(thankfully market dynamics and OSS alternatives will probably stop this but it's not a guarantee, you need like at least six viable firms before you usually see competitive behavior)
Now, imagine a scenario of a typical SWE in todays or maybe not-so-distant future: the agents build your software, you simply a gate-keeper/prompt engineer, all tests pass, you're now doing a production deployment at 12am and something happens but your agents are down. At that point, what do you do if you haven't build or even deployed the system? You're like a L1 support at this point, pretty useless and clueless when it comes to fully understanding and supporting the application .
So you know what do, what I've been doing for about a decade, if the internet goes down? I stop working. And over that time I've worked in many places around the world, developing countries, tropical islands, small huts on remote mountains. And I've lost maybe a day of work because of connectivity issues. I've been deep in a rainforest during a monsoon and still had 4g connection.
If Anthropic goes down I can switch to Gemini. If I run out of credits (people use credits? I only use a monthly subscription) then I can find enough free credits around to get some basic work done. Increasingly, I could run a local model that would be good enough for some things and that'll become even better in the future. So no, I don't think these are any kind of valid arguments. Everyone relies on online services for their work these days, for banking, messaging, office work, etc. If there's some kind of catastrophe that breaks this, we're all screwed, not just the coders who rely on LLMs.
I am genuinely curious about your work lifestyle.
The freedom to travel anywhere while working sounds awesome.
The ability to work anywhere while traveling sounds less so.
Those still have limits, no? Or if there's a subscription that provides limitless access, please tell me which one it is.
I've tried Anthropic's Max plan before, but hit limits after just a couple of hours, same with Google's stuff, but wasn't doing anything radically different when I tried those, compared with Codex, so seems other's limits are way lower.
If you tell me "I lost internet at home and couldn't work there", it's one thing. But that you simply went about a month without internet connection, I find it hard to believe.
Hell, on Tuesday I lost ~2 hours because Starlink was having some issue. When it came up I was on a different ground station and getting very low speeds. Not such a big deal except you never get that time back.
In 2022, funny enough I was at an AWS office (I worked remotely when I worked there) working in ProServe, us-east-1 was having issues that was affecting everything, guess what we all did? Stopped working, the world didn’t come to an end.
Even now that I work from home, on the rare occasions that Internet goes down, I just use my phone if I need to take a Zoom call.
You're an engineer, your goal is to figure stuff out using the best tools in front of you
Humans are resilient, they reliably perform (and throw great parties) in all sorts of chaotic conditions. Perhaps the thing that separates us most from AI is our ability to bring out our best selves when baseline conditions worsen
Actually, the last thing you probably want is somebody reverting back to doing things the way we did them 20 years ago and creating a big mess. Much easier to just declare an outage and deal with it properly according to some emergency plan (you do have one, right?).
CI/CD are relatively new actually. I remember doing that stuff by hand. I.e. I compiled our system on my Desktop system, created a zip file, and then me and our operations department would use an ISDN line to upload the zip file to the server and "deploy" it by unzipping it and restarting the server. That's only 23 years ago. We had a Hudson server somewhere but it had no access to our customer infrastructure. There was no cloud.
I can still do that stuff if I need to (and I sometimes do ;-) ). But I wouldn't dream of messing with a modern production setup like that. We have CI/CD for a reason. What if CI/CD were to break? I'd fix it rather than adding to the problem by panicking and doing things manually.
Take a look at how ridiculously much money is invested in these tools and the companies behind them. Those investments expect a return somehow.
https://boto3.amazonaws.com/v1/documentation/api/latest/inde...
I don’t need to comprehend “the library”. I need to know what I need to do and then look up the API call.
Your teachers had the right goal, but a bad argument. Learning arithmetic isn't just about being able to do a calculation. It's about getting your brain comfortable with math. If you always have to pull out a goddamn calculator, you'll be extremely limited.
Trust me, elementary-age me was dumb to not listen to those teachers and to become so calculator-dependent.
IF I was totally dependent on it, I would be in trouble. Fortunately I am not.
You are at least a decade late to post fears about developers reliance on the internet. It was complete well before the LLM era
Before you go on about kids these days, my first time coding was on an Apple //e in assembly.
I'll happily optimize my life for 99.999% of the time.
If the Internet is down for a long time, I've got bigger problems anyway. Like finding food.
This is why I suggest developers use the free time they gain back writing documentation for their software (preferably in your own words not just AI slop), reading official docs, sharpening your sword, learning design patterns more thoroughly. The more you know about the code / how to code, the more you can guide the model to pick a better route for a solution.
Apply to anything else: you could eat out at restaurants every night, and it would do a great job in feeding you! Think of all the productivity you would gain relying on agential chefs. With restaurants even I can eat like a French chef, they have truly democratized food. And they do a perfect job these days executing dishes, only some mistakes.
What happens when github goes down. You shrug and take a long lunch.
* all services are run at a loss and they increase price to the point the corp doesn’t want to pay for everyone any more.
* it turns out that our chats are used for corporate espionage and the corps get spooked and cut access
* some dispute between EU and US happens and they cut our access.
The solution’s having EU and local models.
Your pizza restaurant is all wonderful and all but what happens when the continual supply of power to the freezer breaks? How will you run your restaurant then?
i would work on the hundreds of non-coding tasks that i need to do. or just not work?
what do you do when github actions goes down?
The loss of competency seems pretty obvious but it's good to have data. What is also interesting to me is that the AI assisted group accomplished the task a bit faster but it wasn't statistically significant. Which seems to align with other findings that AI can make you 'feel' like you're working faster but that perception isn't always matched by the reality. So you're trading learning and eroding competency for a productivity boost which isn't always there.
This is up there with believing tobacco companies health "research" from the 30s, 40s, 50s, 60s, 70s, 80s, and 90s.
I wonder if we're going to have a future where the juniors never gain the skills and experience to work well by themselves, and instead become entirely reliant on AI, assuming that's the only way
That's not what the study says. It says that most users reflect your statement while there is a smaller % that benefits and learns more and faster.
Generalizations are extremely dangerous.
What the article says simply reflect that most people don't care that much and default to the path of least resistance, which is common every day knowledge, but we very well know this does not apply to everyone.
> Among participants who use AI, we find a stark divide in skill formation outcomes between high-scoring interaction patterns (65%-86% quiz score) vs low-scoring interaction patterns (24%-39% quiz score). The high scorers only asked AI conceptual questions instead of code generation or asked for explanations to accompany generated code; these usage patterns demonstrate a high level of cognitive engagement.
This is very much my experience. AI is incredibly useful as a personal tutor
Not much as a tutor.
My hypothesis is that the AI users gained less in coding skill, but improved in spec/requirement writing skills.
But there’s no data, so it’s just my speculation. Intuitively, I think AI is shifting entry level programmers to focus on expressing requirements clearly, which may not be all that bad of a thing.
We're definitely getting better at writing specs. The issue is the labor bottleneck is competent senior engineers, not juniors, not PMs, not box-and-arrow staff engineers.
> I think AI is shifting entry level programmers to focus on expressing requirements clearly
This is what the TDD advocates were saying years ago.
Dramatically improved Jira usage -- better, more descriptive tickets with actionable user stories and clearly expressed requirements. Dramatically improved github PRs. Dramatically improved test coverage. Dramatically improved documentation, not just in code but in comments.
Basically all _for free_, while at the same time probably doubling or tripling our pace at closing issues, including some issues in our backlog that had lingered for months because they were annoying and nobody felt like working on them, but were easy for claude to knock out.
Well, yeah. You were still (presumably) debugging the code you did write in the higher level language.
The linked article makes it very clear that the largest decline was in problem solving (debugging). The juniors starting with AI today are most definitely not going to do that problem-solving on their own.
One of my advantages(?) when it comes to using AI is that I've been the "debugger of last resort" for other people's code for over 20 years now. I've found and fixed compiler code generation bugs that were breaking application code. I'm used to working in teams and to delegating lots of code creation to teammates.
And frankly, I've reached a point where I don't want to be an expert in the JavaScript ORM of the month. It will fall out of fashion in 2 years anyway. And if it suddenly breaks in old code, I'll learn what I need to fix it. In the meantime, I need to know enough to code review it, and to thoroughly understand any potential security issues. That's it. Similarly, I just had Claude convert a bunch of Rust projects from anyhow to miette, and I definitely couldn't pass a quiz on miette. I'm OK with this.
I still develop deep expertise in brand new stuff, but I do so strategically. Does it offer a lot of leverage? Will people still be using it on greenfield projects next year? Then I'm going to learn it.
So at the current state of tech, Claude basically allows me to spend my learning strategically. I know the basics cold, and I learn the new stuff that matters.
I'd kinda like to see this measured. It's obviously not the assembly that matters for nine-9s of jobs. (I used assembly language exactly one time in my career, and that was three lines of inline in 2003.) But you develop a certain set of problem-solving skills when you code assembly. I speculate, like with most problem-solving skills, it has an impact on your overall ability and performance. Put another way, I assert nobody is worse for having learned it, so the only remaining question is, is it neutral?
> everyone's acting like all human coders are better than all AI's
I feel like the sentiment here on HN is that LLMs are better than all novices. But human coders with actual logical and architectural skills are better than LLMs. Even the super-duper AI enthusiasts talk about controlling hoards of LLMs doing their bidding--not the other way around.
I found that Claude wasn't too great at first at it and returned a lot of hallucinated methods or methods that existed in Pandas but not Polars. I chalk this up to context blurring and that there's probably a lot less Polars code in the training corpus.
I found it most useful for quickly pointing me to the right documentation, where I'd learn the right implementation and then use it. It was terrible for the code, but helpful as a glorified doc search.
The models are too good now. One thing I've noticed recently is that I've stopped dreaming about tough problems, be it code or math. The greatest feeling in the world is pounding your head against a problem for a couple of days and waking up the next morning with the solution sketched out in your mind.
I don't think the solution is to be going full natty with things, but to work more alongside the code in an editor, rather than doing things in CLI.
The amount of context switching in my day-to-day work has become insane. There's this culture of “everyone should be able to do everything” (within reason, sure), but in practice it means a data scientist is expected to touch infra code if needed.
Underneath it all is an unspoken assumption that people will just lean on LLMs to make this work.
I also used to get great pleasure from the banging head and then the sudden revelation.
But that takes time. I was valuable when there was no other option. Now? Why would someone wait when an answer is just a prompt away.
They can give plausible architecture but most of the time it’s not usable if you’re starting from scratch.
When you design the system, you’re an architect not a coder, so I see no difference between handing the design to agents or other developers, you’ve done the heavy lifting.
In that perspective, I find LLMs quite useful for learning. But instead of coding, I find myself in long sessions back and forth to ask questions, requesting examples, sequence diagrams .. etc to visualise the final product.
It is a pattern matching problem and that seems to me to be something AI is/will be particularly good at.
Maybe it won’t be the perfect architecture, or the most efficient implementation. But that doesn’t seem to have stopped many companies before.
And how much better than palantir given that musk is a bigot, attempts to buy elections for fascists, meddles in foreign democracies to push far right extremist narratives, used his wealth to steal very sensitive data from government agencies, does Nazi salutes, trains his LLM to be racist...
> bigot
> fascist
> far right extremist
> nazi
> racist
(just pulled a few words from your small comment)
It’s not virtue signaling to say the guy throwing around nazi salutes is in fact a nazi.
Good one
My first thought was that I can abstract what I wrote yesterday, which was a variation of what I built over the previous week. My second thought was a physiological response of fear that today is going to be a hard hyper focus day full of frustration, and that the coding agents that built this will not be able to build a modular, clean abstraction. That was followed by weighing whether it is better to have multiple one off solutions, or to manually create the abstraction myself.
I agree with you 100 percent that the poor performance of models like GPT 4 introduced some kind of regularization in the human in loop coding process.
Nonetheless, we live in a world of competition, and the people who develop techniques that give them an edge will succeed. There is a video about the evolution of technique in the high jump, the Western Roll, the Straddle Technique, and finally the Fosbury Flop. Using coding agents will be like this too.
I am working with 150 GB of time series data. There are certain pain points that need to be mitigated. For example, a different LLM model has to be coerced into analyzing or working with the data from a completely different approach in order to validate. That means instead of being 4x faster, each iteration is 4x faster, and it needs to be done twice, so it still is only 2x faster. I burned $400 in tokens in January. This cannot be good for the environment.
Timezone handling always has to be validated manually. Every exploration of the data is a train and test split. Here is the thing that hurts the most. The AI coding agents always show the top test results, not the test results of the top train results. Rather than tell me a model has no significant results, it will hide that and only present the winning outliers, which is misleading and, like the OP research suggests, very dangerous.
A lot of people are going to get burned before the techniques to mitigate this are developed.
Overfitting has always been a problem when working with data. Just because the barrier of entry for time series work is much lower does not mean that people developing the skill, whether using old school tools like ARIMA manually or having AI do the work, escape the problem of overfitting. The models will always show the happy, successful looking results.
Just like calculators are used when teaching higher math at the secondary level so basic arithmetic does not slow the process of learning math skills, AI will be used in teaching too. What we are doing is confusing techniques that have not been developed yet with not being able to acquire skills. I wrack and challenge my brain every day solving these problems. As millions of other software engineers do as well, the patterns will emerge and later become the skills taught in schools.
For hobby projects though, it's awesome. It just really struggles to do things right in the big codebase at work.
And then you find out someone else had already solved it. So might as well use the Google 2.0 aka ChatGPT.
And I must admit my appetite in learning new technologies has lessened dramatically in the past decade; to be fair, it gets to a point that most new ideas are just rehashing of older ones. When you know half a dozen programming languages or web frameworks, the next one takes you a couple hours to get comfortable with.
I'm a bit younger (33) but you'd be surprised how fast it comes back. I hadn't touched x86 assembly for probably 10 years at one point. Then someone asked a question in a modding community for an ancient game and after spending a few hours it mostly came back to me.
I'm sure if you had to reverse engineer some win32 applications, it'd come back quickly.
That's a skill onto itself, and I mean the general stuff does not fade or at least come back quickly. But there's a lot of the tail end that's just difficult to recall because it's obscure.
How exactly did I hook Delphi apps' TForm handling system instead of breakpointing GetWindowTextA and friends? I mean... I just cannot remember. It wasn't super easy either.
These last few months, however, I've had to spend a lot of time debugging via disassembly for my work. It felt really slow at first, but then it came back to me and now it's really natural again.
If you’ve forgotten your Win32 reverse engineering skills I’m guessing you haven’t done much of that in a long time.
That said, it’s hard to truly forget something once you’ve learned it. If you had to start doing it again today, you’d learn it much faster this time than the first.
For what it’s worth—it’s not entirely clear that this is true: https://en.wikipedia.org/wiki/Hyperthymesia
The human brain seemingly has the capability to remember (virtually?) infinite amounts of information. It’s just that most of us… don’t.
Compression/algorithms don't save you here either. The algorithm for pi is very short, pulling up any particular randomm digit of pi still requires the expenditure of some particular amount of entropy.
The important question is can you learn enough in a standard human lifetime to "fill up your knowledge bank"?
Ok, so my statement is essentially correct.
Most of us can not keep infinite information in our brain.
If you moved back to a country you hadn't lived or spoken its language in for 10 years, you would find yourself that you don't have to relearn it, and it would come back quickly.
Also information is supposedly almost infinite, as with increased efficiency as you learn, it makes volume limits redundant.
I’m not sure if this is in the Wikipedia article, but when I last read about this, years ago, there seemed to be a link between Hyperthymesia and OCD. Brain scans suggested the key was in how these individuals organize the information in their brain, so that it’s easy for them retrieve.
Before the printing press was common, it was common for scholars to memorize entire books. I absolutely cannot do this. When technology made memorization less necessary, our memories shrank. Actually shrank, not merely changing what facts to focus on.
And to be clear, I would never advocate going back to the middle ages! But we did lose something.
We can “store” infinite numbers by using our numeral system as a generator of sorts for whatever the next number must be without actually having to remember infinite numbers, but I do not believe it would be physically possible to literally remember every item in some infinite set.
Sure, maybe we’ve gotten lazy about memorizing things and our true capacity is higher (maybe very much so), but there is still some limit.
Additionally, the practical limit will be very different for different people. Our brains are not all the same.
Think about how we talk about exercise. Yes, there probably is a theoretical limit to how fast any human could run, and maybe Olympic athletes are close to that, but most of us aren’t. Also, if you want your arms to get stronger, it isn’t bad to also exercise your legs; your leg muscles don’t somehow pull strength away from your arm muscles.
No, but the limiting factor is the amount of stored energy available in your body. You could exhaust your energy stores using only your legs and left barely able to use your arms (or anything else).
If we’ve offloaded our memory capacity to external means of rapid recall (ex. the internet) then what have we gained in response? Breadth of knowledge? Increased reasoning abilities? More energy for other kinds of mental work? Because there’s no cheating thermodynamics, even thinking uses energy. Or are we just simply radiating away that unused energy as heat and wasting that potential?
It has two huge benefits: nearly infinite memory for truly interesting stuff and still looking friendly to people who tell me the same stuff all the times.
Side-effect: my wife is not always happy that I forgot about "non-interesting" stuff which are still important ;-)
2) Hyperthymesia is about remembering specific events in your past, not about retaining conceptual knowledge.
APL inventor says that he was developing not a programming language, but notation to express as much problems as one can. He found that expressing more and more problems with the notation first made notation grow, then notation size started to shrink.
To develop conceptual knowledge (when one's "notation" starts to shrink) one has to have some good memory (re-expressing more and more problems).
You can't model systems in your mind using past experiences, at least not reliably and repeatedly.
> When you know half a dozen programming languages or web frameworks, the next one takes you a couple hours to get comfortable with.
Also, relational algebra implements content-addressable storage (CAS), which is essential for data flow computing paradigm. Thus, you will have a window into CPU design.
At 54 (36 years of professional experience) I find these rondos fascinating.
I use remnote for that.
I write cards and quizzes for all kind of stuff, and I tend to retain it for years after having it practiced with the low friction of spaced repetition.
People naturally try to use what they've learned but sometimes end up making things more complicated than they really needed to be. It's a regular problem even excluding the people intentionally over-complicating things for their resume to get higher paying jobs.
One take-away for us from that viewpoint was that knowledge in fact is more important than the lines of code in the repo. We'd rather lose the source code than the knowledge of our workers, so to speak.
Another point is that when you use consultants, you get lines of codes, whereas the consultancy company ends up with the knowledge!
... And so on.
So, I wholeheartedly agree that programming is learning!
Isn't this the opposite of how large tech companies operate? They can churn develops in/out very quickly, hire-to-fire, etc... but the code base lives on. There is little incentive to keep institutional knowledge. The incentives are PRs pushed and value landed.
Isn't large amounts of required institutional knowledge typically a problem?
We had domain specialists with decades of experience and knowledge, and we looked at our developers as the "glue" between domain knowledge and computation (modelling, planning and optimization software).
You can try to make this glue have little knowledge, or lots of knowledge. We chose the latter and it worked well for us.
But I was only in that one company, so I can't really tell.
I could have sworn I was meant to be shipping all this time...
Common example here is learning a language. Say, you learn French or Spanish throughout your school years or on Duolingo. But unless you're lucky enough to be amazing with language skills, if you don't actually use it, you will hit a wall eventually. And similarly if you stop using language that you already know - it will slowly degrade over time.
> "We collect self-reported familiarity with AI coding tools, but we do not actually measure differences in prompting techniques."
Many people drive cars without being able to explain how cars work. Or use devices like that. Or interact with people who's thinking they can't explain. Society works like that, it is functional, does not work by full understanding. We need to develop the functional part not the full understanding part. We can write C without knowing the machine code.
You can often recognize a wrong note without being able to play the piece, spot a logical fallacy without being able to construct the valid argument yourself, catch a translation error with much less fluency than producing the translation would require. We need discriminative competence, not generative.
For years I maintained a library for formatting dates and numbers (prices, ints, ids, phones), it was a pile of regex but I maintained hundreds of test cases for each type of parsing. And as new edge cases appeared, I added them to my tests, and iterated to keep the score high. I don't fully understand my own library, it emerged by scar accumulation. I mean, yes I can explain any line, but why these regexes in this order is a data dependent explanation I don't have anymore, all my edits run in loop with tests and my PRs are sent only when the score is good.
Correctness was never grounded in understanding the implementation. Correctness was grounded in the test suite.
But the fundamentals all cars behave the same way all the time. Imagine running a courier company where sometimes the vehicles take a random left turn.
> Or interact with people who's thinking they can't explain
Sure but they trust those service providers because they are reliable . And the reason that they are reliable is that the service providers can explain their own thinking to themselves. Otherwise their business would be chaos and nobody would trust them.
How you approached your library was practical given the use case. But can you imagine writing a compiler like this? Or writing an industrial automation system? Not only would it be unreliable but it would be extremely slow. It's much faster to deal with something that has a consistent model that attempts to distill the essence of the problem, rather than patching on hack by hack in response to failed test after failed test.
I think being a programmer is closer to being an aircraft pilot than a car driver.
But isn't the corrections of those errors that are valuable to society and get us a job?
People can tell they found a bug or give a description about what they want from a software, yet it requires skills to fix the bugs and to build software. Though LLMs can speedup the process, expert human judgment is still required.
If you know that you need O(n) "contains" checks and O(1) retrieval for items, for a given order of magnitude, it feels like you've all the pieces of the puzzle needed to make sure you keep the LLM on the straight and narrow, even if you didn't know off the top of your head that you should choose ArrayList.
Or if you know that string manipulation might be memory intensive so you write automated tests around it for your order of magnitude, it probably doesn't really matter if you didn't know to choose StringBuilder.
That feels different to e.g. not knowing the difference between an array list and linked list (or the concept of time/space complexity) in the first place.
When it comes to fundamentals, I think it's still worth the investment.
To paraphrase, "months of prompting can save weeks of learning".
Tests only cover cases you already know to look for. In my experience, many important edge cases are discovered by reading the implementation and noticing hidden assumptions or unintended interactions.
When something goes wrong, understanding why almost always requires looking at the code, and that understanding is what informs better tests.
Instead, just learning concepts with AI and then using HI (Human Intelligence) & AI to solve the problem at hand—by going through code line by line and writing tests - is a better approach productivity-, correctness-, efficiency-, and skill-wise.
I can only think of LLMs as fast typists with some domain knowledge.
Like typists of government/legal documents who know how to format documents but cannot practice law. Likewise, LLMs are code typists who can write good/decent/bad code but cannot practice software engineering - we need, and will need, a human for that.
> AI assistance produces significant productivity gains across professional domains, particularly for novice workers. Yet how this assistance affects the development of skills required to effectively supervise AI remains unclear. Novice workers who rely heavily on AI to complete unfamiliar tasks may compromise their own skill acquisition in the process. We conduct randomized experiments to study how developers gained mastery of a new asynchronous programming library with and without the assistance of AI. We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average. Participants who fully delegated coding tasks showed some productivity improvements, but at the cost of learning the library. We identify six distinct AI interaction patterns, three of which involve cognitive engagement and preserve learning outcomes even when participants receive AI assistance. Our findings suggest that AI-enhanced productivity is not a shortcut to competence and AI assistance should be carefully adopted into workflows to preserve skill formation -- particularly in safety-critical domains.
I assistance produces significant productivity gains across professional domains, particularly for novice workers.
We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average.
Are the two sentences talking about non-overlapping domains? Is there an important distinction between productivity and efficiency gains? Does one focus on novice users and one on experienced ones? Admittedly did not read the paper yet, might be clearer than the abstract.
The research question is: "Although the use of AI tools may improve productivity for these engineers, would they also inhibit skill formation? More specifically, does an AI-assisted task completion workflow prevent engineers from gaining in-depth knowledge about the tools used to complete these tasks?" This hopefully makes the distinction more clear.
So you can say "this product helps novice workers complete tasks more efficiently, regardless of domain" while also saying "unfortunately, they remain stupid." The introductiory lit review/context setting cites prior studies to establish "ok coders complete tasks efficiently with this product." But then they say, "our study finds that they can't answer questions." They have to say "earlier studies find that there were productivity gains" in order to say "do these gains extend to other skills? Maybe not!"
I learned a lot more in a short amount of time than I would've stumbling around on my own.
Afaik its been known for a long time that the most effective way of learning a new skill, is to get private tutoring from an expert.
But that's what "impairs learning" means.
Like the architecture work and making good quality specs, working on code has a guiding effect on the coding agents. So in a way, it also benefits to clarify items that may be more ambiguous in the spec. If I write some of the code myself, it will make fewer assumptions about my intent when it touches it (especially when I didn't specify them in the architecture or if they are difficult to articulate in natural language).
In small iterations, the agent checks back for each task. Because I spend a lot of time on architecture, I already have a model in my mind of how small code snippets and feature will connect.
Maybe my comfort with reviewing AI code comes form spending a large chunk of my life reverse engineering human code, to understand it to the extent that complex bugs and vulnerabilities emerge. I've spent a lot of time with different styles of code writing from awful to "this programmer must have a permanent line to god to do this so elegantly". The models is train on that, so I have a little cluster of neurons in my head that's shaped closely enough to follow the model's shape.
Personally, I’ve never been learning software development concepts faster—but that’s because I’ve been offloading actual development to other people for years.
I don't necessarily think that writing more code means you get better coder. I automate nearly all my tests with AI and large chunk of bugfixing as well. I will regularly ask AI to propose an architecture or introduce a new pattern if I don't have a goal in my mind. But in these last 2 examples, I will always redesign the entire approach to be what I consider a better, cleaner interface. I don't recall AI ever getting that right, but must admit I asked AI in the first place cos I didn't know where to start.
If I had to summarize, I would say to let AI implement coding, but not API design/architecture. But at the same time, you can only get good at those by knowing what doesn't work and trying to find a better solution.
How exactly? Do you tell the agent "please write a test for this" or do you also feed it some form of spec to describe what the tested thing is expected to do? And do these tests ever fail?
Asking because the first option essentially just sets the bugs in stone.
Wouldn't it make sense to do it the other way around? You write the test, let the AI generate the code? The test essentially represents the spec and if the AI produces sth which passes all your tests but is still not what you want, then you have a test hole.
I care more about the code than the tests. Tests are verification of my work. And yes, there is a risk of AI "navigating around" bugs, but I found that a lot of the time AI will actually spot a bug and suggest a fix. I also review each line to look for improvements.
Edit: to answer your question, I will typically ask it to test a specific test case or few test cases. Very rarely will I ask it to "add tests everywhere". Yes, these tests frequently fail and the agent will fix on 2nd+ iteration after it runs the tests.
One more thing to add is that a lot of the time agent will add a "dummy" test. I don't really accept those for coverage's sake.
A follow-up:
> I care more about the code than the tests.
Why is that? Your (product) code has tests. Your test (code) doesn't. So I often find that I need to pay at least as much attention to my tests to ensure quality.
I find tests easier to write. Your function(s) may be hundred lines long, but the test is usually setup, run, assert.
I don't have much experience beyond writing unit/integration tests, but individual test cases seem to be simpler than the code they test (linear, no branches).
I can iterate on entire approaches in the same amount of time it would have taken to explore a single concept before.
But AI is an amplifier of human intent- I want a code base that’s maintainable, scalable, etc., and that’s a different than YOLO vibe coding. Vibe engineering, maybe.
Then also switching arch approaches quickly when i find some code strategies that are not correctly ergonomic. Splitting of behaviors and other refactors are much lower cost now.
Sometimes I wonder if people who make statements like this have ever actually casually browsed Twitter or reddit or even attempted a "large" application themselves with SOTA models.
An example: I vibecoded myself a Toggl Track clone yesterday - it works amazingly but if I had to rewrite e.g. the PDF generation code by myself I wouldn't have a clue!
This is my experience exactly. I have never been learning as much as with AI.
It's interesting that numbers show most users degrade but I hate the general assumption that some cannot use it properly to learn faster as well.
Previous title: "Anthropic: AI Coding shows no productivity gains; impairs skill development"
The previous title oversimplified the claim to "all" developers. I found the previous title meaningful while submitting this post because most of the false AI claims of "software engineer is finished" has mostly affected junior `inexperienced` engineers. But I think `junior inexperienced` was implicit which many people didn't pick.
The paper makes a more nuanced claim that AI Coding speeds up work for inexperienced developers, leading to some productivity gains at the cost of actual skill development.
> Novice workers who rely heavily on AI to complete unfamiliar tasks may compromise their own skill acquisition in the process. We conduct randomized experiments to study how developers gained mastery of a new asynchronous programming library with and without the assistance of AI. We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average.
The library in question was Python trio and the model they used was GPT-4o.
1. AI help produced a solution only 2m faster, and
2. AI help reduced retention of skill by 17%
you can also read human code and empathise what were they thinking while writing it
AI code is not for humans, it is just a stream of tokens that do something, you need to build skills to empirically verify that it does what you think it does, but it is pointless to "reason" about it.
> Importantly, using AI assistance didn’t guarantee a lower score. How someone used AI influenced how much information they retained. The participants who showed stronger mastery used AI assistance not just to produce code but to build comprehension while doing so—whether by asking follow-up questions, requesting explanations, or posing conceptual questions while coding independently.
This might be cynically taken as cope, but it matches my own experience. A poor analogy until I find a better one: I don't do arithmetic in my head anymore, it's enough for me to know that 12038 x 912 is in the neighborhood of 10M, if the calculator gives me an answer much different from that then I know something went wrong. In the same way, I'm not writing many for loops by hand anymore but I know how the code works at a high level and how I want to change it.
(We're building Brokk to nudge users in this direction and not a magic "Claude take the wheel" button; link in bio.)
Yes, we can use it 10,000 times to refine our recipes, but "did we learn from it"? I am doubtful about that, given that even after running with the same prompt 10 times, it will give different answers in 8/10 responses.
But I am very confident that I can learn by iterating and printing designs on a 3D printer.
I'm wondering if we could have the best of IDE/Editor features like LSP and LLMs working together. With an LSP syntax errors are a solved problem, if the language is statically typed I often find myself just checking out type signatures of library methods, simpler to me than asking an LLM. But I would love to have LLMs fixing your syntax and with types available or not, giving suggestions on how to best use the libraries given current context.
Cursor tab does that to some extent but it's not fool proof and it still feels too "statistical".
I'd love to have something deeply integrated with LSPs and IDE features, for example VSCode alone has the ability of suggesting imports, Cursor tries to complete them statistically but it often suggest the wrong import path. I'd like to have the twos working together.
Another example is renaming identifiers with F2, it is reliable and predictable, can't say the same when asking an agent doing that. On the other hand if the pattern isn't predictable, e.g. a migration where a 1 to 1 rename isn't enough, but needs to find a pattern, LLMs are just great. So I'd love to have an F2 feature augmented with LLMs capabilities
It reduces the context switching between coding and referencing docs quite a bit.
For example I wanted to add a rate-limiter to an api call with proper http codes, etc. I asked the ai (in IntelliJ it used to be Claude by default but they've since switched to Gemini as default) to generate one for me. The first version was not good so I asked it to do it again but with some changes.
What would take me a couple of hours or more took less than 10 minutes.
I’m starting to believe that people who think AI-generated code is garbage actually don’t know how to code.
I hit about 10 years of coding experience right before AI hit the scene, which I guess makes me lucky. I know, with high confidence, what I want my code to look like, and I make the AI do it. And it does it damn well and damn fast.
I think I sit at a unique point for leveraging AI best. Too junior and you create “working monsters.” Meanwhile, Engineering Managers and Directors treat it like humans, but it’s not AGI yet.
This similarly indicates that reliance on LLM correlates with degraded performance in critical problem-solving, coding and debugging skills. On the bright side, using LLMs as a supplementary learning aid (e.g. clarifying doubts) showed no negative impact on critical skills.
This is why I'm skeptical of people excited about "AI native" junior employees coming in and revamping the workplace. I haven't yet seen any evidence that AI can be effectively harnessed without some domain expertise, and I'm seeing mounting evidence that relying too much on it hinders building that expertise.
I think those who wish to become experts in a domain would willingly eschew using AI in their chosen discipline until they've "built the muscles."
I think being intentional about learning while using AI to be productive is where the stitch is, at least for folks earlier in their career. I touch that in my post here as well: https://www.shayon.dev/post/2026/19/software-engineering-whe...
The three high score patterns are interesting as well. "Conceptual Inquiry" actually results in less time and doesn't improve the score than the other two, which is quite surprising to me.
[1] https://martinfowler.com/articles/llm-learning-loop.html
This is one reason I've been resistant to using it. I don't want my work to go to the companies providing the models. I don't trust them. Not only with my data in the first place, but also that they'll keep providing the service over the long term without totally enshittifying the experience.
I'll be so much more excited by this when local models catch up to (or even exceed) frontier-level quality. How close are we to this?
(In my case, I don't even care if it costs a boatload in hardware capital to deploy.)
You should be concerned if you're outsourcing your work to it, though. There's also no benefit to doing that outside of laziness (the research shows no statistically significant productivity improvement)
[1] plug: this is a video about the Patreon community I founded to do exactly that. Just want to make sure you’re aware that’s the pitch before you do ahead and watch.
I use a web ui to chat with ai and do research, and even then I sometimes have to give up and accept that it won't provide the best solution that I know exists and am just to lazy to flesh out on my own. And to the official docs I go.
But the coding tools, I'm sorry but they constantly disappoint me. Especially the agents. In fact the agents fucking scare me. Thank god copilot prompts me before running a terminal command. The other day I asked it about a cypress test function and the agent asked if it could run some completely unrelated gibberish python code in my terminal. That's just one of many weird things it's done.
My colleagues vibe code things because they don't have experience in the tech we use on our project, it gets passed to me to review with "I hope you understand this". Our manager doesn't care because he's all in on AI and just wants the project to meet deadlines because he's scared for his job, and each level up the org chart from him it's the same. If this is what software development is now then I need to find another career because its pathetic, boring, and stressful for anyone with integrity.
Submission about the arXiv pre-print: https://news.ycombinator.com/item?id=46821360
I am not saying you should be struggling performatively, like a person still proud in 2026 that they are still using Vim for large projects (good for you, eh), but sometimes you need to embrace the discomfort.
I remember a small competition where people do a well-defined "share this content to others" routine to showcase how OS A is way more intuitive than OS B. There was also an OS C, which was way slower than A&B. Then, someone came using OS C, topped the chart with a sizeable time difference.
The point is, sometimes mastery pays back so much that, while there's theoretically better ways to do something, the time you save from that mastery is enough of a reason to not to leave the tool you're using.
I also have a couple of "odd" tools that I use and love, which would cause confused looks from many people. Yet, I'm fast and happy with them.
These large projects are amlmost always in Java, C#, and co. Where the verbosity of the language make it required to use an IDE. Otherwise, it would be a struggle to identify which module to import or what prefix and suffix (Manager, Service, Abstract, Factory, DTO,…) to add to the concept name.
Sure, it sounds good to call for more regulation, or admit that there are downsides to your product, but when you know these things are falling largely on deaf ears and you continue operating business as usual, I wonder how much of it is just theater.
> This suggests that as companies transition to more AI code writing with human supervision, humans may not possess the necessary skills to validate and debug AI-written code if their skill formation was inhibited by using AI in the first place.
I'm reminded of "Kernighan's lever" :
> Everyone knows that debugging is twice as hard as writing a program in the first place. So if you're as clever as you can be when you write it, how will you ever debug it?
AI is writing code in the cleverest way possible which then introduces cognitive load for anyone who hasn't encountered these patterns previously. Although, one might say that AI would also assist in the debugging, you run the risk of adding further complexity in the process of 'fixing' the bugs and before you know it you have a big stinking ball of mud.
On the contrary, without mastery guiding, AI writes code in the most boilerplate way possible, even if that means compromising logic or functionality.
> … which then introduces cognitive load for anyone who hasn't encountered these patterns previously
And for those who have. This is the enterprise Java effect. The old trope is Java was designed to make all devs median and all produce the same median code so enterprises don't have to worry about the individual devs, it's all the same bowl of unflavored oatmeal.
When you read code from vibe coding novice, it's difficult to grok the intended logic because that's buried within these chunks of enterprise pattern boilerplate as if the solution was somehow regex'd at random from StackOverflow until some random combination happened to pass a similarly randomized bag of tests.
The cognitive load to reverse this mess into clean clear expression of logic is very high whether a human or machine "coded" this way.
In both cases, the antidote is caring for craft and mastery first, with an almost pseudocode clarity in expressing the desired outcome.
OK, but -- even this doesn't guarantee the result one wants.
Because even if the master writes the code themselves, they may find their intent was flawed. They expressed the intent clearly, but their intention wasn't helpful for the outcome needed.
This is where rapid iteration comes in.
A master of software engineering may be able to iterate on intent faster with the LLM typing the code for them than they can type and iterate on their own. With parallel work sessions, they may be able to explore intention space faster to reach the outcome.
Each seasonal improvement in LLM models' ability to avoid implementation errors while iterating this way makes the software developer with mastery but lack of perfect pre-visualization of intent more productive. Less time cleaning novice coding errors, more cycles per hour iterating the design in their head.
This type of productivity gain has been meaningful for this type of developer.
At the same time, the "chain of thought" or "reasoning" loops being built into the model are reaching into this intention space, covering more of the prompt engineering space for devs with less mastery being unable to express much less iterate intent. This lets vibe "coders" imagine their productivity is improving as well.
If the output of the vibe coder (usually product managers, if you look closely) is considered to be something like a living mockup and not a product, then actual software engineers can take that and add the *-ilities (supportability, maintainability, etc. that the vibe coder has never specified whether vibing or product managing).
Using a vibed prototype can accelerate the transfer of product conception from the PM to the dev team more effectively than PM just yelling at a dev tech lead that the dev hasn't understood what the PM is saying the product should be. Devs can actually help this process by ensuring the product "idea" person is armed with a claude.md to orient the pattern medianizer machine with the below the waterline stuff engineering teams know are 80% of the cost-through-time.
There's not a lot of discussion of prototype vibing being a new way for product owners and engineering teams to gain clarity above the waterline, or whether it's productive. Here's a dirty secret: it's more productive in that it's more protective of the rarer skilset's time. The vibe time wasted is paid by the product owner (hallelujah), the eng team can start with a prototype the product owner iterated with while getting their intent sorted out, so now engineerings iterations shift from intent (PM headspace) to implementation (eng headspace).
Both loops were tightened.
> you run the risk of adding further complexity in the process of 'fixing' the bugs and before you know it you have a big stinking ball of mud.
Iterating where the problem lies, uncoupling these separate intention and iteration loops, addresses this paradox.
This study is so bad, the sample size is n = 52 and then in some conclusions it goes down to n = 2.
One of the things I worry about is people not even learning what they can ask the computer to do properly because they don't understand the underlying system well enough.
One of my little pet peeves, especially since I do a lot of work in the networking space, is code that works with strings instead of streams. For example, it is not that difficult (with proper languages and libraries) to write an HTTP POST handler that will accept a multi-gigabyte file and upload it to an S3 bucket, perhaps gzip'ing it along the way, such that any size file can be uploaded without reference to the RAM on the machine, by streaming it rather than loading the entire file into a string on upload, then uploading that file to S3, requiring massive amounts of RAM in the middle. There's still a lot of people and code out in the world that works that way. AIs are learning from all that code. The mass of not-very-well-written code can overwhelm the good stuff.
And that's just one example. A whole bunch of stuff that proliferates across a code base like that and you get yet another layer of sloppiness that chews through hardware and negates yet another few generations of hardware advances.
Another thing is that, at the moment, code that is good for an AI is also good for a human. They may not quite be 100% the same but right now they're still largely in sync. (And if we are wise, we will work to keep it that way, which is another conversation, and we probably won't because we aren't going to be this wise at scale, which is yet another conversation.) I do a lot of little things like use little types to maintain invariants in my code [1]. This is good for humans, and good for AIs. The advantages of strong typing still work for AIs as well. Yet none of the AIs I've used seem to use this technique, even with a code base in context that uses this techique extensively, nor are they very good at it, at least in my experience. They almost never spontaneously realize they need a new type, and whenever they go to refactor one of these things they utterly annihilate all the utility of the type in the process, completely blind to the concept of invariants. Not only do they tend to code in typeless goo, they'll even turn well-typed code back into goo if you let them. And the AIs are not so amazing that they overcome the problems even so.
(The way these vibe coded code bases tend to become typeless formless goo as you scale your vibe coding up is one of the reasons why vibe coding doesn't scale up as well as it initially seems to. It's good goo, it's neat goo, it is no sarcasm really amazing that it can spew this goo at several lines per second, but it's still goo and if you need something stronger than goo you have problems. There are times when this is perfect; I'm just about to go spray some goo myself for doing some benchmarking where I just need some data generated. But not everything can be solved that way.)
And who is going to learn to shepherd them through writing better code, if nobody understands these principles anymore?
I started this post with an "if" statement, which wraps the whole rest of the body. Maybe AIs will advance to the point where they're really good at this, maybe better than humans, and it'll be OK that humans lose understanding of this. However, we remain a ways away from this. And even if we get there, it may yet be more years away than we'd like; 10, 15 years of accreting this sort of goo in our code bases and when the AIs that actually can clean this up get here they may have quite a hard time with what their predecessors left behind.
[1]: https://jerf.org/iri/post/2025/fp_lessons_types_as_assertion...
I can start to see the dangers of ai now, whereas before it was more imaginary sci-fi stuff I couldn’t pin down. On the other hand a dystopian sci-fi full of smart everything seems more possible now since code can be whipped up so easily, which means perhaps that the ability for your smart-monocle to find and hack things in every day life is also way more likely now if the world around you is saturated by quick and insecure code.
>We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average.