some all-knowing deity
Seems like a big conflict of your own thoughts.
Or as they say....
> there are known knowns; there are things we know we know. We also know there are known unknowns; that is to say we know there are some things we do not know. But there are also unknown unknowns—the ones we don't know we don't know.
And pretty much your job as an engineer is to mitigate the risk that poses.
We don't build new architectures in a vacuum. We build on what has worked before in similar situations, and we adapt it to the problem at hand.
That adaptation is an ongoing process - but it's not the same as saying "fuck it, let's vibe-code our architecture"
If their job is basically to generate code to close jira tickets I can see the appeal of LLMs.
Honestly, even if I did it that way and then threw it all away and wrote the whole thing manually it'd be worth using. Obviously I don't, because once I've figured out how to scope and coach to get the right result it'd be silly to throw it away, but the same value derives from that step regardless of how you follow it up.
I really like that on IntelliJ I have to approve all changes, so this prompt is unnecessary.
There's a YOLO mode that just changes shit without approval, that I never use. I wonder if anyone does.
Even Claude Code lets you approve each change, but it's already writing code according to a plan that you reviewed and approved.
This makes it move with much less scattered interactions from me, which allows focus time on other tasks. And the committing parts make it easier for me to review what it did just like I would review a feature branch created by a junior colleague.
If it's done and tests pass I'll create a pull request (assigned to myself) from the feature branch. Then thoroughly review it fully, this really requires discipline. And then let Claude fetch the pull request comments from the Github API and fix them. Again as a longer run that allows me to do other things.
YOLO-mode is helpful for me, because it allows Claude to run for 30 minutes with no oversight which allows me to have a meeting or work on something else. If it requires input or approval every 2 minutes you're not async but essentially spending all your time watching it run.
With tools you know ahead of time that they will do the job you expect them to do with very high probability, or fail (with low probability) in some obvious way. With LLMs, there are few tasks you can trust them to do, and you also don't know their failure mode. They can fail yet report success. They work like neither humans nor tools.
An LLM behaves like a highly buggy compiler that too frequently reports success while emitting incorrect code. Not knowing where the bugs are, the only thing you can try to do is write the program in some equivalent way but with different syntax, hoping you won't trigger a bug. That is not a tool programmers often use. Learning to work with such a compiler is a skill, but it's unclear how transferable or lasting that skill is.
If LLMs advance as significantly and as quickly as some believe they will, it may be better to just wait for the buggy compiler to be fixed (or largely fixed). Presumably, much less skill will be required to achieve the same result that requires more skill today.
Copilot has Ask mode, and GPT-5 Codex has Plan/Chat mode for this specific task. They won't change any files. I've been using Codex for a couple of days and it's very good if you give it plenty of guidance.
It's super impressive in my opinion, but if you think it's going to straight up replace humans right now, I think you probably aren't a software developer in the trenches cranking out features.
I'm sort of a Neanderthal when it comes to understanding AI, but I don't think AI in it's current form works like a human. Right now, it kind of just cranks out all the code in one fell swoop. A human on the other hand works more iteratively. You write a little bit of code then you run it and look at an iPhone simulator, look at Figma designs, and see if you're getting closer to what you want. AI doesn't appear to know how to iterate, run code, look at designs, and debug things. I imagine in 100 years it will know how to do all that stuff though. And who knows, maybe in 1 year it will be able to do that. But as of right now, September 28th, 2025 it can't do that yet.
I personally find AI code gen most useful for one-off throwaway code where I have zero intent to learn. I imagine this means that the opposite end of the spectrum where learning is maximized is one where the AI doesn't generate any code for me.
I'm sure there are some people for which the "AI-Driven Engineering" approach would be beneficial, but at least for me I find that replacing those AI coding blocks with just writing the code myself is much more enjoyable, and thus more sustainable to actually delivering something at the end.
[1] https://youtu.be/apREl0KmTdQ?t=4751 (relevant section is about 5 minutes long)
Obviously you have to work to learn, but to me this is a bit like saying learning is maximized when you never talk to anyone or ask for help — too strong.
On the other hand you have AI which, out of the box, seems content to go along with anything and will happily write code for me. And I've never seen it have a single insight on the same level as my teammates. All of which is to say, AI doesn't really feel like something you can properly "ask" something. It's especially far away from that when it's just generating code and nothing else.
I think it boils down to personal preference where some people want to use AI while others don't. I also learn when coding with my AI agent. I learn about using the tool more effectively. As someone who has been coding for 10 years, I find more pleasure in AI assisted coding.
But aside from taste, the product and the business don't care about what I like. It's about shipping quality updates more quickly. And while there might be some tension in saying this, I'm convinced that I can do that much more quickly in AI assisted coding.
That thinking does result in some big tradeoffs…I generally get better results faster but I also have a less complete understanding of my code.
But the notion that Claude Code means an experienced developer spends less thinking carefully is simply wrong. It’s possible (even likely) that a lot of people are using agents poorly…but that isn’t necessarily the agent’s fault.
1) not all coding is the same. You might be working on a production system. I might need a proof of concept
2) not everyone's use of the coding agents is the same
3) developer time, especially good developer time has a cost too
I would like to see an article that frames the tradeoffs of AI assisted coding. Specifically without assigning value judgments (ie goodness or badness). Really hard when your identity is built around writing code.
I have been working in machine-learning for 10 years. I am tired of the computer. I am tired of working. I just want to lay in the grass.
So give them some context. I like Cline's memory bank approach https://docs.cline.bot/prompting/cline-memory-bank which includes the architecture, progress, road map etc. Some of my more complex projects use 30k tokens just on this, with the memory bank built from existing docs and stuff I told the model along the way. Too much context can make models worse but overall it's a fair tradeoff - it maintains my coding style and architecture decisions pretty well.
I also recommend in each session using Plan mode to get to a design you are happy with before generating any code.
Later, you'll realize that there are no new senior coders to burn out, because you got rid of the junior coders, but we're not there yet.
The other thing that makes a big difference is using a language that has strong typing and a strict compiler. You can allow claude to call the complier, analysers, debuggers and so forth and it works amazing to keep code quality consistent if you tell it always follow your existing patterns.
Lastly, I figured out recently, quite by accident, that claude can call adb/logcat on a connected android device, to ingest real debug logs while my app is running, which can be a firehose of logs. It cuts through it like candyfloss and spots runtime errors significantly faster than a human can read all those log. It found a bug for me that was nearly impossible to find with the IDE & breakpoints, just by ingesting the debug/logcats outputs directly.
So overall the article is great, I do think the way we work can and will fundamentally change in the coming years.
Using Plan mode in Cline or other agent based workflows is day and night in the outputs.
Alas, at least in Cline it seems plan mode doesn’t read files just works off context which is insane to me and hinders its usefulness, anyone know why that happens?
Agreed. My tool agnostic workflow is to work up a planning/spec doc in the docs/feature-plans folder. I use one chat thread to make that. First it creates the basic plan, then then we pick it apart together, I manually fix bad assumptions, then in a new chat, we implement.
Before and after implementation, I run my /gilfoyle command for a constructive roast, then my /sec command for a thorough security review. After implementing this, and a bit more, the final LLM output quality is much higher.
edit: adding "make sure we are applying known patterns used in our app for our solution, don't reinvent the wheel." helped a ton. I am on mobile atm, and that might not be the exact wording.
Does including an agent at each stage of this cycle mean "context engineering"? Is this then just more text and assets to feed in at each stage of LLM ussage to provide the context for the next set of tokens to generate for the next stage of the cycle? Is there something deeper that can be done to encode this level of staged development into the agent's weights/"understanding"? Is there an established process for this yet?
- Specification
- Documentation
- Modular Design
- Test-Driven Development
- Coding Standard
- Monitoring & Introspection
The fact then when we read others' code, we don't remember/integrate it into our thinking as well as we do when we're the authors. So mentoring "AI Juniors" provides less growth then doing the job, esp. if it is mostly corrective actions.
I’m personally going to keep myself sharp, only using LLMs very very sparingly.
Once the planning is done, the actual coding is very fast. The human review that follows is again slow, often also leading to minor new tickets.
I can have Claude Code bang out everything from boilerplate to a working prototype to a complex algorithm embedded in a very complex and confusing code base. It’s not correct 100% of the time but it’s pretty damn close. And often times it comes up with algorithms I would have never thought of initially.
These things are at least a 10x multiple of my time.
I'd love for what you say to be possible. Comments like yours often cause me to take another crack at agentic workflows. I'm disappointed every time.
You are just not a true scotsman.
And when they link something is one of those repos full of sloop and no code.
I am becoming paranoid and wonder how many people here can even code.
I want them to be right and me wrong! Please someone show I am wrong.
I don't view LLMs as ways of foregoing the responsibility of writing code and rather see it as my "really smart keyboard". With enough context priming and a well structured codebase I no longer need to spend time writing each line of code and can have Claude do it in a fraction of the time.
I need to start a blog sooner rather than later as I don't agree with the article nor the naysayers. Maybe a year ago I'd say that it's not possible to code with LLM agents. However ever since Cursor's release, LLMs have completely changed my workflow.
obviously good and experienced engineers aren't going to be vibe coders/mollycoddlers by nature. but many good and experienced engineers will be pressured to make poor decisions by impatient business leaders. and that's the root of most AI anxiety: we all know it's going to be used as irresponsibly and recklessly as possible. it's not about the tech. it's about a system with broken incentives.
Which looks to attempt to give better "coding" understanding to the model instead of mere tokens and positioning and hence improve the coding capabilities of these "brilliant but unpredictable junior engineer" coding agents:
- https://ai.meta.com/research/publications/cwm-an-open-weight...
But this is the case also with human teams. Plenty of human teams cannot understand "the full context of an application [sometimes ever]" -- same with contractors who roll on and off teams or entire contracting firms who roll on and off.
For vibe coding you need systems thinking, planning and logic, but less craftmansship.
For PO's the chart looks different, here the traditional flow contains: "polish concepts, make mocks, make stories, dailies, handovers, revisions, waiting for devs busy with other things"
The vibing PO has none of that.
Not saying this is sustainable for big projects already, but it is for ever growing "small" projects (especially if it's a techincal PM that can code a bit). It's just so much faster without devs -_-.
Disclaimer: I am such a PO. What I now wonder: how can I mix vibe coding and properly developed foundations (ai assisted, but not 99% vibed). One answer in my view is splitting services in vibecoded and core, but dependencies on the core already slow you down a lot. Curious to hear how others mix and match both.
What I actually do right now:
- fullstack PoC vibecoded
- specifications for core based on PoC findings
- Build proper vibecoded V1 mocking what moves to core. But here already a more structured vibecoding approach too.
- replace mocks with actual core
Not yet done but planned: build throwaway PoC in a branch of the core so I can also vibe on top of the core directly, including modification of the core.
Or you are just writing the same React components every day.
AI has been insanely helpful for a couple of reasons, but not for writing code.
I discuss my plans with AI and check if I missed out on edge cases I develop to the most common conventions, I ask AI for better naming, ideas for abstractions, to explain me frameworks, find me packages.
But the lease couple of days I’m having sanity problems because there is one AI tool after the other coming up with a new plan on how to use AI efficiently but once you’ve solved a problem is it really much work to type it out?
It shouldn’t be too much code if correctly using orchestration.
It’s almost like all the beautiful conventions and patterns are forgotten.
Good Code is most of the time less code. Yes imho most important is readability or digestibility or even diagnosability like all of these things get completely forgotten when we start talking about AI and I see AI far off in writing such code. Good frameworks and packages will be faster.
AI is a good sparring partner and encyclopaedia especially living in terminal but I just don’t see it should writing a lot of code (except maybe landing pages with css animations) that don’t live in a complex application
People without a true coding background get stuck after a single function because the context window is still so narrow, and code assistants do not get the full picture of the project. It reminds me of the offshore developer's work ethic: "But sir, you told me to add this button there like that, so I deleted the entire codebase". Without thinking about why in general. It keeps saying, "You are absolutely right! I shouldn't have done that!" I just like working with a crappy coder from Fiverr or another freelancer site.
I do see different behavior whether planning or exploration comes first though.
I think that this approach can already get us pretty far. One thing I'm missing is tooling to make it easier to build automation on top of, eg, Claude Code, but I'm sure it's going to come (and I'm tempted to try vibe coding it; if only I had the time).
I will say I always provide an initial context document about the feature/system, to avoid us starting with trivial questions. After about 45minutes I’ll often feel I’ve covered enough ground and given the problem enough thought to really put pen to paper. Off the back of this I’ll ask it to summarise the spec and produce a document. This can be a good point to ditch AI if you are so inclined but still get value from it.
Most recently I was struggling to get an authentication setup working. I spent at least an hour combing through the code looking for the mistake. The issue turned out to be that the VM I was working on had a broken ipv6 configuration.
Like when those in the know could clearly see the internet's path to consuming everything but it just hasn't happened yet so there were countless articles trying to decide if it was a fad or not, a waste of money, bad investment, etc.
Now it works, not like before, gets old after a while.
Just giving it an MCP to test changes also didn't work for me. But the combination with e2e tests was better.
Other than that, I keep hearing the same arguments - "LLMs free up more time for me to think about the 'important things'." Son, your system is not durable, your tests are misleading, and you can't reason about what's happening because you didn't write it. What important things are left to think about??
You just described the majority of pre-LLM enterprise software.
Why didn't you just not accept the change?
1. Write up a detailed spec as a prompt, including a bunch of detail about the actual functionality. Add in some details about wanting to test everything. At this point I did not specifically say "Hey LLM, in my tests, do not mock one of the key idempotency functions in this system". This is obvious to me since 1/4 of the system was concerned with idempotency, and this was super clear in my original spec
2. LLM works for 30 minutes, spits out 2.5K lines
3. I review the product functionality and the test cases at a glance, and then start running tests and iterating on the software.
If I have to mentally split every line of the LLM output and decide whether it's correct or not (remember, I didn't write the rest of the system), then I lose all the productivity gains that the LLM offers. You don't do this, no one does this unless they're modifying a few lines at a time.
My point is that you cannot build robust software from scratch with LLMs right now. They walk right into footguns we've spent decades learning about and building solutions for.
Traditional Coding: [ Thinking & Coding ....................... | Fixing ]
AI-Assisted Coding: [ Coding | Thinking & Fixing ................ ]
Fixing something you have written is much easier than fixing something someone else (or an AI) has written, just because you don't have the mental model for the code, which is the most important part of debugging and refactoring.
LLMs are more like a lightning fast pseudo-random text generator at this point. To see that though you'd need to run it a few times, test it and understand the output. Something that's a bit beyond the casual amateur ability. Maybe what we need is the AI bubble to burst first.
Its sad that we've gotten here though. Every tried having a coffee with your preferred AI as opposed to doing it with a highly motivate junior? Maybe you're not even going to get a chance to experience this.
With classic coding the feedback loop is very straigth forward and fast. More so with and IDE, you get instant feedback as you write.
With LLMs the loop is not as clear.
These chores are exactly what AI is best at
My one concrete pushback to the article is that it states the inevitable end result of vibe coding is a messy unmaintainable codebase. This is empirically not true. At this point I have many vibecoded projects that are quite complex but work perfectly. Most of these are for my private use but two of them serve in a live production context. It goes without saying that not only do these projects work, but they were accomplished 100x faster than I could have done by hand.
Do I also have vibecoded projects that went of the rails? Of course. I had to build those to learn where the edges of the model’s capabilities are, and what its failure modes are, so I can compensate. Vibecoding a good codebase is a skill. I know how to vibecode a good, maintainable codebase. Perhaps this violates your definition of vibecoding; my definition is that I almost never need to actually look at the code. I am just serving as a very hands-on manager. (Though I can look at the code if I need to - have 20 years of coding experience. But if I find that I need to look at the code, something has already gone badly wrong.)
Relevant anecdote: A couple of years ago I had a friend who was incredibly skilled at getting image models to do things that serious people asserted image models definitely couldn’t do at the time. At that time there were no image models that could get consistent text to appear in the image, but my friend could always get exactly the text you wanted. His prompts were themselves incredible works of art and engineering, directly grabbing hold of the fundamental control knobs of the model that most users are fumbling at.
Here’s the thing: any one of us can now make an image that is better than anything he was making at the time. Better compositionality, better understanding of intent, better text accuracy. We do this out of the box and without any attention paid to promoting voodoo at all. The models simply got that much better.
In a year or two, my carefully cultivated expertise around vibecoding will be irrelevant. You will get results like mine by just telling the model what you want. I assert this with high confidence. This is not disappointing to me, because I will be taking full advantage of the bleeding edge of capabilities throughout that period of time. Much like my friend, I don’t want to be good at managing AIs, I want to realize my vision.
It worked fine until recently where I will ask it to tweak some behavior of a command with a flag and it does a diff with like hundreds of lines. So now it's struggling to catch every place it needs to change some hardcoded duplicate values it decided to copy/paste into two dozen random places in the code.
To be fair it is doing a decent job unfucking it now that I noticed and started explicitly showing it how ridiculously cumbersome and unmaintainable it made things with specific examples and refactoring. But if I hadn't bothered to finally sit down and read through it thoroughly it would have just become more broken and inconsistent as it grew exponentially.
In most cases I would never have undertaken those projects at all without AI. One of the projects that is currently live and making me money took about 1 working day with Claude Code. It’s not something I ever would have started without Claude Code, because I know I wouldn’t have the time for it. I have built websites of similar complexity in the past, and since they were free-time type endeavors, they never quite crossed the finish line into commerciality even after several years of on-again-off-again work. So how do you account that with a time multiplier? 100x? Infinite speedup? The counterfactual is a world where the product doesn’t exist at all.
This is where most of the “speedup” happens. It’s more a speedup in overall effectiveness than raw “coding speed.” Another example is a web API for which I was able to very quickly release comprehensive client side SDKs in multiple languages. This is exactly the kind of deterministic boilerplate work LLMs are ideal for, and that would take a human a lot of typing, and looking up details for unfamiliar languages. How long would it have taken me to write SDKs in all those languages by hand? I don’t really know, I simply wouldn’t have done it, I would have just done one SDK in Python and said good enough.
If you really twist my arm and ask me to estimate the speedup on some task that I would have done either way, then yeah I still think a 100x speedup is the right order of magnitude, if we’re talking about Claude Code with Opus 4.1 specifically. In the past I spent about a five years very carefully building a suite of tools for managing my simulation work and serving as a pre/post-processor. Obviously this wasn’t full-time work on the code itself, but the development progressed across that timeframe. I recently threw all that out and replaced it with stuff I rebuilt in about a week with AI. In this case I was leveraging a lot of the learnings I gleaned from the first time I built it, so it’s not a fair one-to-one comparison, but you’re really never going to see a pure natural experiment for this sort of thing.
I think most people are in a professional position where they are sort of externally rate limited. They can’t imaging being 100x more effective. There would be no point to it. In many cases they already sit around doing nothing all day, because they are waiting for other people or processes. I’m lucky to not be in such a position. There’s always somewhere I can apply energy and see results, and so AI acts as an increasingly dramatic multiplier. This is a subtle but crucial point: if you never try to use AI in a way that would even hypothetically result in a big productivity multiplier (doing things you wouldn’t have otherwise done, doing a much more thorough job on the things you need to do, and trying to intentionally speed up your work on core tasks) then you can’t possibly know what the speedup factor is. People end up sounding like a medieval peasant suddenly getting access to a motorcycle and complaining that it doesn’t get them to the market faster, and then you find out that they never actually ride it.
I wonder, have you sat down and tried to vibecode something with Claude Code? If so, what kind of multiplier would you find plausible?
I’ve found LLMs just as useful for the "thankless" layers (e.g. tests, docs, deployment).
The real failure mode is letting AI flood the repo with half-baked abstractions without a playbook. It's helpful to have the model review the existing code and plan out the approach before writing any new code.
The leverage may be in using LLMs more systematically across the lifecycle, including the grunt work the author says remains human-only.
It's also great for things that aren't creative, like 'implement a unit test framework using google test and cmake, but don't actually write the tests yet'. That type of thing saves me hours and hours. It's something I rarely do, so it's not like I just start editing my cmake and test files, I'd be looking up documentation, and a lot of code that is necessary, but takes a lot of time.
With LLMs, I usually get what I want quickly. If it's not what I want, a bit of time reviewing what it did and where it went wrong usually tells me what I need to give it a better prompt.
All software engineering is left aside in exchange for coding -> fixing -> coding -> fixing.
Current AI models, especially small ones, are quite smooth-brain by themselves.
But, if we can refine them to a point where thinking AI systems can run on smart glasses and smartwatches in a usable, stable state, we can assume by then larger AI systems with higher compute are bound to be much more capable.
We can claim AGI milestone then. And, the real race for superintelligence begins.
I’ve found this concept trips CC up—- assertions are backwards, confusing comments in the test, etc. Just starting a prompt with “Use TDD to…” really helps.
I will raise that LLMs are pretty good at some of the non-coding tasks too.
eg. "I'm currently creating an AI for a turn based board game. Without doing any implementation, create a plan for the steps that need to be done including training off real world game data".
The LLM creates a tasklist for iterative steps to accomplish the above. It usually needs correction specific to the business/game needs but it's a great start and i recommend doing this just so the LLM has a doc with context on what its trying to achieve in a bigger picture as you have it complete tasks.
Article being discussed in this thread isn't intended to be a luddite rejection of AI. It's just a mistake I see people keep making (and have made myself) and some thoughts on how to avoid it with the tools we have today.
They can code your startup...
Like the LLM was a person, lol.
Like for example nobody says:
They cars can be fast.
People usually say:
The cars are fast.
Or sometimes:
Them cars are fast.
But never:
They cars are fast.
Now, if you say the problem is that you don't want to be inflicted with other people's AI code, just enforce more meaningful tests. There has always been bad code in the past, and there always will be.
I, for one, am doing great with AI coding - and my feeling is that more emphasis on project structure is the way forward for better end to end results.
You might be a little bit out of touch with the current zeitgeist -- and what's happening in a lot of corporations.
I cannot express how tired I am of seeing this beyond stupid take.
If you truly believe that, you have either only ever worked with the most piss-poor junior engineers, or you simply have never worked with junior engineers.
LLMs do not learn, LLMs do not ask clarifications, LLMs do not wonder if they're going in the wrong direction, LLMs do not have taste, LLMs do not have opinions, LLMs write absolutely illogical nonsense, LLMs do not ask themselves what is the best for the customer, LLMs have no context beyond what you have explicitely fed them for that one specific task, and much more.
This is the fundamental issue with LLM's imo.
LLMs aren't effective when used this way.
You still have to think.
IMO a vibe coder who is speaking their ideas to an agent which implements them is going to have way more time to think than a hand coder who is spending 80% of their time editing text.
This man should probably be a poet.
The assumption he is making is that the tools are good, and that the tools will eventually be used by everyone. So the competitive advantage is time limited.
He is largely correct, although I think productivity gains are overblown.