We were so spoiled. We got fat salaries to sit in air conditioned Herman Millers all day learning about computers. Now we discover a way to synthesis intelligence and the only thing we can think to do with it is ruin the most fun career most of us could have ever hoped for.
Sure, we're all more productive now, but how much of that is because we leverage AI on top of the intelligence we gained from all of that manual work? Who is to say that in 36 months you're not a worse developer over all because that systems knowledge starts to atrophy too?
This isn't me saying you shouldn't use AI. I use it all of the time to do useful side tasks like to setup GitHub Workflows while I write a feature, or with my agent on a VPS to do internet tasks for me. It's nice to have a little synthesized intelligence.
What isn't nice is to supplement your own intelligence. I think the gains are in the work there--similar to how you can be absolutely ripped from taking steroids while destroying your body. Often it's the shortcuts that are the most treacherous path.
I’ve been trying to ask people a different question: sure, we’re more productive now but to me, the AI era is only serving to plunge us deeper than ever into producing more, more, more, faster, faster, faster. And for what? What’s it all for? I became a software engineer because I have a lot of fun writing code, thinking through and solving complicated problems, and experiencing the reward of seeing what I’ve built by hand working for the first time.
Do people really have fun managing a fleet of agents that generate the code instead? Or is it just the rush of producing something extremely quickly, much more quickly than you might be able to alone, regardless of how well (or poorly) it might work? For me, being able to move quickly was never the fun part.
It’s one thing to utilize AI to lessen the drudgework, the boilerplate, but I look at people who have gone all in on agentic development and it just really makes me wonder.
> I’ve been trying to ask people a different question: sure, we’re more productive now but to me, the AI era is only serving to plunge us deeper than ever into producing more, more, more, faster, faster, faster. And for what? What’s it all for?
Many devs here have stated that the fun part for them is seeing the end product, not the act of creating it. Using AI is an act of need satisfaction.
Unfortunately, cloning a GitHub repository or downloading a Squarespace template doesn't hit the spot, because you can see exactly where the code came from, so your brain knows you were not the one responsible. AI's greatest feature is that it obfuscates provenance. You can now happily clone that repository or download that template without the feeling that you're cloning someone else's repository or downloading someone else's template.
People who use coding as a means to an end, producing a product
People who enjoy process over product and coding is the enjoyable part, not the end product
Company executives fall into bucket 1. Even if you love your cushy air-conditioned job, doesn’t mean the people above you don’t see you as a means to an end, a better product.
Solo founders and small startups are in bucket 1 as well but that doesn’t mean that don’t enjoy coding, just the product being made is much more satisfying.
Brother, there was slop before AI and slop after AI.
Pre-AI slop got a pass because the creator probably worked tirelessly to make it and that process gave it value.
Post-AI slop get criticized because AI can create it so fast and easily.
But the builder has not changed. Because the builder knows that the product depends on iteration and polish. And knowing what to iterate and what to polish requires intuition and taste. And thats the thing about taste, its forever fleeting. What we think is AI slop, was actually novel and coveted at a point before AI existed.
You're either using it wrong, or work somewhere that sucks. I'm having a ton of fun coding with AI.
I feel like I am learning a whole new branch of the programming skill tree. LLMs and their harnesses are like a mew set of constraints to work around when designing systems, but if you get them right you can build bigger, better things than ever before.
I say all of this as someone who spent the last two days rebuilding RBAC for my application after Claude royally messed it up.
I attempted to build a K8s based image catalog and management system a few years ago, and progress was very slow. Took 2 months to get the first vaporware demo put together.
This year I'm building and IoT data management platform and I've already built two demos of the product and am adding a bunch of features for a third. Nothing is vaporware about it. All in about 4 months.
The big wins have been system config, debugging, and exploration. I was able to build an Arrow Flight SQL backend to use as an interface and try it out with some use cases, decide it wasn't going to work, and replace it with something else, all in about two weeks. Would have taken far longer before, if I would have been able to do it at all. I knew nothing about Flight SQL before trying it out.
I shit you not I have been told to "stop trying to understand it, just make sure it works" when mandated to use LLMs at work, like the absurdity of that statement isn't self-evident.
Needless to say I have been looking for a way out of this job (and likely career) ever since.
What's going to happen is people who give a shit are going to get filtered and self-select out of the job, and then the entire field is going to be dominated by dunning kruger AI maximalists. The "AI will replace engineers" is true, but for an entirely different reason and entirely different way than most people making that argument think.
Please stop saying we - because many of us didn't ever desire or contribute (outside of having our code / data / IP stolen) to these models or the companies developing them. Many of us weren't willingly singing up to use AI - we were essentially forced to by our employers. Not everyone is more productive now because of AI - many people spend the extra time they save writing code, reviewing code produced by AI.
Yep. My job was code review heavy before AI. Now it's even worse because I can't even trust that the developer understood what they were "developing", so I have to be even more vigilant. I may have to institute a "prompt review" process for some.
if computers themselves are bicycles for the mind, LLMs are cars. or even motorcycles, if we talk smaller models.
- they require specific roads being paved for them. for example, if your tooling is proprietary and not accessible from the CLI, your agent is pretty much fucked. if your tool is not represented in training data (think, `jj` VCS or your proprietary/tailor-made tooling), you require duct tapes such as "skills" and "memories". a bicycle (that is, your own mind + computer) handles such off-roads much better.
- they get you from A to B faster, sure, but along the way you may encounter something curious - a different road to take, an interesting vista. not to mention, bicycles are actually good for your health, and professional drivers suffer from all the sedentary job diseases we programmers do, unless they actively counter it. with LLMs, we get a "sedentary job disease" of skill atrophy, on top of all the other atrophies us coders should counter with a proper exercise set at least three times a week.
- finally, when you crash in a car (Opus/Sonnet, GPT-5.5) or, worse, on a motorcycle (smaller Qwens, DeepSeek, Haiku/GPT-5.4-nano), you crash very loudly and with a high chance of irrecoverable casualties.
The bike/motorbike/car analogy is good. In each you can find enjoyment that suits your character.
I have been drawn to coding from the first day I took a course at college learning Pascal and assembly on an IBM PC. Hell I even wrote an IEEE/488 driver for an Osborne 1.
As an experimental physicist, coding was always central to my work. I always felt guilty because I had more pleasure coding than tackling the physics itself.
This is a new age we are entering. AI changes how we do things, but I believe that the human passion to be creative and do intellectual work will always be part of it. It's in our nature.
So I'm not worried that we'll all dry up and turn into zombies.
And, like cars, our federal/state governments are lighting astronomical amounts of cash to subsidize the infrastructure that makes them operate at the scale they currently operate at.
Just one more ~~lane~~ datacenter, bro. That'll fix it; trust. Please. One more ~~lane~~ datacenter. It'll be so good.
There's one more: The "bicycle for the mind" analogy was one Jobs chose specifically in reference to the fact that a human on a bicycle moves more efficiently than a human on foot—or any other land animal for that matter. But it's still the human's own effort that propels the bike. Motor vehicles require an external energy source, the human only directs it. Similarly, with LLMs, it's no longer the user's own effort that's doing the thinking.
> There's one more: The "bicycle for the mind" analogy was one Jobs chose specifically in reference to the fact that a human on a bicycle moves more efficiently than a human on foot—or any other land animal for that matter.
One of the canonical references of Jobs’ use of “bicycle for the mind” also compared the efficiency of locomotion to the California Condor which was (I’m working from memory, here) 17X more efficient than a human walking.
A human on a bicycle, according to Jobs then, is more efficient than the most efficient animal locomotion known to humankind. The comparison included animals moving through non-terrestrial environments.
I wonder if this same effect happens for very wealthy and/or very senior executives? Those sorts of people have always had numerous people they could 'outsource' their thinking to; delegating work, asking for research/summaries, assigning tasks.
Does handing off that sort of work to people also ruin your skills in the same way? Or are AIs fundamentally different, and if so, why? Because we have no moral or social pressure to not delegate everything?
Yes. The moment you move from daily IC work to management your skills begin to atrophy. Even if you are magical and somehow they do not - you are not keeping up/practicing on the newer developments.
If you are doing your job as a manager it’s the scariest career possible. There is no realistic fallback to lower paid IC work after a certain amount of years. Your job is to enable others, not do things yourself.
There are shades of grey and you try to keep skills at least honed a little here and there doing R&D and side projects - but it’s just not the same as day to day production line work.
And of course some people start at a much higher baseline of skill than others. The impact is largely the same though over time.
This is by far the single largest caution I give to skilled engineers who talk to me about moving into a management track. It’s a decision not to be taken lightly.
Yes. It even happens to e.g. professors and the deans of universities.
This is why I think even the fuss about Noam Shazeer joining OpenAI needs to be seen within a context; as good as this hire is, there is no inherent reason to believe he still brings some secret undiscovered magic that others do not have in a more current form.
I find this hard to parse, maybe because English is not my first language. Do senior SWEs provide more value than they provide principles, or do they provide more value than principles do?
intellectual work (or really any discipline with a 'skills tree' that you can progress up or down) degrades over time without practice. same way that, if you don't run for a while, you're not going to be able to hit your last PR when you pick it back up.
it's definitely easier to catch up after some time away than it would be if you'd never developed the skills in the first place or didn't have a natural talent, but you'll definitely atrophy without exercise. every leader i've ever worked for who graduated to a purely managerial/'strategic' position and didn't keep up their IC skills eventually got pretty slow on the uptake.
i appreciate that this study was done (AI and its inverse relationship to human wellbeing is one of the biggest challenges of our time IMO) but this also seems obvious
I've discussed this in my company, where I've argued that we are losing critical thinking skills for AI.
My friend said being an engineer is not just writing code, but designing and architecting the system. I agree, but we are also moving that to AI. We are offloading the thought process to a LLM, which confirms our biases and tells us what we want to hear.
Essentially, we are also losing the software architecture and design skills because we aren't talking to other engineers/architects/designers who may approach the problen in a complete different way; we are just asking the AI to confirm our thoughts.
The ownership class does not really do work. What they delegate (high level plans) is generally easier to do for AI agents than things like software engineering because it does not need to be as precise or executable. However, there the work has very broad scope scope and the roles are very high risk, so it might need a Fable 5 or Fable 6 level vision language model to remove the jaggedness to make it 'safe' enough to drop the humans But what is going to blow up in 2027 are Automated AI Companies. Human CEOs and owners will not be able to compete with these AI-run companies.
It will be interesting to see how AI and company leadership/ownership stuff shakes out. In some sense I would expect it to be quite different than what exists currently, because an AI can’t be responsible for anything legally, and (arguably?) can’t own anything. But then, C-levels don’t actually seem to have much personal legal liability toward their company’s decisions, or maybe it won’t be so big of a deal.
I’d be curious to see alternatives ownerships structures. Like an AI-coordinated collection of guilds, unions, or co-ops. If it can’t accumulate upwards, maybe the fundamental unit of ownership will stay with the workers.
Yeah, the idea that C-suites ever took any significantly liability has always been a fantasy for as long as I can remember. At worst/best case a single lower level individual will be used as a scapegoat for the systemic abuses of the company.
AI needs a great deal of handholding which is different than just offloading tasks. You spend a lot of cognitive energy playing a slot machine hoping the RNG works out this time.
There’s definitely a skill to using AI but it just doesn’t generalize very well.
I still believe the slot machine analogy holds to some extent, but I can honestly say my winning percentage is at least 90% for one shot generated code now.
I think if you know it's limitations (inlcluding your own), I don't think about hoping anymore.
I should note that when I say AI, I mean the collective models from all the major providers. The most important lesson is, you need to ask around.
> There’s definitely a skill to using AI but it just doesn’t generalize very well.
This I agree with. The only way working with AI can really be benefical outside of dealing with AI is, we are visited by extremely intelligent beings that will fuck up in the weirdest ways.
> There isn’t a magic prompt I could have written before knowing the solution. A retrospective prompt can reproduce the 200 lines only because it embeds the invariants you discovered. The useful AI task would have been systematically finding counterexamples and iterating toward those invariants, not generating the final file in one shot.
For me the value with AI isn't the one shot, it is the fact that it can help you iterate faster. My biggest take awway so far with AI is it can help you fail fast. The faster you can fail, the faster you can iterate.
When it works the first time you’re done, when it fails you spend more time on it. This biases the cognitive investment around failure.
> I don't think about hoping anymore.
Running the exact same prompt again that already failed doesn’t have a that high success rate, but it’s also very low effort. So IMO it’s often worth attempting.
mine def has to be the initial impl of my python to my personal fave compiler ir. i got claude chat to write it in two sentences. 4 turns helping it rmeemver where it put shit becsuse of transcript issues
The code in http://github.com/gitsense/gsc-cli contains over 300 go files and I would say 90% were one shots. I can even prove that they are AI generated since I attribute the AI in each file, as shown in the example below:
Since I don't know go, I can't say which is the best. I am also sure they would not live up to a go experts expectations. I just know that since I don't know go, I would not have been able to create the tool. I can also prove that you can easily maintain AI generated code since I have been expanding and evolving it for over 6 months with pretty much no issues.
Yes, but what happens to the professions and classes you've mentioned that wasn't looked at in this study is that new skills developed. For instance, something was assuredly lost as programming on average deviated from low level programming, but new skills and optimizations were gained. Fewer people knew how to construct a real-time bitmap from scratch, but more advanced software was being built on average as people were armed with greater and more manageable abstractions.
Having run into quite a few managers who used to be programmers who can barely do fizzbuzz after a few years of managing: yes, I think it's pretty similar. Much of your brain is "use it or lose it" and that's not using it.
I do think there's more than enough room to claim that LLMs are probably significantly worse than that kind of human delegation though, in part because you have such a rapid cycle time that not-incredibly-rich people can't afford from humans.
The difference is what you delegate to. An executive delegates to people like himself. A computer user delegates to another kind of existence so those think don't think the same way the user does.
Programming normally highlights this difference. LLM programming makes it much less apparent but its still there, LLM are not thinking the way humans do and therefore struggle to solve many problems humans easily solve. So letting all human programming skills rot and just use LLM will halt our progress unless we reach AGI before our programmings skills are mostly gone.
The "I have people skills; I am good at dealing with people!" character in Office Space was memorably hilarious for a reason. Many people lose the skills that got them where they are from disuse as they rise through the ranks.
>Because we have no moral or social pressure to not delegate everything?
"We" depends on where you are. Countries like Japan or Germany have maintained a gray collar, rather than a blue/white collar culture for exactly that reason. You will find business owners on the workshop floors frequently because there is an understanding that divorcing management from tacit work is going destroy leadership ability. That's the basis of vocational work culture, having general expertise across all domains of your job rather than being a kind of over-specialized idiot.
This is why the rich and powerful often seem so out of touch. There's no one around them willing to tell them they're wrong or push back on bad ideas. It doesn't just ruin skills but fundamentally distorts one's perception of the world.
The main skill you need to reach that level is the one they always keep practicing: shmoozing. Thinking is not required as an executive, so you cannot lose what you never had.
And being born wealthy requires zero skill or practice.
This seems to assume that a skill is only important if you need it for your job.
If that is the case, then wouldn't this whole thing be a non-issue? We lose all the skills we used to have, but we don't need them because our entire job now is interacting with AI, and that skill we will continue to develop because it is what we do all day.
I don't know if I fully agree with that. Some skills are important even if you don't need them for your day job.
> Or are AIs fundamentally different, and if so, why?
Literally: the context window.
With the human you have a window that possibly extends up to _years_. With your language model you have maybe a few megabytes which is always preceded by instructions from the model maker.
That's a misconception. Context window of humans is measured in seconds and limited to single-digit amount of entities. Human attention is truly ephemeral, with a ridiculously short span. What you're talking about cannot be emulated by putting everything into the context, you need models with deeper cognitive capabilities, and the longer your context the smarter it needs to be. It can be done by increasing the depth and the size of the model, and by training it better. Test-time scaling does work (longer reasoning, agentic compression, ICL, etc), but hits the wall rather quickly if not assisted by better cognition.
That's a simple neurologic fact, there's nothing to believe. You can use a trivial experiment to verify that you can't keep details in your sliding attention window for more than a few seconds or focus on more than a few things simultaneously. Human memory and cognition is layered and this immediate layer is what resembles the model's context the most.
You're possibly mixing it up with long-term memory which doesn't keep immediate facts and details, it's for heavily processed and compressed summaries, for the lack of a better analogy from the LLM world. You aren't keeping the entire codebase in your memory, just its highly processed and conceptualized version. This conceptualization can be somewhat emulated as an agentic loop, but it can only go so far, current models quickly lose coherency and aren't good enough to predict what's important.
Models don't need to remember more details, they need stronger processing of what they already remember.
>"Oh yea, just do it _better_. That's your problem."
I think we're talking about different things. Models can be trained better to cram more intelligence into the same amount of parameters, that's what I mean. Similarly to how your ability to learn (and perform) math depends on your prior math training.
Is this true, though? AIs already interact with way more data than their context window, they can make queries to external data sources, and they can make context references to external data that can be pulled in as needed.
The context window is the working memory, which is not ‘years’ for humans either.
I haven't written a full function of code in over a year. That being said, I've been spending a lot more time thinking about architecture and system properties.
So, yes, I do feel like I've lost some of that very low-level skill. But maybe I've also been able to spend more time on a higher level skill? Maybe the doctors got worse with the images but had more cognitive resources to think about the patient's context?
Not sure.
But yes, I can't physically get myself to write code without an AI anymore. It feels so much slower, almost painful.
I’m curious if someone coming into that fresh, without having had the mental exercise of, for example, grappling with data structures and (lower-level) algorithms in practical applications by hand, could achieve a mindset that useful and productive?
When I was in design school, as much of our work was in physical media — graphite, cut paper, paint, vine charcoal — as was practicing great kerning and getting experience with the digital tools. Even though you still had to make the individual strokes and choose appropriate tools in the digital realm, there was still a perception of the process that was obviously lacking among those that came from strictly digital backgrounds. It’s similar to seeing someone who’s only worked with photo references try life drawing — there’s an entire part of the cognitive process not being used when you’re drawing something that’s already 2D. Sure, they can learn, but unless they’re forced to, they’ll probably just keep taking a picture and drawing that. But image generating, even with extremely granular inpainting and such, is so different it’s not even comparable. I’d hesitate to say that someone with a lot of experience doing very advanced image generation would be dramatically further along than a complete beginner if they learned to draw which is not true of the photo reference artist, and even less true of a purely digital artist that did life drawing on a tablet.
Kind of like how millennials, many of whom always had access to technology, but also experienced dial-up-era computer use, are generally more technically savvy than the your stereotypical “iPad kid” that can’t even traverse a directory structure.
Yes, and I struggle to give junior developers feedback because of this.
The code they submit for review no longer represents their thinking or their skill. The feedback loop for beginners is a bit broken.
What is their skill now? How is it displayed? How and where do we provide valuable feedback? I find myself just approving large PRs because I don't have the time to read it all.
I feel like I've let down a lot of developers this year.
> I feel like I've let down a lot of developers this year.
I think the industry has, and we don’t get much out of individualizing it. Let’s not forget that this progression is a series of deliberate choices made by a handful of very influential people that couldn’t give a teaspoon of shit about developers beyond what their subscriptions bring in.
It’s not a super well-received opinion around here, but I think the economics of the large-scale services will mean on-device models are going to be where this all goes sooner or later. My gut says setting up a local LLM coding assistant will be more like choosing and setting up an ide. I don’t have a crystal ball though.
I was thinking about this (while drawing), and there's a few points to make:
* There's meaning in all the little marks. Drawings are potentially layered with meaning. I was drawing a stick I found in the park, and thinking: this is the extent and the way in which the stick is straight or curved, and the specific way and form that it's knobbly, and that's all explained by how it grew; this green on it is the peculiar blue-green of lichen; it's a oak stick, I like oak; my drawing is just a stick, good, I might frame it, call it anti-art maybe. Some or all of that is expressed in the marks, though this becomes more apparent if I tell you that it is.
* So, an AI - diffusion - could undoubtedly draw a stick in an arty way. But to draw one that fits the above meanings, it has to be prompted with the meaning. The AI can't prompt itself, so there's a role for artist-as-prompt-writer. Whoopee, right, what fun. These prompts, if about graphical matters and a sort of meditation on the stick, would be essentially lies, because no such exploration of the form and significance of the stick would have really taken place without trying to draw it. It can't prompt itself or be honest.
* So to come up with the above words that might do as a prompt, I had to draw a stick. I had to learn by drawing a stick, even though I've been drawing for decades. Experience isn't the whole deal, otherwise the artist is only churning out filler material, which would be like AI art. Instead the artist has to explore all the time, while leaning on experience. The viewers are into that, they sense the excitement of exploration. They want to see an artist. Well, not exactly, most artists are unrewarding to look at (Brian Froud for instance), but they want to see creativity unfolding, over several pictures.
You can have non-graphical (non-mark-making) prompt-art, a bit like collage or photography, sure, but that's its own thing. Like you're saying, you can't just fake the craft forever, and even faking it once is less than ideal, unless you're narrowly focussed on output that meets targets, instead of meaning.
With programming, this might be different, since hitting targets and getting functionality might be all that's wanted, but I'm sure it depends.
The way I see it is with every stroke, an artist makes multiple decisions, most of them unconscious. They’re based on everything from the media, to our experience, mood, inspiration, and very much on our physiology… the arc of my shoulder and the length of forearm are just as much a part of the process as what inspired me about something to draw. To some extent, much of that is even true in photography.
Diffusion models basically record, classify, and amalgamate those decisions, which is why it’s so dang difficult to get generated art to look like something distinct. Not distinct like an existing artist, but genuinely unique.
The workflow of the prompter is very similar to an existing workflow in the art world: someone commissioning art. There’s often a discussion where the customer gives the artist a textual description, back-and-forth with sketches and preliminary versions, and sometimes revisions if the proposed final product isn’t what they wanted. Commissioning a piece of art is a creative process, but it’s not the same thing as being the artist. Even an art director who has extremely granular control over what they commission would never claim to be the artist. They’d get run out of town. I believe creating a collection though juxtaposition or even curation could be art, but you’d still never be the author of the contained pieces.
Ai is decent at it but it still struggles to find and balance and choose the right context. This will of course improve but I’m starting to think it’ll be like the self-driving car history where the last 10% needed is a long slow road.
10-20 years ago we had such high hopes for data and IoT, but it’s still just banal and mildly useful, full of walled gardens and property formats. General purpose AI may end up similar. My company has like 10 different MCPs to enterprise software we use that has context, but AI still needs lots of guidance.
I think the main aspect is that a human can be held responsible. A human can be fired.
Then we are pretending agents have agency when that is also what the human has.
It seems like the meaning of economy of scale should almost flip to mean the economy of very small, highly specialized companies.
It seems hard to see that the person in charge of cybersecurity is going to also be in charge of marketing.
I think there is also this aspect that if you told me in high school I was someday going to be a project manager, I would have thought you were insane. You mean like Oppenheimer? No one is ever going to put me in charge of the Manhattan Project.
I think you can also see the vague outlines that someone is going to have to be in charge of token cost.
>But maybe I've also been able to spend more time on a higher level skill? Maybe the doctors got worse with the images but had more cognitive resources to think about the patient's context?
That's not the way the economics behind this work.
Supposing the AI priests are right (they aren't) and using AI creates a thought surplus on the user, freeing cognitive capacity to think of higher things. What do you think will said user's boss want to do with that surplus? Let the user develop higher-level cognitive abilities? I don't think so.
The doctors in the article performed worse post-AI: suppose AI saved them so much time that they did 100 exams in the time they used to take doing 10 exams. What will their employers do with that freed up labour time? They'll of course have the doctors do more exams and perhaps fire some now-redundant doctors that are no longer needed. The surviving doctors are left deskilled, doing the same or more work, and society gets worse quality medical care. But hey, its not all bad - the employer gets to save on labour, and shareholders will be happy.
Absolutely not, not until LLMs stop being unreliable bullshit generators, and I believe that the current generation of AI is simply not capable of it. Even the best available (lol) model, Fable/Mythos, is still a bullshit generator, just one that's a bit more powerful.
You’re fucked once you have to start getting AI to produce code for languages you’ve never actually worked in.
You will have to write code to understand deeply what goes on at the low level. Without a solid understanding of the low level, you won’t truly know what optimal solutions look like on a high level. You will be flying by the seat of your pants, churning out code that works but has bad low level quirks sprinkled through the code base. The AI will say it’s fine, but you’re just building up shitty software. Feels like shit, run likes shit.
Once someone new comes along who has worked with the language manually and can get AI to produce effective high quality code, you’ve lost competitive advantage. They can build way better versions of whatever you do. You’re finished. You’re a low quality engineer.
Your argument has many, many leaps but I infer your point.
The lack of attention does reduce the overall correctness of the system.
But, and unfortunately, in most situations the loss of correctness is more than made up for by the speed gains.
Unless you're building software that could kill someone, it's hard to reject the improved speed. The cost of delay is usually more damaging than the cost of incorrectness.
In real business, all important software is capable of killing someone or cause monetary loss through incorrectness or loss of functionality due to performance failures.
If that doesn’t apply, you are just writing toy apps or front ends that are of no real significance, so yes you could carry on with incorrect code in those cases, but it’s not much of a career either.
This is new, the scope of it, its not just about individual "skills" because its all of them; we are being challenged at the very fundamentals of our ability to think deeply and widely and persistently. That has never happened before like this.
It is quite extraordinary and breath-taking at times to see the agents in action; the flipside is that very power renders us both vulnerable to its seduction and enfeeblement on an equal scope - its almost hard-drug like in its potential long-term psychological effects.
> we are being challenged at the very fundamentals of our ability to think deeply and widely and persistently. That has never happened before like this.
Social media and content algorithms come to mind as an early wave that changed the landscape here that defines the horrible status quo leading into the AI era.
These days it's trivial to slide into an echo chamber and very hard to break out of the silo.
There might be a double-edged sword here where AI, trusted by most people as an omniscient oracle, can offer the only pushback we encounter on positions we picked up passively by scrolling social media, Youtube, TikTok.
For example, ask Claude, ChatGPT, and even Grok about the "space lasers" that started wildfires in Hawaii in 2018, something people like Marjorie Taylor Greene floated on social media. It quickly debunks it as bullshit.
Now, maybe it will pan out such that everyone will have their own AI that tells them what they want to hear. But so far I've watched people abandon arguments on Twitter because Grok rejected their claim. So it feels like there's a glimmer of hope.
You should possibly spend some time reading what people used to say about the invention of Radio and Television.
> It is quite extraordinary and breath-taking at times to see the agents in action;
So is any magic trick. The unsettling notion that it may all just be an illusion that you've failed to correctly understand doesn't seem to weigh on people.
> its almost hard-drug like in its potential long-term psychological effects.
That might have more to do with how the owners of these products choose to market and deploy them. Perhaps if they peeled back the covers just slightly your euphoria would change to dread. There's an Upton Sinclair moment coming.
"They will cease to exercise memory because they rely on that which is written, calling things to remembrance no longer from within themselves, but by means of external marks"
And he was right, that is why we don't let people bring their textbooks to most tests. That rule wasn't made because we hate students, its because it improves the outcomes if we force people to commit things to memory instead of just looking them up.
You would need similar restrictions during the learning phase to get people to learn things in an LLM world.
So writing and LLM are still a net positive, but only if you acknowledge the problem and add fixes so people still learn rather than just rely on this crutch.
Human mind as much as body needs challenge. That's the only way for growth, heck even sustaining some higher cognitive levels.
Nurses, doctors and family members know damn well how life trajectory nosedives for somebody ie suddenly bound to bed, when stimulus and doable challenges are reduced to minimum.
llms remove challenges, or minimize them. I can't image any added value for any engineer apart from cost cutting for employer. Sure, next come folks who are doing 10x compared to before, and some actually do. Even there, I have my doubts. For rest of us, its not good and won't get better unless they price it out of most markets.
Im learning new things at a pace I never imagined at 40 years old. New sports, new businesses, new academic pursuits. Technology is a lever and AI is the biggest lever we've ever had. It enables laziness or incredible productivity. Choose your own path forward.
I don't know this applies to you, but I've had several friends who convinced themselves they're exploring the frontiers of science, and it always turns out that their conversations spiraled into some sort of weird quantum-metaphysical gobbledygook.
LLMs are sycophants, and in long conversations, their sycophancy produces a positive feedback loop: the context window contains affirmations of incorrect interpretations / analogies, so the chatbot continues down that path because, well, that's the most likely completion of previous text. And before you know it, you're discovering the hidden fabric of the universe, which is always some Minkowski fractal spacetime tensor lattice manifold with subharmonic DNA nanotubes.
That is to say, unless you have a robust way to evaluate what you're learning, and to confirm that you're actually learning, I'd tread carefully.
You really just need to augment with tight prompting and know how to extract information and links to peer reviewed literature and well sourced information. Once again, technology here is a lever. Separating wheat from chaff has been key in academic and information pursuits forever and it is becoming ever more important.
And you're gonna review that voluminous academic literature in a field you're not familiar with, right?...
That is my point: an LLM can be great if you know the field and can spot errors. Or, to a lesser extent, if you have some automatic feedback loop that the model can't easily game ("does this code pass unit tests?"). It's a lot less great if there's a risk that you won't detect the early drift.
Sifting through information to separate truth from fiction is a modern, experience-based skill that I’ve developed YES.
However, your perspective on how LLMs are used might be too narrow. While no one is suggesting using an AI to find a one-shot cure for Alzheimer's, LLMs are incredibly effective tools when paired with textbooks to master subjects like undergraduate physics.
100% you are so right. Don't rely on the LLM. It is a robot you give tools and context, and then you ask it to do something with that. It's amazing at that.
>LLMs are incredibly effective tools when paired with textbooks to master subjects like undergraduate physics
I agree that they can be helpful for explaining things if you're stuck. But since you used the example of physics, the majority of hard won physics knowledge comes from working lots of problems yourself.
Too many people are already used to LLMs circumventing the laborious aspects of things, so it takes a certain person to withhold from just having the LLM solve the problem for you when you're stuck.
Until we can figure out to teach people to fight against their instincts, then I don't think LLMs will lead to better long term outcomes.
If you’re cognizant of the sycophancy and not a total narcissist then it’s not too hard to figure out with a little experience when the models are just trying to hype you up
Learning by doing is often more effective because it naturally involves forced recall.
But it's not the only way to learn and not even the only way to do a lot of forced recall.
It's downright crazy to say you aren't learning anything by reading. You likely won't retain that high a % of the content without repeated drilling, but it's not like nonfiction exists for no reason!
>It's downright crazy to say you aren't learning anything by reading. You likely won't retain that high a % of the content without repeated drilling, but it's not like nonfiction exists for no reason!
How much of a nonfiction book can you recall a few months after reading it? Probably very little.
However, I haven't taken a calculus class in nearly a decade and I still know how to solve derivatives and integrals.
Point being: there's levels to this, and reading is not nearly as effective as drilling exercises.
You can have it write a program that generates drills for you.
I wanted to become better at reading sheet music so I generated a sheet music reading program. You can have it generate maths drills, then ask questions about it if you get stuck or whatever. If you genuinely want to get better at something then AI will help you learn it faster. Obviously its going to hamper more people's cognitive ability that it will enhance but that is a separate problem.
The people who say they are learning faster with it aren't mentioning drills, they're mentioning outcomes.
I actually did have an LLM ingest some material and generate drills. It worked well. It's rare that happens, though.
The difference between humans and other animals on the planet had always been the ability to reason. If we, as a species, lose that ability, we're looking at an extinction-level event.
For example: AI has helped me get into restoring retro tech, specifically resoldering leaky caps on retro Macintosh logic boards. Before AI, I didn't know how to use a multimeter (I knew theoretically how it worked), I didn't know how to use flux, solder wick, heat gun. I also didn't understand how bromine radicals yellowed plastic and how to reverse it by using blue light similar to what they use for indoor aquariums.
No, not everyone, there's no easily accessible video showing how to setup aquarium lights to slowly de-yellow ABS plastic -- at least I didn't find one; AI helped me find an obscure reference to it and I then read a few message boards where it was mentioned.
I also use AI to take in-progress pictures as I desolder to help me check for traces that need to be repaired or help identifying specific chips. I probably could try and find a video where the same chip is featured and someone explains it, and/or retrieve the schematic for the specific logic board, but that's very painful and does slow the process. Think of AI, in this specific case, as enabling skill development for me in a field I wouldn't have necessarily have gotten into, because of being short on time and AI helps me consolidate that information quickly.
You are very closed minded. LLMs can absolutely give pragmatic and fast information in a useful (non-sycophantic form). You need to examine your extreme biases here.
Just because you can't use a tool doesn't mean the tool isnt useful.
Your bias on display here is frankly silly. Im not saying LLMs are ALWAYS the best way of learning something just like they aren't always the best at anything. They are a valuable tool though. Yes so is youtube and textbooks, and professsors, and peer review literature, and pen and paper, and block training etc.
Should we also go to ancient scrolls for learning, because videos are too easy? The poster showed you how they learn with the help of AI. You can learn from videos, they learn from AI. What's the problem?
Nonsense. Absolute nonsense. Ive crafted LLMs to create drills and practice exams etc. Im not just reading about new sport activities, or videography/photography or linear algebra or physics. I'm putting it into action. As I said above, technology is a lever. Maybe you are just reading stuff and not drilling problems. That isnt me.
Not sure how good LLM are at that given they are so biased towards the user being right. You would need an LLM trained to disagree with the user to properly grade your results of those drills.
At most you could do multiple choice questions, or questions with a single objective answer such as "what is the capital of France". But such tests aren't very helpful to learn new things.
Students cannot grade their own answers for more complex problems, they make mistakes but say they are correct since they don't understand the material, or they are correct but since it doesn't say the exact same thing as the example answer they say they are wrong and correct it even though it was already correct. And an LLM would be even worse than that at correcting tests.
This is not a bias against you, its just a general thing that applies to all students and people. Nobody is good at correcting their own work, even the most esteemed professor gets his work checked by other people. And an LLM is not another person here, they aren't good enough to check your work.
Note its much harder to accurately grade an answer than to answer a question.
You are so incorrect here. If you feed it an evaluation rubric it will go after it! You really understimate the technology when in good and well intentioned hands. (Yes lazy people will use it to cheat and get to the answer faster. But it is like a grad student tutor who will do an evaluation in 30 seconds! Rapid iteration and laddering up...)
You've described school. You get told about things and given a hammer to leverage what you've been told.
An LLM absolutely shortens the research part of learning. If I had a human of who had a moderate level of skill who would endlessly answer all my questions, the result would be the same.
You might have a point when it comes to software development because the AI can tell you things but it also just do them for you, at which point, you've learned a lot less. But for non-software things I have to learn things so I can then go and do them.
But even for software development, I've learned a lot of esoteric crap to get interop working on projects that I will probably quickly forget just the same as when I had to spend hours skimming through stackoverflow.
> An LLM absolutely shortens the research part of learning
No, it doesn't. Because in any scenario where you are using AI in a potentially appropriate manner, you are verifying every single source it spits out and cross referencing everything it says. If you do not do this you are failing the process entirely.
If I ask my buddy for help, I don't cross reference everything he says either even though he could just as easily be wrong. A lot of what I'm talking about is every day stuff not nuclear physics.
If I ask my buddy for help I'm not considering that part of research or learning I consider that leaning on their potential expertise. You're not learning shit on how to build a house just because you ask your friend to build one for you.
If we're talking about research as in actually attempting to learn and get to the heart of a matter then yes. You should be cross referencing everything.
So never learn anything? Be entirely beholden to other people for everything? People who somehow themselves also have complete knowledge? I've done so many things myself now and that has given me the confidence to do even more.
If I had to have a complete idea of what I was doing to do it, I wouldn't even have a job. Doing stuff and making mistakes is exactly how you learn.
And I see people say this, that they've built a house but what they've built is something that structurally collapses immediately and fails due to problems that even an average person would've avoided had they learned how to build a house.
So did you build a house, or did you build a house?
How can I argue about such a hypothetical?!? If I ask an LLM or my mother-in-law about how to get wine out of the carpet, and I get wine out of the carpet, maybe I have done it wrong and my house will burn down! Who knows?! Your argument is meaningless; there's nothing anyone can say to that.
And if you ask an LLM about an effective cure for cancer and it spits out to drink bleach, are you going to follow it?
When you ask 'who knows' that's the point of research which was my original comment here. The same goes true for some random asshole telling you to drink bleach as it does an LLM, except people seemingly have hyped themselves into believing the LLM is more right somehow instead of being trained on every random asshole ever.
Time helps to some degree, but I've worked with professionals who've earned a living for 30+ years, enjoy strong reputations among their peers, and who absolutely do not know what they're doing.
You think you're learning things at a pace you never imagined, but all you're really doing is learning superficially. You never really go deep into things, and it "seems" like you're learning faster, but how much can you actually recall of the things you learned? By the way, I'm saying all this because this has been my experience. I can prompt like no tomorrow asking questions, but I forget most of it because it's not deep learning. I'm not actually thinking myself, I'm allowing the AI to think for me.
How exactly is AI helping you with sports? Have you become a better athlete? Have you learned more about the strategy of each sport from the perspective of a coach for example? Care to elaborate?
recently vibe-coded a personalized skateboarding block-training app. The platform maps curated YouTube tutorials to specific tricks I want to learn, logs my daily progress, and tracks failure rates alongside the specific reasons for missed attempts.It also functions as a structured video journal that integrates seamlessly into my coaching routine. For instance, during a session of 10 daily kickflip attempts, the app logs my success metrics and video capture so my coach can review my exact mechanics. Over time, this data-driven feedback loop visualizes my progression curve and systematically improves my landing consistency.
In another case I follow a bodybuilding cutting regimen and it helps me create and track recipes consistent with my diet plan and macro guidelines. It helps me also create tasty recipes that fit my criteria based on the ingredients I have on hand.
Those are just 2 examples.
I also recently built a backyard jib setup with a platform, ramp, PVC jib rail for snowboarding, and it helped me architect the design for it.
I actually agree with you; I definitely feel I'm able to use LLMs to learn and explore concepts, but I've always been a self-taught and highly motivated person in the first place. Everything technical that I know, I know because I put in the work to learn it on my own. That's why I wish these tools were being advertised as models that help you do BETTER work, not MORE work. They're using them as excuses to lay off swaths of workers, instead of allowing them to uplift people's skillsets. And, of course, it invited a whole mass of individuals who use them to artificially elevate their perceived skills.
Are you actually learning them or are you just letting the AI kinda do them for you? There was a time I knew how to take a path integral, but kinda-sorta knowing what something is and how to ask a calculator for it is different than knowing yourself.
I literally just started playing around with Lathe (https://github.com/devenjarvis/lathe) which was shared here on HN a few weeks back. There are some things I have been wanting to learn more about (some for work, some for personal desire), and I used it to generate a tutorial about some of the things.
I have been very pleased with the results so far. I was able to tune the tutorial to exactly what I want to learn, and it did a very good job (at least that i have seen so far). It has made learning fun, since I get to learn exactly what I want, and I can ask the AI questions and have it make changes to the tutorial in real time as i am working through it.
Now, will I keep using this at a rate to fully offset all of the thinking i have stopped doing since i started using AI? I am not sure, I guess time will tell.
Yea, this is a serious problem but honestly it was present before LLMs. Actually knowing, deeply, how to do something was always very difficult and very time consuming. People would substitute it with “edutainment”: youtube videos, Ted Talks, blogs, etc. The thing is nobody was really learning deeply. We were training our brains for recognition, but mastery requires training for generation which was always the harder of the 2 cognitive processes.
"[Learning] new sports" strikes me as an especially odd one. I can see how an AI tool could help to learn the theory or perhaps come up with better training or match-day/racing strategies, but it won't short-cut the work of developing the necessary physical skills, will it?
i could not get through the hurdles of installing an IDE and js/python modules before.
now i am learning basic scripting and data modeling etc.
it is phenomenal for learning languages.
i built a chicken coop and some furniture. the skills and confidence i gained are real. am i failing to learn certain skills in the process? of course. but I'm getting further then i would on my own, and that is truly meaningful.
you can keep dismissing it; but I'm genuinely using it to break down barriers, give me confidence, and highlight my ignorance in very productive ways.
i find it bizarre how unwilling some people are to recognize that.
> i could not get through the hurdles of installing an IDE and js/python modules before
You want us to believe you couldn't overcome the puddle-deep challenge of installing an IDE and using Pip or Node in the past, but now you're actually learning how to write functions?
Cool for you if true I guess but I'm pretty seriously skeptical
Python environment management is notoriously reticulate. I can totally imagine that they got stuck on some path bullshit so that the installed packages were on a different Python than the one they launched in the IDE. I've been using and configuring computers since they had DIP switches and if I were teaching someone to code manually I would absolutely let them use Claude to get their environment set up.
I can sort of buy this, honestly. When was the last time you picked up a new language with unfamiliar tooling? Especially something older that won't hold your hand.
A lot of the most miserable parts of getting started coding have nothing to do with programming and everything to do with like, trying to apt-install the right compiler version or figure out which build headers you need or some other equally trivial bullshit that gets in the way of writing code.
> I can sort of buy this, honestly. When was the last time you picked up a new language with unfamiliar tooling?
I started learning Monogame with C# literally yesterday after being a chiefly JavaScript dev for over a decade. I'm having a lot of fun learning it from scratch with no AI
I'll concede that before yesterday it had been a while though
> Especially something older that won't hold your hand
Python and JavaScript are two of the most "will hold your hand" languages these days aren't they?
> A lot of the most miserable parts of getting started coding have nothing to do with programming and everything to do with like, trying to apt-install the right compiler version or figure out which build headers you need or some other equally trivial bullshit that gets in the way of writing code.
I agree but absolutely not for these two languages in particular
You can open the developer tools in any major browser and start typing in JavaScript right there if you want
Python you can install it and write it right in any editor
The upfront burden for both languages is absolutely trivial imo
I agree with the sentiment of the parent comment. But I also sometimes question what you have posed exactly. Am I really learning or getting the curiosity itch scratched. The line is very thin sometimes. But I think I am more in the camp of learning. I can ask so dumb question that I never had the courage to ask nor were they entertained where I was educated. It has been a boon to finally ask.
For example I am following Dirac's book "The Principles of quantum mechanics" to study QM. Pre-AI wouldn't have been able to do so, I am just that dumb. Even with AI its tough. but the thing is I can keep asking questions until I get that concept drilled in. Now I am doing it at a pace that's unfathomable to me.
But now that I am getting to grips with QM, I can get to things that I am really interested to learn like spin resonance and so on. This is something I am so grateful for.
Now it can be questioned that is it making me wise, intelligent or just "giving me answers" that I should strive to discover myself. I dont know the answer to that. But studying what I want, how I want and not getting judged is something i deeply enjoy. Srry the comment might have taken some tangents.
> For example I am following Dirac's book "The Principles of quantum mechanics" to study QM. Pre-AI wouldn't have been able to do so, I am just that dumb
I have the same experience, but there's another dimension I want to throw out: breadth versus depth.
I've wildly increased my breadth of learning. If I'm ever curious about anything, even a passing thought, I can scratch that itch in a way I never could before.
But am I going deep? Acquiring new skills? Eh... I usually go far enough to unblock myself and/or settle a curiosity. I don't think that's good or bad, but it does present a certain set of tradeoffs that are different than going deep.
Yeah, it is like going to a restaurant to taste a foreign meal rather than looking up how to cook that foreign meal yourself, cooking it and that way getting to experience it.
Tool use typically follows this curve. If you want to preserve a skill you have to actually preserve it. This isn't inherently bad by itself, tools enable us to do much more than we can without them and its a point of contention whether or not any skill is inherently important when a tool comes along that does it for us.
One of the challenges here is that the skillset we are in danger of letting atrophy is essentially unbounded. It’s not a specialized tool like a calculator, where you have a well scoped domain of problems you are offloading. granted, in practice many people are using ai for specialized domains (like coding or producing visual designs). But whatever level of abstraction they are currently working at is not, in principle, something that they couldn’t also offload to ai.
The instinct of many people (myself included, a lot of the time) is "feed the AI the problems from the domains I'm least good at / familiar with" because those ones are the most frustrating and where I'm the most likely to spin my wheels. It's easy to imagine how this feeds into a cycle of "only put forth effort at things I'm already good at" and a consequent narrowing of professional / intellectual development.
The huge problem in this specific case is that to use this tool well you also need the underlying skill to be developed and preserved. It's very different from a power drill.
The difference if this is okay is how reliable the tool is. If it is a calculator or compiler it is okay. The example in the article also sounds okay (machine learning image classification), though I am not sure.
LLM output is unreliable, so we still need to judge it. If I want to be able to judge code, I must have worked with it to a certain extent. So the unreliable tool does not help me much if I don't want to accept the unreliability.
Sure. But if you don't own the tool and it is held by a cabal of centralist (even political state-adjacent) parties, you're having a bad day when computer says no.
If they can keep up. Unfortunately, we learn from previous technology shifts that the masses will always favor ease of use (to the point of infinite scroll 5 second videos dopamine puddle, or echo chamber social networking in lieu of critical media consumption), which does not bode well for the market for alternative hardware: one which is already expensive.
I fear on-prem AI is likely to become as popular as on-prem servers without Cloudflare using self-hosted email are today: that is to say, people have heard of it but the skillset is almost popularly eviscerated, external policies make it progressively impractical, and anyone who does it is 'niche'. While basic guides will exist, obtaining top-level output will probably require many moons of concerted effort.
Consider another perspective: They don't have to keep up. Once a model is good enough for a task, the model can stay. A hammer is a hammer and a hammer from 100 years ago still has most of its utility.
Similar to the hammer, it's not unreasonable to think that some classes of work will simply be solved by some model generation and whatever happens at the frontier after that does not matter all that much for work that puny humans do.
Then, of course, there will be a time where all of this is moot: Absolutely no human will want a human to diagnose their medical issues. That is not a skill deterioration issue. We simply will concede that we are not able to do it as well as a more capable system can, and without much fanfare, increasingly delegate, as we have always done.
Carrying your analogy further, let's assume all human jobs fall under good enough open source models. All human problems (food, shelter, not clobbering each other over the head because of monkey genes) are solved through a combination of AI and robotics. Maybe we even remove governments, police, and live in a future post-capitalist ecotopia.
Even if this occurs, and I don't trust well-resourced humans to allow their existing apex-predator positions in present era capitalism to be overturned, the action - as far as either humanity or AI is concerned - will still be at the forefront of possibility: a front by definition invisible to old models. And someone has to pay for the hardware to be there. Do we (a) allow private-sector dominance, effectively depowering traditional nation states and empowering a private cabal beyond historically conceivable levels (b) nationalize thought (c) head in sand and pretend it will all go away?
Most of the world seems to be with strategy C right now, strategy A is the advancing default and has already achieved extra-terrestrial reach with a threat of extra-terrestrial persistence, and strategy B is potentially scarier than the other outcomes if it goes wrong but might be lovely, if you believe in nordic state funds, solarpunk futures and socialist utopia.
Interesting times. By the way, if anyone with AI capitalization reads this, I'm looking for investment to feed humans more efficiently and have a NASDAQ reverse merger under negotiation and effectively priced out with board buy in. Just need capital support. https://infinite-food.com/
On-prem versus cloud inference doesn't matter for concentration of power.
Concentration of power exists when the model makers are the same as (or control) the inference providers. Making a model is capital intensive, so there aren't many of them. Providing inference is not: I don't even need to own GPUs; I can rent them from those who do and then sell by the token. B300s cost less than $4 an hour currently.
Cloud can even be more effective at lowering concentration of power than on premise. Asking people to individually buy $20,000 of compute equipment plus power and cooling equipment to run a frontier model is not something they're going to do if they can just pay four-tenths of a cent per output token. If the only cloud inference providers are the big proprietary US titans, that means you're going to get far more power concentration than if open source inference providers are an alternative, because then I can just switch my API endpoint.
Eh, this is our species first contact with that type of technology. A good number of voices see how deleterious these things are, and it’s all still very new. Future humans will tell parables about the evil tech bros and their silly obsessions, and the unequal accumulation of capital. This will be seen as a dark and stupid time, but I think we’ll persevere - the tech bro set is much weaker than they imagine, and certainly than they project.
We love our open source models. GLM 5.2 came out recently and the timeline for closing the gap to closed source shrank to something like 2 weeks by popular measurements.
Edit: Mis remembered the timeline I saw not 2 weeks, 3 months, but still I think my point stands.
Just to make sure I understand your argument. Are you saying that today's open source models are on par with frontier closed models of two weeks ago? By what criteria?
Commenters there were saying GLM 5.2 was roughly equivalent to Opus 4.8 in coding prowess, based on personal experience of the people commenting. Opus 4.8 came out on May 28 this year (so more like 3 weeks ago), GLM 5.2 came out 2 days ago.
I haven't tried this specific model, but you can understand that a lot of HN testimonials are bound to swing extremely pro open source. I certainly hope they are as good. But I have personally tried a lot of models that are supposedly as good as the frontier ones and have found them lacking.
After a while, you do start to start to skip a couple rounds of open source models until there's a notable release. That, and the resources needed to run them are increasingly bought up by the owners of frontier models
The biggest negative I'm seeing is that people are moving _too fast_ to evaluate the stuff they're slinging. When you move too fast, you do not take the time to build up your taste or appreciate the nuances of different approaches to the same problem.
I am using LLMs quite a lot, but the amount of time I spend sitting on some slopped out code is I think on average much longer than a lot of my peers. What I've found is that while the original thing "works", it usually winds up being another 2-3 cycles of iterating on the original idea after I've let it settle in my head before I actually feel confident about merging.
As a result, when I add it all up, for actual "this is important" design-level concerns, I do not feel significantly more productive.
This is hardly surprising. Think about Staff+ engineers: their work is a lot like commanding AI. Most climb the corporate ladder through hard work, excellent engineering skills, soft skills, and a bit of luck. But as their companies grow, they gradually spend less time coding, debugging, or doing deep design.
Instead, they act more like highly technical product managers. They help VPs plan and write high-level product requirements sprinkled with technical terms. They draw boxes on whiteboards and create pretty slides. They write polished documents that keep their leadership happy, and they are either in meetings or on their way to the next one. When they have a technical idea, they dispatch a team to test it out.
Naturally, they still feel deeply technical, until the day they have to resolve a production issue, pass a technical interview, or write extensive code. That is when they realize their skills have grown rusty.
I point this out not to criticize, but to highlight a genuine career challenge. As an engineer, I would rather hone my technical skills. Yet, if you want to climb the corporate ladder, you have to take on more organizational work. The only solution I can think of is to become more like a researcher or a professor. Over the years, good professors spend less time writing papers or deriving formulas. However, their insights are so deep that they still produce amazing results by advising PhD students. But that path is much easier said than done.
Welcome software engineers to the world of automated production. Your skills will drop, the usefulness of your skills will drop, the pay for your skills will drop, the quality of output you produce will drop, everything will be worse for everybody except for the person that legally claims ownership over the machine and what it produces.
100 years from now people will wonder at how anybody did anything manually writing all their code, but at the same time will have their mind boggled that a game like Roller Coaster Tycoon is possible without 1024 CPU cores running a few terabytes of memory in order to control every single person walking around the park.
Of course. If you don’t use something it atrophies not only in non use but in losing interest in keeping up with the state of the art in said tech.
What we gain though is for people don’t possess that knowledge in the first place, now have this superpower. I know several individuals who have vast experience in specific disciplines and they are now able to solve real problems where there were previously struggling and having to make existing solutions work.
In the context of software engineering it allows people that have great institutional knowledge bypass the software market and construct stuff on their own - or at least prototype something and turn it over to an SE if the situation dictates.
I’ve been using CC for several months now and have noticed an increasing quality of output - Fable 5 I think was 85% there. At 95% SE’s are going to be increasingly looking for work to do.
To the title though, I’ve noticed while my desire to actually write code is decreasing CC is forcing me to improve my high level thought processes in the context of overarching goals in a project through discussion with CC. The software often introduces things that had escaped me or just think more outside the box.
My concerns are that this technology will be restricted at some point and the people making the restrictions will have a lot of control - and we know how that works out. But I believe they are inevitable, first obvious example being Fable 5. Are guardrails needed - yeah sure. Common sense says that I don’t want someone able to concoct an easily transmittable Ebola virus that has a 90 day incubation period in their kitchen but I do want an entrepreneur to be able to build a competitor to MS Office, or a cure for Ebola, for example.
> What we gain though is for people don’t possess that knowledge in the first place, now have this superpower.
Indeed. The great innovation of AI is giving people with wealth access to vast amounts of knowledge, while limiting the amount of wealth that people with knowledge can access.
All the benefits you're describing apply to the present moment; people with knowledge, self-discipline and expertise can leverage LLMs to great effect.
How many people like this will exist in a decade? Two?
Meanwhile the people with mountains of money can leverage them to enough effect to make even bigger mountains of money, by hiring less-than-skilled people who will do way more with it than they could without it.
Is this a situation where AI will go away and we will regret the loss of skills? At worse, we will be forced to use open weight models instead of the cutting edge, so I don't think it's a big deal. I'm sure people got worse at arithmetic after the invention of the calculator.
I don't think the real threat is at the individual level, but at the societal level.
Building skills over time leads to insights that lead to innovation.
AI does many interesting things, but it doesn't innovate (yet).
The real threat isn't that we'll all lose our skills (possibly) and then lose access to AI (unlikely), it is that AI will remain at roughly currently levels and we'll dull our skills due to reliance on it and innovation will stall because we've offloaded too much of the thinking to the non-innovative machine.
I'm not saying this is what definitely will happen, but it does seem like a very possible outcome.
The number of people now involved in software development has now increased because of a lower barrier to entry. I know many people who would previously use a no-code tool or hire offshore devs, or simply not have their problem solved, who are now vibe coding. Many of these people couldn't write very much code manually if they had to, but they're closer to understanding software than they were previously.
Yeah, this describes me perfectly. I can't really program, but I'm now building a bunch of different projects and submitting PRs that people seem to appreciate.
One of them, a Home Assistant integration for controlling adjustable beds, would be borderline impossible to do well manually - I've vibe-reverse engineered the Bluetooth protocols of more than a 100 Android apps.
> I've vibe-reverse engineered the Bluetooth protocols of more than a 100 Android apps
Impressive coverage, but how do you verify that the reverse engineering matches what the hardware expects? Do you find people who own those beds and get them to test stuff for you?
> The number of people now involved in software development has now increased because of a lower barrier to entry.
I don't see what this has to do with what I posted.
Yes, AI lowers the floor on software development, and there are positive aspects of that, but that doesn't change the possibility of an innovation stall.
re: concerns at a societal level. More people are making software than before. People who have been coding for a long time have moved up an abstraction layer and are further from the code. But many people are actually closer to the code than before.
> But many people are actually closer to the code than before.
More people are coding, I wouldn't say they are closer to the code if they are vibe coding. Are any of them going to produce the next breakthrough in computer language/framework/method of development/etc?
The risk of AI is that we dull the skills of enough people at the high end of the state of the art of the nuts and bolts of software development that we slow down innovation on that end. That's the concern.
Previously-non-programmers vibe coding CRUD apps they never could have before is all well and good but really has nothing to do with this concern. They may create wonderful and successful businesses but they are irrelevant to computer science related innovation.
There are a lot of mundane coding skills I consciously put off learning in case they'd ever become obsolete, and now I'm glad I did. Like sure learning React was good, but Angular? That boilerplate is Claude's job now. Ruby? Forget that.
For me personally, holding a dialogue with several agentic dev streams in parallel keeps me very much on my toes mentally. It's quite exhausting, and certainly not letting my brain atrophy.
I'm probably losing some coding skills, but replacing them with different ones, and honing some others.
I used to manage dev teams of 20+ people inside high pressure, high stakes projects.
I've been coding all my adult life, on big things and small.
To me, agentic engineering is a deja vue of managing teams, except in real time.
One thing I have found out immensely useful for is exploring new niche areas.
One example is I got it to setup a new terminal (wezterm) and configure it quickly to my liking. This would have been a lot of googling and reading before, generally too much time for me to invest.
I don't think it's valuable knowledge for me to have apart from a high level to evaluate the possibilities.
Humans will become individually and independently less skilled while having access to tools that allow them to do far more than even the most skilled human could, before having access to these tools.
I'm not sure if we'll become less intelligent. I think our sacks of neurons are gonna keep on making associations, just across a different set of topics.
MDs NPs and PAs currently are offered free access to a medical AI app called Open Evedidence. Was just discussing this with our chief medical officer and while he is quite enthusiastic about it, he could already see this same dumbing down effect emerging among his providers.
The solution I think, is 5 or 10 ten year re-certification exams for MDs, currently reqd for "Physician Extenders" but not the Physicians who supervise them.
Continuing compulsory education and re-certification works, I suspect, in all highly skilled fields that are both augmented by AI but also degraded by it.
My compiler writing skills atrophied with the advent of high-level languages, but in exchange I got more done. There is still a very well paid market for compiler writers, but the fact that not everyone needs to be one has made the world richer overall.
As jatins says: Why bother with source code? just deploy prompts and specs. It’s rare for C programmers to care about assembly, or python programmer to care about C. Why do you care that much about source code?
They really are though. There seems to be some confusion about "determinism" as if everyone writes perfect code and every compiler interprets that code perfectly. There's bugs and nondeterminism in every step of the chain with human coders and, say, a .NET compiler too. The idea that we've got coding down to a pristine and exact science and AI is some random chaos element that muddies it makes no sense to me after seeing how most human created code works.
New tool that does task better than worker leads to workers being less good at task. Net outcome for patient is positive. Next?
Programming: "for a given task, if you take a shortcut then you will not have the familiarity and expertise that someone who took the veritable and righteous path would have".
The question is then, what did you do with the extra time. If it's fuck all, then yes, that's a liability.
Like any technology, it comes down to the disposition of any given person in how they plan on applying it.
Not trying to say it's all going to be awesome. Definitely maybe the opposite. These arguments are weak tho.
If your argument boils down to "individuals just need to use it the right way" then I ask you to kindly use that same framing for driving cars, owning guns, etc. and then rethink your position.
It turns out that when the use of a tool has external consequences for misusing it, it's important that there are structures in place for penalizing the misuse of that tool.
I don't feel like AI coding has ruined my skills, and I could go back to manual coding any time.
However, I cannot build a good mental model of a software component that I didn't write myself, and that can affect future maintenance if that component is not properly decoupled from the rest of the system.
Autocomplete of entire functions and methods. Nice, but also really boring. Takes the fun out it. It's all about fixing sup-par code now, a line here or there.
It's just boring. I tried writing some code by hand today after a few months hardly thinking about things and it was really hard to do even the simplest stuff.
I suppose tractors and cars ruined our skills at feeding and caring for horses.
Losing a specific skill to automation isn't necessarily a bad thing. Losing the ability to learn things would be however, and that would be my fear with AI, but I'm not sure it's well-founded. Humans learn naturally by interacting with the world.
Some effect is real, but it's likely overstated by poor metrics.
"Currency" in all fields relates to the recency and frequency with which you dealt with a particular issue. Whether flying on auto-pilot or coding with AI, automated reduces some currency. But is that a reduction in capability?
Measuring concrete tasks makes currency the operative skill; that's why it works to cram for standardized and mid-level tests.
(Indeed, the 2010's interviewing "wisdom" about people being quick to answer simple questions veered into measuring currency, not skill.)
I think this effect is strongest in time-impacted professionals. Doctors doing dozens of endoscopies a week and developers churning out code will use what tool leverage they can, and forget as much as possible to focus on what they need to. I suspect the effect is weaker in personal or research projects.
People riding bikes won't be able to run long distances - because they won't have to, and will be able to outdo any runners. That's only a problem if the supply of bikes is someone constrained. So the risk is not skill loss, but losing control of the means of production.
How do you measure the bundle of "skills" that comprises critical thinking? And if that's your analogy's distance running, what's the "riding a bike" analogue?
No, AI isn't doing anything. When someone gets stabbed, do you say "are knives stabbing people?" No, just like you don't give credit to the knife for cutting up your vegetables.
"Just being aware that this phenomenon exists hopefully provokes some self-reflection about which skills people want to maintain and which they’re willing to outsource” to AI tools. Right. Obviously.
So we need to be teaching that core lesson to children -- they don't retain skills that they don't practice. And we need to be careful to decide what skills and verify they are learning them. We also should absolutely be using AI to provide personalized instruction to every single student.
Blaming the tools for things that humans do is incredibly stupid and dangerously misguided. Because it shirks responsibility onto the technology, when technology is the best lever humans and society have to improve things! It just happens to also be the best lever available to make things worse.
This negative view of improving technology starts from a warped and very unrealistic concept of the state of the world, where it has been, and the role technology has played.
1. Technologies, starting with fire, the printing press, etc. have been critical in raising life expectancy, standard of living, etc.
2. The world is still a profoundly unequal and exploitive place.
3. AI and robotics have the potential to provide everyone on earth who wants it with extremely inexpensive labor to help them with anything they need or can imagine. This will be a dramatic shift in quality of living.
Human society is the source of our problems, not technology. Part of this is that I think deep down people believe that any tools or developments that arise will just be used to exploit and suppress them more, and there is no alternative. In this case, I guess people think the best outcome is to go back to feudalism or some nonsense because technology just makes things worse.
But why stop there? Why not go back to, I don't know.. fire? Or maybe no one should ever eat any red fruit?
AI should demand tokens of attention from proffesionals little challenges to continue, turn the toolise into an adventure. The gold standard would be the primer from diamond age
The bit about computer science (behind the paywall) starts:
> To investigate whether skills are being lost in the field of computer science, researchers at the AI firm Anthropic in San Francisco, California, designed a randomized controlled trial in which 52 software engineers were asked to perform a basic coding task
Of course it's ruining skills. Anyone who has a really well honed mental skill knows that we get rusty really damned fast (e.g. musician at a professional musician level, which happens to be my reference point). You keep shedding, you stay sharp. You take more than a week off, it's going downhill.
There's no way this isn't happening with other skills. No goddamned way. Anyone who tells you otherwise is grifting.
Back in the mid-90s I was doing desktop support. It was a lot of work because PCs were relatively new (and they were garbage!), and people broke shit all the time. Sometime around '96 a disk-cloning utility called Ghost was released. It was great - one could provision a fully working PC with all required apps and config settings in minutes! Sounds lame now, but back then it was revolutionary. It had a dark side, though. After about a year most people I worked with had lost the ability to troubleshoot even the most basic problems. The solution to every problem was to just re-apply the standard Ghost disk image (we called it 'Ghosting' back then) ... Can't print? Ghost it! Not receiving emails? Ghost it! Word is too slow? Ghost it!
Happened with all of tech support, really, or at least in my corner of the world: you take your PC to a technician and you receive in return a fresh Windows install, a folder with most of your files (including a copy of "C:\Windows"), and none of the programs or shortcuts you had before.
> To investigate whether skills are being lost in the field of computer science, researchers at the AI firm Anthropic in San Francisco, California, designed a randomized controlled trial in which 52 software engineers were asked to perform a basic coding task3. During the exercise, all 52 participants could search the web and access instructions on how to do the task. Half of the participants were prompted to use an AI assistant as well.
> Afterwards, all of the software engineers were asked to complete a quiz about what they had learnt from the task. The participants who had used an AI assistant did significantly worse on the quiz than those who hadn’t: the average score was 50% in the AI group versus 67% in the non-AI group.
This doesn't strike me as a great test? Most engineers aren't going to learn anything from a basic coding task anyway, so I do wonder exactly what they were testing there. If it was just recall about what the issue was, then it doesn't really strike me as a problem - using AI to handle simple problems that it's clearly capable of dealing with is the right way to use it, and of course you're not going to spend time poring over the details because then you haven't saved any time by using AI.
There are other examples that don't strike me as particularly problematic, like GPS eroding people's sense of direction. It's totally reasonable to let a skill atrophy that you no longer really need because you have an ever-present tool to handle it. I'm a lot worse at doing long division than I was when I was <whatever grade one learns long division in>.
The whole skill atrophy thing seems like much less of a problem than it's made out to be. We've been letting skills atrophy for good reason long before the advent of AI. If you start at McDonald's as a fry cook and work your way up to regional manager, if you suddenly have to work a shift on the fry station you're going to be worse than you were when you were doing it all the time. MDs at investment banks almost certainly can't put together a pitch deck as well as the junior bankers who are doing that task regularly. These things are fine - part of moving up in the world and having a broader impact is being able to successfully delegate tasks, and when you delegate tasks your skill at those tasks will atrophy. No real difference whether you're delegating them to AI or not.
To be clear, there are of course cases where skill atrophy is bad. iLoveOncall posted about senior engineers in their org who have lost all of those skills and their judgment along with them. That's definitely bad! If you delegate so much that you lose the ability to even judge good work, now you can't even delegate effectively any more.
I think the real lesson with AI is that you need to be self-aware about what skills you should practice and retain vs. what skills you can let atrophy, since it's easier than ever to hand things off. I've lost most of my ability to write a SQL query, but that's fine because it was only a skill I used intermittently and AI can always do the job fine at the level of complexity I need. I have not let my skill of writing product specs atrophy (I am a PM, in case you haven't read my username), because that's critical to using AI correctly in the first place.
How many of you feel this, anecdotally? I had to come up with a class hierarchy for a roguelike game the other day. This should be something that's dead simple, off the top of my head, no problem.
And suddenly I was stuck! It was like thoughts weren't forming properly. My instinct was to use Claude to help brainstorm, but I resisted. 5 minutes later, I finally broke free and instantly came up with the plan.
What the hell?
I realized I'd offloaded my planning onto AI. I would ask it for plans and then choose the best one, but that's a different skill than coming up with the plans in the first place. My skills were rotting.
I don't know what to tell you bro. I type much less code but I haven't stopped thinking altogether. I still write to myself and to others. I read longform. I walk and introspect. I think about high level problems. Any general cognitive decline I ever notice, I attribute to having a bad day or getting old...
I think it’s like shoes. Shoes are generally better than callouses at protecting your feet, providing traction, etc. Almost everyone accepts the trade-off and wears shoes as opposed to going barefoot. To be sure, your feet lose some of their toughness. You may become reliant to some degree on the tool. I know that when my kid started wearing shoes, walking became actually a bit harder and be fell a few times. The shoes were a little clunky and lacked the same sensation. He changed he gait to accommodate, but over-compensated. They took a while to get used to wearing. AI also often feels a bit clunky and lacks the same feel. However, I think some people are confusing the shoes for the feet. The shoes don’t do the walking. The LLM is the words, but not really the ideas. If you stand or pose in shoes, move your legs around and wait for the shoes to take you somewhere… that’s not going to work.
At least for writing, I think AI is mostly useful for the types of writing that aren’t particularly interesting or worthwhile in having to begin with.
In concrete terms, AI isn’t all that useful for writing a personal blog, because no one wants to read obvious AI slop. But it is useful for creating boilerplate product pages, FAQs, and other types of writing that weren’t very interesting pre-AI.
So it’s not really a huge deal to me that my skill for writing descriptive product page text or FAQs is atrophying, assuming that it is.
Not worthwhile feels a bit strong (a good FAQ is definitely worthwhile!), but I definitely agree that there is a big difference between any kind of art (writing, playing music, creation of images/videos/etc.) for its own sake and for commercial purposes. AI is terrible for the former but perfectly fine for the latter.
There will always be value in a human writing fiction or a memoir or even a Substack. The human perspective is inherently valuable there. Much less so with ad copy that's just going to get A/B tested ad infinitum until a winner is picked out based entirely on data.
Same with visual art. Art painters aren't going to lose their jobs to AI, but once you've got a robot that can paint a house reliably, house painters are done for.
What is most troubling for me is seeing kids just switch off when LLMs are available. Doing homework they will have zero interpretation or contemplation, just enter the question as a prompt and record the result. LLMs appear to have the ability to interfere with the most basic aspects of attention and executive function.
That’s got nothing to do with llm’s and it’s got nothing go do with ‘how technology is used’.
I’m not even sure why this has to be said.
Universal fact 1) most people lack the self discipline to choose what really benefits them and takes care of their well being. We already know this - obesity rates, excessive screen time etc.
The solution is to control the actions of people. But most of you ain’t ready for that conversation.
AI has allowed me to keep shipping features and system even when holding a normally managerial position, so if anything it preserved some of my coding skills. I'd not have seen any code otherwise (writing code is a huge time sink compared to managing things around an org).
I pity those who need to contend with that as ICs, though.
Hopefully this is not your case but as an IC the “manager, who didn’t write code for a couple of years” that decides to come back and “help” is one of the worst experiences. Code is usually subpar, uses old idioms, and comes with added pressure. With AI this can become much worse, because of the sheer volume of code that can be thrown at the IC.
My current position is more on the operational side and AI allowed me to create a pretty significant software system for our technical ops, that the wider org liked and given me the resources to recruit flesh-made engineers to support.
AI/AR Glasses that show you every piece of knowledge about whatever your looking at might not help, especially if the AI is wrong. Otherwise everyones a know it all?
It's going to get a lot worse than AI being unintentionally wrong. We're only at the start of the curve for this, but it's obvious that controlling what LLMs show in response to politically-charged topics is huge lever of power for those in charge.
So now imagine you're using Chineses AI/AR glasses that you've come to rely on for "knowledge" and you look at the famous picture from Tianemen square: "Doesn't look like anything to me".
That's a great analogy - everybody knows you forget pretty much everything after a few years away from the line. It should be seen like one of those studies that proves smiling is correlated with happiness. :-)
The two senior engineers in my org (in a FAANG) who vibe-code the most have lost literally all of their skills. Their code has become terrible and their judgment even worse.
All of this makes me selfishly excited for my own future. It's glaringly obvious that anyone who's a heavy user of LLMs is atrophying their skills in real-time. I have yet to meet a single person for whom it's not the case.
But I essentially completely stopped using them for software engineering (why isn't really relevant, but it's not because od this skill atrophy). So as the skills of everyone else is diminishing, mine is proportionally raising.
It has never been easier to get better than others. You don't need to put in more effort, just the same effort as you always have, and others will do the job of losing their skills for your own benefit.
Idk man my system design game is better than its ever been because I put in the effort to use these tools and recognize they can't do software design better than I can and because I've increased the scope of what I'm building I often have to think more deeply about the problem up front. A typical speccing sessions lasts a few hours for me on big work before I have AI start writing that work where I'm just going back and forth on what I want, points of consideration for performance, usability, structure etc pushing back on where AI (always) chooses the most naive way it wants to do something.
Every time I see an anecdote like this, A it reaffirms my belief that FAANG devs are fairly mediocre on the whole (not saying this is you, obviously there are good FAANG devs) and B it reaffirms my belief that the developers who kind of give up their thinking like this are really using the tool wrong or didn't really care about the work before AI either so its now just a quick means to an end.
+100 on this. In addition, if you don't outsource your thinking and you're willing to go through all this, you absolutely don't need the top tier models.
> Every time I see an anecdote like this, A it reaffirms my belief that FAANG devs are fairly mediocre on the whole
I think another (partially causal?) problem is how they're managed. The whole perf circus is just ridiculous, especially the stuff recently reported about facebook. But they're all more or less like that. Steeped in that cocktail of incentives, who even knows what might happen to an otherwise excellent engineer.
But also just numerically, they can't be much above average, on average, because there are so many.
I strongly believe that you cannot evaluate how good a system design is if you don't implement it by hand.
LLMs will implement what you ask them to, even if it is the wrong approach. They can be lazy and take shortcuts all the time, but they do not feel PAIN (obviously they don't feel anything and aren't lazy, I'm just personifying them but you get the point). Only when you implement by hand can you feel if the implementation of your design is painful or not, and only this signal can tell you if your design if truly good or not.
I do think LLMs are useful for design work, they are good at asking clarifications and probing questions which actually do push you to approach problems differently, but leaving implementation of designs to LLM is a recipe for disaster, and judging your own design skills when you're not implementing such designs is seriously laughable. And to be clear, it already was before LLMs, when "software architects" were just designing and then had peons implement for them.
LLMs are enabling a whole new level of bad code that is best describe by the following Jurassic Park quote: “Your scientists were so preoccupied with whether or not they could, they didn't stop to think if they should.”.
> A it reaffirms my belief that FAANG devs are fairly mediocre on the whole
Off-topic but having worked in other companies as well, I can guarantee you that this is not the case. The skill of engineers in FAANGs and other "top tier" companies is much higher than average.
> I strongly believe that you cannot evaluate how good a system design is if you don't implement it by hand.
Fwiw we both agree that LLMs should not design systems. I do the design, but otherwise I don't get how this is true, the success of a design is indicated by long term success in the system it built. You can measure this against success in the task it was deployed for via performance metrics for one. And then from a developer standpoint how easy it was to maintain later on. Success of a system is a measurement over time, but it's not some quality that can only be measured by those who built it.
> Off-topic but having worked in other companies as well, I can guarantee you that this is not the case. The skill of engineers in FAANGs and other "top tier" companies is much higher than average.
I have first hand knowledge of this so I agree to disagree. Being surrounded by google, aws, and meta folks my understanding is the best people leave faang when they get the itch to do something better with their time.
> my understanding is the best people leave faang when they get the itch to do something better with their time.
Your understanding is wrong. The extreme majority of people who leave FAANGs don't leave willingly, and are not the best, unless it's to go to even better companies like Jane Street and the like, but those are rare.
Turns out for most people, there's not much better to do during your 9 to 5 than solving problems for half a million a year.
> Turns out for most people, there's not much better to do during your 9 to 5 than solving problems for half a million a year.
Lol. I don't know I was talking with a guy yesterday who left a FAANG clearing 1.4 million / yr who's now running his own successful startup. My sample is successful founders backed by top tier VCs or exited founders who have done much better than FAANG. If you can't do more you stick around the FAANG, those who can go and do it.
You're talking about what happens to engineers who decide they want to be small business owners rather than engineers. That's orthogonal to this thread about where engineers with the best technical skills end up.
> Only when you implement by hand can you feel if the implementation of your design is painful or not, and only this signal can tell you if your design if truly good or not.
I think this touches obliquely on a point I keep coming back to, that one of the most important things a codebase does is to communicate ideas about how a process should work. Yes, it also produces some binary that runs on a bunch of servers or whatever, but that's a really temporary, ephemeral artifact. The lasting thing is the idea. Making your ideas (expressed in code) easy to understand, easy to work with, and easy to evolve in time is the art of software engineering. I 100% agree, from my own experimentation with LLMs, glancing at something a model has produced and checking that it has some test coverage isn't enough to know whether it's well-engineered. You'll only find out later when you try to work with the code.
I recently did a manual exercise to force myself to keep my skills from decaying too much after about a year of using agents exclusively. My ability to go from a blank slate to software was indeed in the toilet, but my ability to reason over and edit code seems to be surviving fine. I suspect that your LLM-pilled coworkers' judgment issues are related to laziness that LLMs have enabled, rather than an inherent property of LLM use.
You can't be able to do one but not the other, they're the same thing.
> I suspect that your LLM-pilled coworkers' judgment issues are related to laziness that LLMs have enabled, rather than an inherent property of LLM use.
Does it matter? The end result is the same. Maybe the studies that the article mentions are simply showcasing the exact same effect that you're suggesting here, but it doesn't change the fact that there is indeed a negative outcome.
Not true, blank page syndrome, inertia and write vs edit flows are a known thing across disciplines. You might not think about it as much because you have a command in your package manager or framework to bootstrap projects, but imagine if you had to hand code _everything_. I'll guess there are languages/frameworks where you consider yourself quite proficient but where you couldn't bootstrap a project without checking the docs if not for the tooling.
It matters. If LLMs are like alcohol, useful but potentially dangerous, they should be treated very differently than if they're just poison. If someone's a drunk and they can't keep their lives together, it's not the alcohol's fault, it's the person's.
Yes, feeling like I had to relearn to walk. The first week was rough, everything was wired for LLM usage and autocomplete. Couldn't even type right anymore.
> I had the same experience over the past year with early coding harness at the beginning of the year, then Claude code since its release date. But after 1+year going that direction I really don’t want to continue. The novelty is gone, dealing with AI now feels frustrating and boring, I miss engaging deeply with the actual lower level technical challenges. I do not want to manage fleets of agents. I do not want to rediscover for the hundredth time that in fact all this time an agent took shortcuts for acceptance tests I rely upon and didn’t catch. Or once again get the agent to understand why and what I want it to do after its context got bloated and it start to drift completely. While I got artifacts I can use (libraries, tools, docs), including some things that I’m pretty confident are SoA I do not feel satisfied anymore knowing that I used a model to generate them, even if I was the one designing every part of it. I do feel that I’m lying anytime I come to a colleague to share a new cool tool I have made.
> YMMV but I’m personally feeling burnt out with AI coding agents and ready to go back to the old ways for my next personal project
Are you actually seeing any signs that we’re going back to how software was written before, and needing those skills in the same way? Because I sure am not seeing that right now. As someone who vibe codes 100% and has become managements favorite, while being more or less allowed to break the platform every other release I know my skills are atrophying. But it’s taking me different places in my career entirely. There’s a path to managing other engineers now that opened years before it would have previously. Even writing this makes it sound ridiculous, but that’s what’s infront of me right now. There is an entirely other set of skills that I’m interested in sharpening now. Definitely no more sitting down several hours per day and meeting about system design and integrations with others.
It's not really about going back. Evolution happens within a pattern of ebb and flow, back and forth. We never get anything perfectly right. We overdo, then course correct, rinse and repeat. Right now, we're embracing AI, but we're also noticing atrophy of skill as an effect. These may be the last generations of such craftspeople that can notice, compare, and inform as to whether there's actual loss. That future you're seeing for yourself is still being written. Stay tuned.
The article is saying that using AI degrades certain skills when AI is not available. You're claiming that AI is making people less effective even when they have access to AI. I'm skeptical of your claim.
The article's claim is probably true, but not really an argument against AI. Using keyboards degrades my ability to write by hand but that's not a good argument against keyboards. AI will become another tool that allows us to operate more effectively and at a higher level of abstraction. Just like keyboards and Python.
Now, we still occasionally need people who can write assembly (and do calligraphy). But mostly we don't.
As someone who was self taught as a programmer and has a reasonable high level understanding of some CS concepts but not lots of experience applying them, and no good mentor, I’ve found working with an LLM really englightening. Asking Claude to think about “good ways to structure this” or asking how similar problems get solved in industry or high profile projects has really helped me design better solutions and avoid painfully reinventing wheels (recent eg was for a plugin type architecture).
I think a lot of academics and researchers who code but aren’t software engineers or CS majors are going to benefit, provided they take the time to prove what the model does and are curious about whether it’s doing something sensible!
Relative to a 1% coder hand rolling something then yes it’s AI slop etc. but it’s prob still raising the bar generally.
> Asking Claude to think about “good ways to structure this” or asking how similar problems get solved in industry or high profile projects has really helped me design better solutions
I think this highlights the difference between the “how do I make a ham sandwich?” approach of chat vs the “sudo make me a ham sandwich” of agentic coding.
Another tactic is to use LLMs to help you learn. That's another way to approach "It has never been easier to get better than others."
Avoiding tool use because you're afraid you won't be able to use the tool responsibly is not likely to be a winning strategy in the end. Learning to use the tool well is much more effective.
But they're also unreliable in what they present, they still hallucinate. I rather do my own research or listen to a real human on the topic who actually has an internal concept and structure of what they're talking about.
"My friends can't reason as well as they used to, they've been using LLMs to do their reasoning for them." is going to be the eventual outcome. What's the point in getting more done if it ruins your mind in the process?
Writing code seems more like walking to me; at least it is the most manual way of getting a computer program. Horses might be more like one of those low-code/no-code solutions (it really fits, they are useful but very opinionated, so not always cooperative). And, the situation with AI seems a bit worrying for them.
To continue the modified analogy, if your friends lost the ability to walk, you’d be quite worried, right?
It’s absolutely atrocious in the Fortune 500 tech sector. Multiple times a day I see people debating each other on teams about things they know nothing about and it’s obvious that it is just copilot v copilot. Crazy.
I call it Alien Slop Intelligence (ASI) that gradually turns your mind into slop. AI is sloppy on details, so at first you polish its slop by hand, then you start ignoring small imperfections, and eventually you lose taste and skills to evaluate AI output. At that point your mind has become slop.
A skill which is now done better by a machine is no longer a skill, it is technology. It is just a matter of time before most of our logical and language reasoning skills are replaced by frontier model-agents, which will at some point be far superior (if not already) to human capability.
So I totally disagree with this premise that human skills are being ruined by the use of AI technology. No, many human skills are being made obsolete. That's a good thing for economic productivity as a whole, but for those who only have skills that are being automated, their labor value decreases (which is usually bad for them as individuals).
You miss the point, the point is that by using AI, our skills (let say "coding in Rust") are diminishing and even without reading this article, we can feel it to some extent already if we aren't lying to ourselves, especially very heavy AI users.
We do however create new skills, skills that might be more relevant for the future, but still, it is controversial.
It's not good news if the AI companies have to raise their prices 10x to deliver the same service while paying down the crippling debt they've incurred getting everyone hooked on AI.
If my mechanic started charging 10x more to fix my car, I'd learn to fix my car.
We were so spoiled. We got fat salaries to sit in air conditioned Herman Millers all day learning about computers. Now we discover a way to synthesis intelligence and the only thing we can think to do with it is ruin the most fun career most of us could have ever hoped for.
Sure, we're all more productive now, but how much of that is because we leverage AI on top of the intelligence we gained from all of that manual work? Who is to say that in 36 months you're not a worse developer over all because that systems knowledge starts to atrophy too?
This isn't me saying you shouldn't use AI. I use it all of the time to do useful side tasks like to setup GitHub Workflows while I write a feature, or with my agent on a VPS to do internet tasks for me. It's nice to have a little synthesized intelligence.
What isn't nice is to supplement your own intelligence. I think the gains are in the work there--similar to how you can be absolutely ripped from taking steroids while destroying your body. Often it's the shortcuts that are the most treacherous path.
> Sure, we're all more productive now, but…
I’ve been trying to ask people a different question: sure, we’re more productive now but to me, the AI era is only serving to plunge us deeper than ever into producing more, more, more, faster, faster, faster. And for what? What’s it all for? I became a software engineer because I have a lot of fun writing code, thinking through and solving complicated problems, and experiencing the reward of seeing what I’ve built by hand working for the first time.
Do people really have fun managing a fleet of agents that generate the code instead? Or is it just the rush of producing something extremely quickly, much more quickly than you might be able to alone, regardless of how well (or poorly) it might work? For me, being able to move quickly was never the fun part.
It’s one thing to utilize AI to lessen the drudgework, the boilerplate, but I look at people who have gone all in on agentic development and it just really makes me wonder.
> I’ve been trying to ask people a different question: sure, we’re more productive now but to me, the AI era is only serving to plunge us deeper than ever into producing more, more, more, faster, faster, faster. And for what? What’s it all for?
Many devs here have stated that the fun part for them is seeing the end product, not the act of creating it. Using AI is an act of need satisfaction.
Unfortunately, cloning a GitHub repository or downloading a Squarespace template doesn't hit the spot, because you can see exactly where the code came from, so your brain knows you were not the one responsible. AI's greatest feature is that it obfuscates provenance. You can now happily clone that repository or download that template without the feeling that you're cloning someone else's repository or downloading someone else's template.
2 types of people:
People who use coding as a means to an end, producing a product
People who enjoy process over product and coding is the enjoyable part, not the end product
Company executives fall into bucket 1. Even if you love your cushy air-conditioned job, doesn’t mean the people above you don’t see you as a means to an end, a better product.
Solo founders and small startups are in bucket 1 as well but that doesn’t mean that don’t enjoy coding, just the product being made is much more satisfying.
The difference is about slop tolerance not "process over product".
Brother, there was slop before AI and slop after AI.
Pre-AI slop got a pass because the creator probably worked tirelessly to make it and that process gave it value.
Post-AI slop get criticized because AI can create it so fast and easily.
But the builder has not changed. Because the builder knows that the product depends on iteration and polish. And knowing what to iterate and what to polish requires intuition and taste. And thats the thing about taste, its forever fleeting. What we think is AI slop, was actually novel and coveted at a point before AI existed.
Yea, you can create code slop much quicker these days. That much is true.
Whether the demand will match the supply is still an open question. Did the apple app store need another 4,000 to do apps? I guess we'll see.
I use Ai to remove large amounts of code from our code base. Do refactorings that would not have made business sense before.
I also use Ai to be more ambitious. Online evaluation for our in app flows instead of offline.
So for us the entire quality of the product has been increased a lot.
You're either using it wrong, or work somewhere that sucks. I'm having a ton of fun coding with AI.
I feel like I am learning a whole new branch of the programming skill tree. LLMs and their harnesses are like a mew set of constraints to work around when designing systems, but if you get them right you can build bigger, better things than ever before.
I say all of this as someone who spent the last two days rebuilding RBAC for my application after Claude royally messed it up.
What's an example of "bigger and better" and what projects did you have before?
I attempted to build a K8s based image catalog and management system a few years ago, and progress was very slow. Took 2 months to get the first vaporware demo put together.
This year I'm building and IoT data management platform and I've already built two demos of the product and am adding a bunch of features for a third. Nothing is vaporware about it. All in about 4 months.
The big wins have been system config, debugging, and exploration. I was able to build an Arrow Flight SQL backend to use as an interface and try it out with some use cases, decide it wasn't going to work, and replace it with something else, all in about two weeks. Would have taken far longer before, if I would have been able to do it at all. I knew nothing about Flight SQL before trying it out.
I shit you not I have been told to "stop trying to understand it, just make sure it works" when mandated to use LLMs at work, like the absurdity of that statement isn't self-evident.
Needless to say I have been looking for a way out of this job (and likely career) ever since.
What's going to happen is people who give a shit are going to get filtered and self-select out of the job, and then the entire field is going to be dominated by dunning kruger AI maximalists. The "AI will replace engineers" is true, but for an entirely different reason and entirely different way than most people making that argument think.
You (and many others) have become a "reverse centaur" [0]. AI has been forced on you and reduced you to little more than a button pusher.
[0] https://pluralistic.net/2025/12/05/pop-that-bubble/#u-washin...
Please stop saying we - because many of us didn't ever desire or contribute (outside of having our code / data / IP stolen) to these models or the companies developing them. Many of us weren't willingly singing up to use AI - we were essentially forced to by our employers. Not everyone is more productive now because of AI - many people spend the extra time they save writing code, reviewing code produced by AI.
Yep. My job was code review heavy before AI. Now it's even worse because I can't even trust that the developer understood what they were "developing", so I have to be even more vigilant. I may have to institute a "prompt review" process for some.
if computers themselves are bicycles for the mind, LLMs are cars. or even motorcycles, if we talk smaller models.
- they require specific roads being paved for them. for example, if your tooling is proprietary and not accessible from the CLI, your agent is pretty much fucked. if your tool is not represented in training data (think, `jj` VCS or your proprietary/tailor-made tooling), you require duct tapes such as "skills" and "memories". a bicycle (that is, your own mind + computer) handles such off-roads much better.
- they get you from A to B faster, sure, but along the way you may encounter something curious - a different road to take, an interesting vista. not to mention, bicycles are actually good for your health, and professional drivers suffer from all the sedentary job diseases we programmers do, unless they actively counter it. with LLMs, we get a "sedentary job disease" of skill atrophy, on top of all the other atrophies us coders should counter with a proper exercise set at least three times a week.
- finally, when you crash in a car (Opus/Sonnet, GPT-5.5) or, worse, on a motorcycle (smaller Qwens, DeepSeek, Haiku/GPT-5.4-nano), you crash very loudly and with a high chance of irrecoverable casualties.
The bike/motorbike/car analogy is good. In each you can find enjoyment that suits your character.
I have been drawn to coding from the first day I took a course at college learning Pascal and assembly on an IBM PC. Hell I even wrote an IEEE/488 driver for an Osborne 1.
As an experimental physicist, coding was always central to my work. I always felt guilty because I had more pleasure coding than tackling the physics itself.
This is a new age we are entering. AI changes how we do things, but I believe that the human passion to be creative and do intellectual work will always be part of it. It's in our nature.
So I'm not worried that we'll all dry up and turn into zombies.
And, like cars, our federal/state governments are lighting astronomical amounts of cash to subsidize the infrastructure that makes them operate at the scale they currently operate at.
Just one more ~~lane~~ datacenter, bro. That'll fix it; trust. Please. One more ~~lane~~ datacenter. It'll be so good.
There's one more: The "bicycle for the mind" analogy was one Jobs chose specifically in reference to the fact that a human on a bicycle moves more efficiently than a human on foot—or any other land animal for that matter. But it's still the human's own effort that propels the bike. Motor vehicles require an external energy source, the human only directs it. Similarly, with LLMs, it's no longer the user's own effort that's doing the thinking.
> There's one more: The "bicycle for the mind" analogy was one Jobs chose specifically in reference to the fact that a human on a bicycle moves more efficiently than a human on foot—or any other land animal for that matter.
One of the canonical references of Jobs’ use of “bicycle for the mind” also compared the efficiency of locomotion to the California Condor which was (I’m working from memory, here) 17X more efficient than a human walking.
A human on a bicycle, according to Jobs then, is more efficient than the most efficient animal locomotion known to humankind. The comparison included animals moving through non-terrestrial environments.
I wonder if this same effect happens for very wealthy and/or very senior executives? Those sorts of people have always had numerous people they could 'outsource' their thinking to; delegating work, asking for research/summaries, assigning tasks.
Does handing off that sort of work to people also ruin your skills in the same way? Or are AIs fundamentally different, and if so, why? Because we have no moral or social pressure to not delegate everything?
Yes. The moment you move from daily IC work to management your skills begin to atrophy. Even if you are magical and somehow they do not - you are not keeping up/practicing on the newer developments.
If you are doing your job as a manager it’s the scariest career possible. There is no realistic fallback to lower paid IC work after a certain amount of years. Your job is to enable others, not do things yourself.
There are shades of grey and you try to keep skills at least honed a little here and there doing R&D and side projects - but it’s just not the same as day to day production line work.
And of course some people start at a much higher baseline of skill than others. The impact is largely the same though over time.
This is by far the single largest caution I give to skilled engineers who talk to me about moving into a management track. It’s a decision not to be taken lightly.
Yes. It even happens to e.g. professors and the deans of universities.
This is why I think even the fuss about Noam Shazeer joining OpenAI needs to be seen within a context; as good as this hire is, there is no inherent reason to believe he still brings some secret undiscovered magic that others do not have in a more current form.
This is why senior swes are known to provide so much more value than principals.
I find this hard to parse, maybe because English is not my first language. Do senior SWEs provide more value than they provide principles, or do they provide more value than principles do?
He means "Principal engineers", its a title for higher up engineers.
fixed, my bad.
intellectual work (or really any discipline with a 'skills tree' that you can progress up or down) degrades over time without practice. same way that, if you don't run for a while, you're not going to be able to hit your last PR when you pick it back up.
it's definitely easier to catch up after some time away than it would be if you'd never developed the skills in the first place or didn't have a natural talent, but you'll definitely atrophy without exercise. every leader i've ever worked for who graduated to a purely managerial/'strategic' position and didn't keep up their IC skills eventually got pretty slow on the uptake.
i appreciate that this study was done (AI and its inverse relationship to human wellbeing is one of the biggest challenges of our time IMO) but this also seems obvious
I've discussed this in my company, where I've argued that we are losing critical thinking skills for AI.
My friend said being an engineer is not just writing code, but designing and architecting the system. I agree, but we are also moving that to AI. We are offloading the thought process to a LLM, which confirms our biases and tells us what we want to hear.
Essentially, we are also losing the software architecture and design skills because we aren't talking to other engineers/architects/designers who may approach the problen in a complete different way; we are just asking the AI to confirm our thoughts.
The ownership class does not really do work. What they delegate (high level plans) is generally easier to do for AI agents than things like software engineering because it does not need to be as precise or executable. However, there the work has very broad scope scope and the roles are very high risk, so it might need a Fable 5 or Fable 6 level vision language model to remove the jaggedness to make it 'safe' enough to drop the humans But what is going to blow up in 2027 are Automated AI Companies. Human CEOs and owners will not be able to compete with these AI-run companies.
It will be interesting to see how AI and company leadership/ownership stuff shakes out. In some sense I would expect it to be quite different than what exists currently, because an AI can’t be responsible for anything legally, and (arguably?) can’t own anything. But then, C-levels don’t actually seem to have much personal legal liability toward their company’s decisions, or maybe it won’t be so big of a deal.
I’d be curious to see alternatives ownerships structures. Like an AI-coordinated collection of guilds, unions, or co-ops. If it can’t accumulate upwards, maybe the fundamental unit of ownership will stay with the workers.
Yeah, the idea that C-suites ever took any significantly liability has always been a fantasy for as long as I can remember. At worst/best case a single lower level individual will be used as a scapegoat for the systemic abuses of the company.
AI needs a great deal of handholding which is different than just offloading tasks. You spend a lot of cognitive energy playing a slot machine hoping the RNG works out this time.
There’s definitely a skill to using AI but it just doesn’t generalize very well.
> playing a slot machine hoping
I still believe the slot machine analogy holds to some extent, but I can honestly say my winning percentage is at least 90% for one shot generated code now.
I think if you know it's limitations (inlcluding your own), I don't think about hoping anymore.
I should note that when I say AI, I mean the collective models from all the major providers. The most important lesson is, you need to ask around.
> There’s definitely a skill to using AI but it just doesn’t generalize very well.
This I agree with. The only way working with AI can really be benefical outside of dealing with AI is, we are visited by extremely intelligent beings that will fuck up in the weirdest ways.
I wish I could've one shoted this 200 LoC PR: https://github.com/ncruces/wasm2go/pull/42/changes#diff-5181...
I'd really like to know the prompt, because I did try and, this took me way too long to figure out.
The AI agreed
> There isn’t a magic prompt I could have written before knowing the solution. A retrospective prompt can reproduce the 200 lines only because it embeds the invariants you discovered. The useful AI task would have been systematically finding counterexamples and iterating toward those invariants, not generating the final file in one shot.
For me the value with AI isn't the one shot, it is the fact that it can help you iterate faster. My biggest take awway so far with AI is it can help you fail fast. The faster you can fail, the faster you can iterate.
When it works the first time you’re done, when it fails you spend more time on it. This biases the cognitive investment around failure.
> I don't think about hoping anymore.
Running the exact same prompt again that already failed doesn’t have a that high success rate, but it’s also very low effort. So IMO it’s often worth attempting.
AI has definitely gotten good enough that its one shot success rate is better than most developers.
whats your fave/best one shot code gen?
mine def has to be the initial impl of my python to my personal fave compiler ir. i got claude chat to write it in two sentences. 4 turns helping it rmeemver where it put shit becsuse of transcript issues
The code in http://github.com/gitsense/gsc-cli contains over 300 go files and I would say 90% were one shots. I can even prove that they are AI generated since I attribute the AI in each file, as shown in the example below:
https://github.com/gitsense/gsc-cli/blob/main/internal/cli/r...
Since I don't know go, I can't say which is the best. I am also sure they would not live up to a go experts expectations. I just know that since I don't know go, I would not have been able to create the tool. I can also prove that you can easily maintain AI generated code since I have been expanding and evolving it for over 6 months with pretty much no issues.
Yes, but what happens to the professions and classes you've mentioned that wasn't looked at in this study is that new skills developed. For instance, something was assuredly lost as programming on average deviated from low level programming, but new skills and optimizations were gained. Fewer people knew how to construct a real-time bitmap from scratch, but more advanced software was being built on average as people were armed with greater and more manageable abstractions.
Having run into quite a few managers who used to be programmers who can barely do fizzbuzz after a few years of managing: yes, I think it's pretty similar. Much of your brain is "use it or lose it" and that's not using it.
I do think there's more than enough room to claim that LLMs are probably significantly worse than that kind of human delegation though, in part because you have such a rapid cycle time that not-incredibly-rich people can't afford from humans.
The difference is what you delegate to. An executive delegates to people like himself. A computer user delegates to another kind of existence so those think don't think the same way the user does.
Programming normally highlights this difference. LLM programming makes it much less apparent but its still there, LLM are not thinking the way humans do and therefore struggle to solve many problems humans easily solve. So letting all human programming skills rot and just use LLM will halt our progress unless we reach AGI before our programmings skills are mostly gone.
The "I have people skills; I am good at dealing with people!" character in Office Space was memorably hilarious for a reason. Many people lose the skills that got them where they are from disuse as they rise through the ranks.
>Because we have no moral or social pressure to not delegate everything?
"We" depends on where you are. Countries like Japan or Germany have maintained a gray collar, rather than a blue/white collar culture for exactly that reason. You will find business owners on the workshop floors frequently because there is an understanding that divorcing management from tacit work is going destroy leadership ability. That's the basis of vocational work culture, having general expertise across all domains of your job rather than being a kind of over-specialized idiot.
This is why the rich and powerful often seem so out of touch. There's no one around them willing to tell them they're wrong or push back on bad ideas. It doesn't just ruin skills but fundamentally distorts one's perception of the world.
The main skill you need to reach that level is the one they always keep practicing: shmoozing. Thinking is not required as an executive, so you cannot lose what you never had.
And being born wealthy requires zero skill or practice.
This seems to assume that a skill is only important if you need it for your job.
If that is the case, then wouldn't this whole thing be a non-issue? We lose all the skills we used to have, but we don't need them because our entire job now is interacting with AI, and that skill we will continue to develop because it is what we do all day.
I don't know if I fully agree with that. Some skills are important even if you don't need them for your day job.
> Or are AIs fundamentally different, and if so, why?
Literally: the context window.
With the human you have a window that possibly extends up to _years_. With your language model you have maybe a few megabytes which is always preceded by instructions from the model maker.
That's a misconception. Context window of humans is measured in seconds and limited to single-digit amount of entities. Human attention is truly ephemeral, with a ridiculously short span. What you're talking about cannot be emulated by putting everything into the context, you need models with deeper cognitive capabilities, and the longer your context the smarter it needs to be. It can be done by increasing the depth and the size of the model, and by training it better. Test-time scaling does work (longer reasoning, agentic compression, ICL, etc), but hits the wall rather quickly if not assisted by better cognition.
> Human attention is truly ephemeral, with a ridiculously short span.
I do not believe this at all. I think you'd have to have a very limited experience working with other human beings to be able to believe this.
> and by training it better.
"Oh yea, just do it _better_. That's your problem." Perhaps some people operate without any context but most of us find the experience lacking.
That's a simple neurologic fact, there's nothing to believe. You can use a trivial experiment to verify that you can't keep details in your sliding attention window for more than a few seconds or focus on more than a few things simultaneously. Human memory and cognition is layered and this immediate layer is what resembles the model's context the most.
You're possibly mixing it up with long-term memory which doesn't keep immediate facts and details, it's for heavily processed and compressed summaries, for the lack of a better analogy from the LLM world. You aren't keeping the entire codebase in your memory, just its highly processed and conceptualized version. This conceptualization can be somewhat emulated as an agentic loop, but it can only go so far, current models quickly lose coherency and aren't good enough to predict what's important.
Models don't need to remember more details, they need stronger processing of what they already remember.
>"Oh yea, just do it _better_. That's your problem."
I think we're talking about different things. Models can be trained better to cram more intelligence into the same amount of parameters, that's what I mean. Similarly to how your ability to learn (and perform) math depends on your prior math training.
Is this true, though? AIs already interact with way more data than their context window, they can make queries to external data sources, and they can make context references to external data that can be pulled in as needed.
The context window is the working memory, which is not ‘years’ for humans either.
Okay. Can you take _all_ the data you have out of your mind and put into a computerized format? Is there nothing intangible in your mind?
The human context window includes actual _context_ not just _data_.
I haven't written a full function of code in over a year. That being said, I've been spending a lot more time thinking about architecture and system properties.
So, yes, I do feel like I've lost some of that very low-level skill. But maybe I've also been able to spend more time on a higher level skill? Maybe the doctors got worse with the images but had more cognitive resources to think about the patient's context?
Not sure.
But yes, I can't physically get myself to write code without an AI anymore. It feels so much slower, almost painful.
I’m curious if someone coming into that fresh, without having had the mental exercise of, for example, grappling with data structures and (lower-level) algorithms in practical applications by hand, could achieve a mindset that useful and productive?
When I was in design school, as much of our work was in physical media — graphite, cut paper, paint, vine charcoal — as was practicing great kerning and getting experience with the digital tools. Even though you still had to make the individual strokes and choose appropriate tools in the digital realm, there was still a perception of the process that was obviously lacking among those that came from strictly digital backgrounds. It’s similar to seeing someone who’s only worked with photo references try life drawing — there’s an entire part of the cognitive process not being used when you’re drawing something that’s already 2D. Sure, they can learn, but unless they’re forced to, they’ll probably just keep taking a picture and drawing that. But image generating, even with extremely granular inpainting and such, is so different it’s not even comparable. I’d hesitate to say that someone with a lot of experience doing very advanced image generation would be dramatically further along than a complete beginner if they learned to draw which is not true of the photo reference artist, and even less true of a purely digital artist that did life drawing on a tablet.
Kind of like how millennials, many of whom always had access to technology, but also experienced dial-up-era computer use, are generally more technically savvy than the your stereotypical “iPad kid” that can’t even traverse a directory structure.
Yes, and I struggle to give junior developers feedback because of this.
The code they submit for review no longer represents their thinking or their skill. The feedback loop for beginners is a bit broken.
What is their skill now? How is it displayed? How and where do we provide valuable feedback? I find myself just approving large PRs because I don't have the time to read it all.
I feel like I've let down a lot of developers this year.
> I feel like I've let down a lot of developers this year.
I think the industry has, and we don’t get much out of individualizing it. Let’s not forget that this progression is a series of deliberate choices made by a handful of very influential people that couldn’t give a teaspoon of shit about developers beyond what their subscriptions bring in.
It’s not a super well-received opinion around here, but I think the economics of the large-scale services will mean on-device models are going to be where this all goes sooner or later. My gut says setting up a local LLM coding assistant will be more like choosing and setting up an ide. I don’t have a crystal ball though.
I was thinking about this (while drawing), and there's a few points to make:
* There's meaning in all the little marks. Drawings are potentially layered with meaning. I was drawing a stick I found in the park, and thinking: this is the extent and the way in which the stick is straight or curved, and the specific way and form that it's knobbly, and that's all explained by how it grew; this green on it is the peculiar blue-green of lichen; it's a oak stick, I like oak; my drawing is just a stick, good, I might frame it, call it anti-art maybe. Some or all of that is expressed in the marks, though this becomes more apparent if I tell you that it is.
* So, an AI - diffusion - could undoubtedly draw a stick in an arty way. But to draw one that fits the above meanings, it has to be prompted with the meaning. The AI can't prompt itself, so there's a role for artist-as-prompt-writer. Whoopee, right, what fun. These prompts, if about graphical matters and a sort of meditation on the stick, would be essentially lies, because no such exploration of the form and significance of the stick would have really taken place without trying to draw it. It can't prompt itself or be honest.
* So to come up with the above words that might do as a prompt, I had to draw a stick. I had to learn by drawing a stick, even though I've been drawing for decades. Experience isn't the whole deal, otherwise the artist is only churning out filler material, which would be like AI art. Instead the artist has to explore all the time, while leaning on experience. The viewers are into that, they sense the excitement of exploration. They want to see an artist. Well, not exactly, most artists are unrewarding to look at (Brian Froud for instance), but they want to see creativity unfolding, over several pictures.
You can have non-graphical (non-mark-making) prompt-art, a bit like collage or photography, sure, but that's its own thing. Like you're saying, you can't just fake the craft forever, and even faking it once is less than ideal, unless you're narrowly focussed on output that meets targets, instead of meaning.
With programming, this might be different, since hitting targets and getting functionality might be all that's wanted, but I'm sure it depends.
The way I see it is with every stroke, an artist makes multiple decisions, most of them unconscious. They’re based on everything from the media, to our experience, mood, inspiration, and very much on our physiology… the arc of my shoulder and the length of forearm are just as much a part of the process as what inspired me about something to draw. To some extent, much of that is even true in photography.
Diffusion models basically record, classify, and amalgamate those decisions, which is why it’s so dang difficult to get generated art to look like something distinct. Not distinct like an existing artist, but genuinely unique.
The workflow of the prompter is very similar to an existing workflow in the art world: someone commissioning art. There’s often a discussion where the customer gives the artist a textual description, back-and-forth with sketches and preliminary versions, and sometimes revisions if the proposed final product isn’t what they wanted. Commissioning a piece of art is a creative process, but it’s not the same thing as being the artist. Even an art director who has extremely granular control over what they commission would never claim to be the artist. They’d get run out of town. I believe creating a collection though juxtaposition or even curation could be art, but you’d still never be the author of the contained pieces.
The question is what higher level skills there will be to focus on?
I am not convinced that there are tasks, like project management or architecture, that the Ai is inherently worse at.
Ai is decent at it but it still struggles to find and balance and choose the right context. This will of course improve but I’m starting to think it’ll be like the self-driving car history where the last 10% needed is a long slow road.
10-20 years ago we had such high hopes for data and IoT, but it’s still just banal and mildly useful, full of walled gardens and property formats. General purpose AI may end up similar. My company has like 10 different MCPs to enterprise software we use that has context, but AI still needs lots of guidance.
I think the main aspect is that a human can be held responsible. A human can be fired. Then we are pretending agents have agency when that is also what the human has. It seems like the meaning of economy of scale should almost flip to mean the economy of very small, highly specialized companies. It seems hard to see that the person in charge of cybersecurity is going to also be in charge of marketing. I think there is also this aspect that if you told me in high school I was someday going to be a project manager, I would have thought you were insane. You mean like Oppenheimer? No one is ever going to put me in charge of the Manhattan Project.
I think you can also see the vague outlines that someone is going to have to be in charge of token cost.
Managing AI context.
>But maybe I've also been able to spend more time on a higher level skill? Maybe the doctors got worse with the images but had more cognitive resources to think about the patient's context?
That's not the way the economics behind this work.
Supposing the AI priests are right (they aren't) and using AI creates a thought surplus on the user, freeing cognitive capacity to think of higher things. What do you think will said user's boss want to do with that surplus? Let the user develop higher-level cognitive abilities? I don't think so.
The doctors in the article performed worse post-AI: suppose AI saved them so much time that they did 100 exams in the time they used to take doing 10 exams. What will their employers do with that freed up labour time? They'll of course have the doctors do more exams and perhaps fire some now-redundant doctors that are no longer needed. The surviving doctors are left deskilled, doing the same or more work, and society gets worse quality medical care. But hey, its not all bad - the employer gets to save on labour, and shareholders will be happy.
We never learned how to use IBM punch cards. Writing functions by hand for modern languages is fast becoming anachronistic.
Absolutely not, not until LLMs stop being unreliable bullshit generators, and I believe that the current generation of AI is simply not capable of it. Even the best available (lol) model, Fable/Mythos, is still a bullshit generator, just one that's a bit more powerful.
In the eyes of many ceo’s, you are slowly becoming useless.
You’re fucked once you have to start getting AI to produce code for languages you’ve never actually worked in.
You will have to write code to understand deeply what goes on at the low level. Without a solid understanding of the low level, you won’t truly know what optimal solutions look like on a high level. You will be flying by the seat of your pants, churning out code that works but has bad low level quirks sprinkled through the code base. The AI will say it’s fine, but you’re just building up shitty software. Feels like shit, run likes shit.
Once someone new comes along who has worked with the language manually and can get AI to produce effective high quality code, you’ve lost competitive advantage. They can build way better versions of whatever you do. You’re finished. You’re a low quality engineer.
Your argument has many, many leaps but I infer your point.
The lack of attention does reduce the overall correctness of the system.
But, and unfortunately, in most situations the loss of correctness is more than made up for by the speed gains.
Unless you're building software that could kill someone, it's hard to reject the improved speed. The cost of delay is usually more damaging than the cost of incorrectness.
>But, and unfortunately, in most situations the loss of correctness is more than made up for by the speed gains.
Welcome to the future, where we can create buggy, broken software at a higher velocity than ever before!
In real business, all important software is capable of killing someone or cause monetary loss through incorrectness or loss of functionality due to performance failures.
If that doesn’t apply, you are just writing toy apps or front ends that are of no real significance, so yes you could carry on with incorrect code in those cases, but it’s not much of a career either.
Thank you for illuminating the realities of real business.
It made me realize I've only had the misfortune of being employed at fake ones.
This is new, the scope of it, its not just about individual "skills" because its all of them; we are being challenged at the very fundamentals of our ability to think deeply and widely and persistently. That has never happened before like this.
It is quite extraordinary and breath-taking at times to see the agents in action; the flipside is that very power renders us both vulnerable to its seduction and enfeeblement on an equal scope - its almost hard-drug like in its potential long-term psychological effects.
> we are being challenged at the very fundamentals of our ability to think deeply and widely and persistently. That has never happened before like this.
Social media and content algorithms come to mind as an early wave that changed the landscape here that defines the horrible status quo leading into the AI era.
These days it's trivial to slide into an echo chamber and very hard to break out of the silo.
There might be a double-edged sword here where AI, trusted by most people as an omniscient oracle, can offer the only pushback we encounter on positions we picked up passively by scrolling social media, Youtube, TikTok.
For example, ask Claude, ChatGPT, and even Grok about the "space lasers" that started wildfires in Hawaii in 2018, something people like Marjorie Taylor Greene floated on social media. It quickly debunks it as bullshit.
Now, maybe it will pan out such that everyone will have their own AI that tells them what they want to hear. But so far I've watched people abandon arguments on Twitter because Grok rejected their claim. So it feels like there's a glimmer of hope.
> That has never happened before like this.
You should possibly spend some time reading what people used to say about the invention of Radio and Television.
> It is quite extraordinary and breath-taking at times to see the agents in action;
So is any magic trick. The unsettling notion that it may all just be an illusion that you've failed to correctly understand doesn't seem to weigh on people.
> its almost hard-drug like in its potential long-term psychological effects.
That might have more to do with how the owners of these products choose to market and deploy them. Perhaps if they peeled back the covers just slightly your euphoria would change to dread. There's an Upton Sinclair moment coming.
Relatedly, this is Plato on writing:
"They will cease to exercise memory because they rely on that which is written, calling things to remembrance no longer from within themselves, but by means of external marks"
And he was right, that is why we don't let people bring their textbooks to most tests. That rule wasn't made because we hate students, its because it improves the outcomes if we force people to commit things to memory instead of just looking them up.
You would need similar restrictions during the learning phase to get people to learn things in an LLM world.
So writing and LLM are still a net positive, but only if you acknowledge the problem and add fixes so people still learn rather than just rely on this crutch.
Human mind as much as body needs challenge. That's the only way for growth, heck even sustaining some higher cognitive levels.
Nurses, doctors and family members know damn well how life trajectory nosedives for somebody ie suddenly bound to bed, when stimulus and doable challenges are reduced to minimum.
llms remove challenges, or minimize them. I can't image any added value for any engineer apart from cost cutting for employer. Sure, next come folks who are doing 10x compared to before, and some actually do. Even there, I have my doubts. For rest of us, its not good and won't get better unless they price it out of most markets.
Im learning new things at a pace I never imagined at 40 years old. New sports, new businesses, new academic pursuits. Technology is a lever and AI is the biggest lever we've ever had. It enables laziness or incredible productivity. Choose your own path forward.
I don't know this applies to you, but I've had several friends who convinced themselves they're exploring the frontiers of science, and it always turns out that their conversations spiraled into some sort of weird quantum-metaphysical gobbledygook.
LLMs are sycophants, and in long conversations, their sycophancy produces a positive feedback loop: the context window contains affirmations of incorrect interpretations / analogies, so the chatbot continues down that path because, well, that's the most likely completion of previous text. And before you know it, you're discovering the hidden fabric of the universe, which is always some Minkowski fractal spacetime tensor lattice manifold with subharmonic DNA nanotubes.
That is to say, unless you have a robust way to evaluate what you're learning, and to confirm that you're actually learning, I'd tread carefully.
You really just need to augment with tight prompting and know how to extract information and links to peer reviewed literature and well sourced information. Once again, technology here is a lever. Separating wheat from chaff has been key in academic and information pursuits forever and it is becoming ever more important.
And you're gonna review that voluminous academic literature in a field you're not familiar with, right?...
That is my point: an LLM can be great if you know the field and can spot errors. Or, to a lesser extent, if you have some automatic feedback loop that the model can't easily game ("does this code pass unit tests?"). It's a lot less great if there's a risk that you won't detect the early drift.
Sifting through information to separate truth from fiction is a modern, experience-based skill that I’ve developed YES.
However, your perspective on how LLMs are used might be too narrow. While no one is suggesting using an AI to find a one-shot cure for Alzheimer's, LLMs are incredibly effective tools when paired with textbooks to master subjects like undergraduate physics.
100% you are so right. Don't rely on the LLM. It is a robot you give tools and context, and then you ask it to do something with that. It's amazing at that.
>LLMs are incredibly effective tools when paired with textbooks to master subjects like undergraduate physics
I agree that they can be helpful for explaining things if you're stuck. But since you used the example of physics, the majority of hard won physics knowledge comes from working lots of problems yourself.
Too many people are already used to LLMs circumventing the laborious aspects of things, so it takes a certain person to withhold from just having the LLM solve the problem for you when you're stuck.
Until we can figure out to teach people to fight against their instincts, then I don't think LLMs will lead to better long term outcomes.
If you’re cognizant of the sycophancy and not a total narcissist then it’s not too hard to figure out with a little experience when the models are just trying to hype you up
> Im learning new things at a pace I never imagined at 40 years old.
You aren't learning anything. Learning involves doing.
We've known this for ages: simply reading a maths book without drilling on the problems will not get a student to pass.
Best case scenario, you're reading stuff. For users of coding agents, they're not even doing that.
Learning by doing is often more effective because it naturally involves forced recall.
But it's not the only way to learn and not even the only way to do a lot of forced recall.
It's downright crazy to say you aren't learning anything by reading. You likely won't retain that high a % of the content without repeated drilling, but it's not like nonfiction exists for no reason!
>It's downright crazy to say you aren't learning anything by reading. You likely won't retain that high a % of the content without repeated drilling, but it's not like nonfiction exists for no reason!
How much of a nonfiction book can you recall a few months after reading it? Probably very little.
However, I haven't taken a calculus class in nearly a decade and I still know how to solve derivatives and integrals.
Point being: there's levels to this, and reading is not nearly as effective as drilling exercises.
>You aren't learning anything. Learning involves doing.
You can have it write a program that generates drills for you.
I wanted to become better at reading sheet music so I generated a sheet music reading program. You can have it generate maths drills, then ask questions about it if you get stuck or whatever. If you genuinely want to get better at something then AI will help you learn it faster. Obviously its going to hamper more people's cognitive ability that it will enhance but that is a separate problem.
The people who say they are learning faster with it aren't mentioning drills, they're mentioning outcomes.
I actually did have an LLM ingest some material and generate drills. It worked well. It's rare that happens, though.
The difference between humans and other animals on the planet had always been the ability to reason. If we, as a species, lose that ability, we're looking at an extinction-level event.
Your initial response post was to me, and you absolutely injected the heavily biased assumption that I was doing no active learning.
I'm learning by doing things.
For example: AI has helped me get into restoring retro tech, specifically resoldering leaky caps on retro Macintosh logic boards. Before AI, I didn't know how to use a multimeter (I knew theoretically how it worked), I didn't know how to use flux, solder wick, heat gun. I also didn't understand how bromine radicals yellowed plastic and how to reverse it by using blue light similar to what they use for indoor aquariums.
So AI unlocked doing for me.
You realise that every single one of those things had easily accessible videos to learn from, right?
You're happy using AI instead of other material because it will constantly tell you how brilliant you are, or how quick you're learning.
No, not everyone, there's no easily accessible video showing how to setup aquarium lights to slowly de-yellow ABS plastic -- at least I didn't find one; AI helped me find an obscure reference to it and I then read a few message boards where it was mentioned.
I also use AI to take in-progress pictures as I desolder to help me check for traces that need to be repaired or help identifying specific chips. I probably could try and find a video where the same chip is featured and someone explains it, and/or retrieve the schematic for the specific logic board, but that's very painful and does slow the process. Think of AI, in this specific case, as enabling skill development for me in a field I wouldn't have necessarily have gotten into, because of being short on time and AI helps me consolidate that information quickly.
You are very closed minded. LLMs can absolutely give pragmatic and fast information in a useful (non-sycophantic form). You need to examine your extreme biases here.
Just because you can't use a tool doesn't mean the tool isnt useful.
Your bias on display here is frankly silly. Im not saying LLMs are ALWAYS the best way of learning something just like they aren't always the best at anything. They are a valuable tool though. Yes so is youtube and textbooks, and professsors, and peer review literature, and pen and paper, and block training etc.
Should we also go to ancient scrolls for learning, because videos are too easy? The poster showed you how they learn with the help of AI. You can learn from videos, they learn from AI. What's the problem?
When did I say I wasn't doing and just reading. That is your own heavy bias creeping in.
This.
Without doing you may as well read some fiction. The result is mostly the same.
Nonsense. Absolute nonsense. Ive crafted LLMs to create drills and practice exams etc. Im not just reading about new sport activities, or videography/photography or linear algebra or physics. I'm putting it into action. As I said above, technology is a lever. Maybe you are just reading stuff and not drilling problems. That isnt me.
Not sure how good LLM are at that given they are so biased towards the user being right. You would need an LLM trained to disagree with the user to properly grade your results of those drills.
Once more, you are making massively blanketed biased assumptions about me.
You seem to believe LLMs cannot even produce simple exams and rubrics is that right?
At most you could do multiple choice questions, or questions with a single objective answer such as "what is the capital of France". But such tests aren't very helpful to learn new things.
Students cannot grade their own answers for more complex problems, they make mistakes but say they are correct since they don't understand the material, or they are correct but since it doesn't say the exact same thing as the example answer they say they are wrong and correct it even though it was already correct. And an LLM would be even worse than that at correcting tests.
This is not a bias against you, its just a general thing that applies to all students and people. Nobody is good at correcting their own work, even the most esteemed professor gets his work checked by other people. And an LLM is not another person here, they aren't good enough to check your work.
Note its much harder to accurately grade an answer than to answer a question.
You are so incorrect here. If you feed it an evaluation rubric it will go after it! You really understimate the technology when in good and well intentioned hands. (Yes lazy people will use it to cheat and get to the answer faster. But it is like a grad student tutor who will do an evaluation in 30 seconds! Rapid iteration and laddering up...)
You're not learning them. You're being told about them and given a hammer to leverage them with mediocre to low skill level.
Learning requires a huge time investment. Using an LLM doesn't shorten that.
You've described school. You get told about things and given a hammer to leverage what you've been told.
An LLM absolutely shortens the research part of learning. If I had a human of who had a moderate level of skill who would endlessly answer all my questions, the result would be the same.
You might have a point when it comes to software development because the AI can tell you things but it also just do them for you, at which point, you've learned a lot less. But for non-software things I have to learn things so I can then go and do them.
But even for software development, I've learned a lot of esoteric crap to get interop working on projects that I will probably quickly forget just the same as when I had to spend hours skimming through stackoverflow.
> An LLM absolutely shortens the research part of learning
No, it doesn't. Because in any scenario where you are using AI in a potentially appropriate manner, you are verifying every single source it spits out and cross referencing everything it says. If you do not do this you are failing the process entirely.
If I ask my buddy for help, I don't cross reference everything he says either even though he could just as easily be wrong. A lot of what I'm talking about is every day stuff not nuclear physics.
LLMs are not even close to that unreliable.
If I ask my buddy for help I'm not considering that part of research or learning I consider that leaning on their potential expertise. You're not learning shit on how to build a house just because you ask your friend to build one for you.
If we're talking about research as in actually attempting to learn and get to the heart of a matter then yes. You should be cross referencing everything.
My point is if you ask your friend/LLM how to build a house and then build a house then you've learned how to build a house.
I've build a lot of "houses" recently that I wouldn't have attempted if I didn't have a friend/LLM to ask.
You probably just built a shanty town according to a professional.
What's your point? I could just not ask an LLM and not do anything. Is that better somehow?
Yes. You could not do it. Doing stuff isn’t essential. Especially when you have no idea what you are doing.
So never learn anything? Be entirely beholden to other people for everything? People who somehow themselves also have complete knowledge? I've done so many things myself now and that has given me the confidence to do even more.
If I had to have a complete idea of what I was doing to do it, I wouldn't even have a job. Doing stuff and making mistakes is exactly how you learn.
I think you might have reached a nonsensical level of AI contrarianism when it leads you to the statement "doing stuff isn't essential"
Spot on analogy.
And I see people say this, that they've built a house but what they've built is something that structurally collapses immediately and fails due to problems that even an average person would've avoided had they learned how to build a house.
So did you build a house, or did you build a house?
How can I argue about such a hypothetical?!? If I ask an LLM or my mother-in-law about how to get wine out of the carpet, and I get wine out of the carpet, maybe I have done it wrong and my house will burn down! Who knows?! Your argument is meaningless; there's nothing anyone can say to that.
And if you ask an LLM about an effective cure for cancer and it spits out to drink bleach, are you going to follow it?
When you ask 'who knows' that's the point of research which was my original comment here. The same goes true for some random asshole telling you to drink bleach as it does an LLM, except people seemingly have hyped themselves into believing the LLM is more right somehow instead of being trained on every random asshole ever.
Time helps to some degree, but I've worked with professionals who've earned a living for 30+ years, enjoy strong reputations among their peers, and who absolutely do not know what they're doing.
You think you're learning things at a pace you never imagined, but all you're really doing is learning superficially. You never really go deep into things, and it "seems" like you're learning faster, but how much can you actually recall of the things you learned? By the way, I'm saying all this because this has been my experience. I can prompt like no tomorrow asking questions, but I forget most of it because it's not deep learning. I'm not actually thinking myself, I'm allowing the AI to think for me.
How exactly is AI helping you with sports? Have you become a better athlete? Have you learned more about the strategy of each sport from the perspective of a coach for example? Care to elaborate?
recently vibe-coded a personalized skateboarding block-training app. The platform maps curated YouTube tutorials to specific tricks I want to learn, logs my daily progress, and tracks failure rates alongside the specific reasons for missed attempts.It also functions as a structured video journal that integrates seamlessly into my coaching routine. For instance, during a session of 10 daily kickflip attempts, the app logs my success metrics and video capture so my coach can review my exact mechanics. Over time, this data-driven feedback loop visualizes my progression curve and systematically improves my landing consistency.
In another case I follow a bodybuilding cutting regimen and it helps me create and track recipes consistent with my diet plan and macro guidelines. It helps me also create tasty recipes that fit my criteria based on the ingredients I have on hand.
Those are just 2 examples.
I also recently built a backyard jib setup with a platform, ramp, PVC jib rail for snowboarding, and it helped me architect the design for it.
I actually agree with you; I definitely feel I'm able to use LLMs to learn and explore concepts, but I've always been a self-taught and highly motivated person in the first place. Everything technical that I know, I know because I put in the work to learn it on my own. That's why I wish these tools were being advertised as models that help you do BETTER work, not MORE work. They're using them as excuses to lay off swaths of workers, instead of allowing them to uplift people's skillsets. And, of course, it invited a whole mass of individuals who use them to artificially elevate their perceived skills.
Are you actually learning them or are you just letting the AI kinda do them for you? There was a time I knew how to take a path integral, but kinda-sorta knowing what something is and how to ask a calculator for it is different than knowing yourself.
I literally just started playing around with Lathe (https://github.com/devenjarvis/lathe) which was shared here on HN a few weeks back. There are some things I have been wanting to learn more about (some for work, some for personal desire), and I used it to generate a tutorial about some of the things.
I have been very pleased with the results so far. I was able to tune the tutorial to exactly what I want to learn, and it did a very good job (at least that i have seen so far). It has made learning fun, since I get to learn exactly what I want, and I can ask the AI questions and have it make changes to the tutorial in real time as i am working through it.
Now, will I keep using this at a rate to fully offset all of the thinking i have stopped doing since i started using AI? I am not sure, I guess time will tell.
Yea, this is a serious problem but honestly it was present before LLMs. Actually knowing, deeply, how to do something was always very difficult and very time consuming. People would substitute it with “edutainment”: youtube videos, Ted Talks, blogs, etc. The thing is nobody was really learning deeply. We were training our brains for recognition, but mastery requires training for generation which was always the harder of the 2 cognitive processes.
"[Learning] new sports" strikes me as an especially odd one. I can see how an AI tool could help to learn the theory or perhaps come up with better training or match-day/racing strategies, but it won't short-cut the work of developing the necessary physical skills, will it?
See my response above on this.
I'm using it to learn coding.
i could not get through the hurdles of installing an IDE and js/python modules before.
now i am learning basic scripting and data modeling etc.
it is phenomenal for learning languages.
i built a chicken coop and some furniture. the skills and confidence i gained are real. am i failing to learn certain skills in the process? of course. but I'm getting further then i would on my own, and that is truly meaningful.
you can keep dismissing it; but I'm genuinely using it to break down barriers, give me confidence, and highlight my ignorance in very productive ways.
i find it bizarre how unwilling some people are to recognize that.
> i could not get through the hurdles of installing an IDE and js/python modules before
You want us to believe you couldn't overcome the puddle-deep challenge of installing an IDE and using Pip or Node in the past, but now you're actually learning how to write functions?
Cool for you if true I guess but I'm pretty seriously skeptical
Python environment management is notoriously reticulate. I can totally imagine that they got stuck on some path bullshit so that the installed packages were on a different Python than the one they launched in the IDE. I've been using and configuring computers since they had DIP switches and if I were teaching someone to code manually I would absolutely let them use Claude to get their environment set up.
I can sort of buy this, honestly. When was the last time you picked up a new language with unfamiliar tooling? Especially something older that won't hold your hand.
A lot of the most miserable parts of getting started coding have nothing to do with programming and everything to do with like, trying to apt-install the right compiler version or figure out which build headers you need or some other equally trivial bullshit that gets in the way of writing code.
> I can sort of buy this, honestly. When was the last time you picked up a new language with unfamiliar tooling?
I started learning Monogame with C# literally yesterday after being a chiefly JavaScript dev for over a decade. I'm having a lot of fun learning it from scratch with no AI
I'll concede that before yesterday it had been a while though
> Especially something older that won't hold your hand
Python and JavaScript are two of the most "will hold your hand" languages these days aren't they?
> A lot of the most miserable parts of getting started coding have nothing to do with programming and everything to do with like, trying to apt-install the right compiler version or figure out which build headers you need or some other equally trivial bullshit that gets in the way of writing code.
I agree but absolutely not for these two languages in particular
You can open the developer tools in any major browser and start typing in JavaScript right there if you want
Python you can install it and write it right in any editor
The upfront burden for both languages is absolutely trivial imo
It's kinda easy to measure if you've learned something. You can test yourself if you can do what you've learned without AI.
https://terrytao.wordpress.com/career-advice/theres-more-to-...
i'm using LLMs to do personal aspnet web project. i'm not skill in javascript. i let AI handle it and learn.
Sure. It doesn't mean that you're learning javascript though. This is where the OP, hodder, is wrong.
I agree with the sentiment of the parent comment. But I also sometimes question what you have posed exactly. Am I really learning or getting the curiosity itch scratched. The line is very thin sometimes. But I think I am more in the camp of learning. I can ask so dumb question that I never had the courage to ask nor were they entertained where I was educated. It has been a boon to finally ask.
For example I am following Dirac's book "The Principles of quantum mechanics" to study QM. Pre-AI wouldn't have been able to do so, I am just that dumb. Even with AI its tough. but the thing is I can keep asking questions until I get that concept drilled in. Now I am doing it at a pace that's unfathomable to me.
But now that I am getting to grips with QM, I can get to things that I am really interested to learn like spin resonance and so on. This is something I am so grateful for.
Now it can be questioned that is it making me wise, intelligent or just "giving me answers" that I should strive to discover myself. I dont know the answer to that. But studying what I want, how I want and not getting judged is something i deeply enjoy. Srry the comment might have taken some tangents.
> For example I am following Dirac's book "The Principles of quantum mechanics" to study QM. Pre-AI wouldn't have been able to do so, I am just that dumb
Why exactly wouldn’t you be able to learn pre-AI?
are you learning new things or are you just reading LLMs output? there might be some overlap, but those are very distinct activities
I have the same experience, but there's another dimension I want to throw out: breadth versus depth.
I've wildly increased my breadth of learning. If I'm ever curious about anything, even a passing thought, I can scratch that itch in a way I never could before.
But am I going deep? Acquiring new skills? Eh... I usually go far enough to unblock myself and/or settle a curiosity. I don't think that's good or bad, but it does present a certain set of tradeoffs that are different than going deep.
Yeah, it is like going to a restaurant to taste a foreign meal rather than looking up how to cook that foreign meal yourself, cooking it and that way getting to experience it.
Isn't this effectively the same argument an addict would put forward?
>Im learning new things at a pace I never imagined at 40 years old
Are you learning, or are you simply consuming?
I think we're missing long form studies that show if it's possible to learn deeply from compressed AI generated summaries of topics.
Tool use typically follows this curve. If you want to preserve a skill you have to actually preserve it. This isn't inherently bad by itself, tools enable us to do much more than we can without them and its a point of contention whether or not any skill is inherently important when a tool comes along that does it for us.
One of the challenges here is that the skillset we are in danger of letting atrophy is essentially unbounded. It’s not a specialized tool like a calculator, where you have a well scoped domain of problems you are offloading. granted, in practice many people are using ai for specialized domains (like coding or producing visual designs). But whatever level of abstraction they are currently working at is not, in principle, something that they couldn’t also offload to ai.
The instinct of many people (myself included, a lot of the time) is "feed the AI the problems from the domains I'm least good at / familiar with" because those ones are the most frustrating and where I'm the most likely to spin my wheels. It's easy to imagine how this feeds into a cycle of "only put forth effort at things I'm already good at" and a consequent narrowing of professional / intellectual development.
The huge problem in this specific case is that to use this tool well you also need the underlying skill to be developed and preserved. It's very different from a power drill.
This is true of most tools, it's just that we generally consider the skills simple.
I'm very bad at using power drills.
The difference if this is okay is how reliable the tool is. If it is a calculator or compiler it is okay. The example in the article also sounds okay (machine learning image classification), though I am not sure.
LLM output is unreliable, so we still need to judge it. If I want to be able to judge code, I must have worked with it to a certain extent. So the unreliable tool does not help me much if I don't want to accept the unreliability.
Sure. But if you don't own the tool and it is held by a cabal of centralist (even political state-adjacent) parties, you're having a bad day when computer says no.
The tools aren't owned by a cabal of state-adjacent parties. Specific implementations are. But open-source models have saved the day here.
If they can keep up. Unfortunately, we learn from previous technology shifts that the masses will always favor ease of use (to the point of infinite scroll 5 second videos dopamine puddle, or echo chamber social networking in lieu of critical media consumption), which does not bode well for the market for alternative hardware: one which is already expensive.
I fear on-prem AI is likely to become as popular as on-prem servers without Cloudflare using self-hosted email are today: that is to say, people have heard of it but the skillset is almost popularly eviscerated, external policies make it progressively impractical, and anyone who does it is 'niche'. While basic guides will exist, obtaining top-level output will probably require many moons of concerted effort.
Basically: AI is SaaS for thinking.
> If they can keep up.
Consider another perspective: They don't have to keep up. Once a model is good enough for a task, the model can stay. A hammer is a hammer and a hammer from 100 years ago still has most of its utility.
Similar to the hammer, it's not unreasonable to think that some classes of work will simply be solved by some model generation and whatever happens at the frontier after that does not matter all that much for work that puny humans do.
Then, of course, there will be a time where all of this is moot: Absolutely no human will want a human to diagnose their medical issues. That is not a skill deterioration issue. We simply will concede that we are not able to do it as well as a more capable system can, and without much fanfare, increasingly delegate, as we have always done.
Carrying your analogy further, let's assume all human jobs fall under good enough open source models. All human problems (food, shelter, not clobbering each other over the head because of monkey genes) are solved through a combination of AI and robotics. Maybe we even remove governments, police, and live in a future post-capitalist ecotopia.
Even if this occurs, and I don't trust well-resourced humans to allow their existing apex-predator positions in present era capitalism to be overturned, the action - as far as either humanity or AI is concerned - will still be at the forefront of possibility: a front by definition invisible to old models. And someone has to pay for the hardware to be there. Do we (a) allow private-sector dominance, effectively depowering traditional nation states and empowering a private cabal beyond historically conceivable levels (b) nationalize thought (c) head in sand and pretend it will all go away?
Most of the world seems to be with strategy C right now, strategy A is the advancing default and has already achieved extra-terrestrial reach with a threat of extra-terrestrial persistence, and strategy B is potentially scarier than the other outcomes if it goes wrong but might be lovely, if you believe in nordic state funds, solarpunk futures and socialist utopia.
Interesting times. By the way, if anyone with AI capitalization reads this, I'm looking for investment to feed humans more efficiently and have a NASDAQ reverse merger under negotiation and effectively priced out with board buy in. Just need capital support. https://infinite-food.com/
On-prem versus cloud inference doesn't matter for concentration of power.
Concentration of power exists when the model makers are the same as (or control) the inference providers. Making a model is capital intensive, so there aren't many of them. Providing inference is not: I don't even need to own GPUs; I can rent them from those who do and then sell by the token. B300s cost less than $4 an hour currently.
Cloud can even be more effective at lowering concentration of power than on premise. Asking people to individually buy $20,000 of compute equipment plus power and cooling equipment to run a frontier model is not something they're going to do if they can just pay four-tenths of a cent per output token. If the only cloud inference providers are the big proprietary US titans, that means you're going to get far more power concentration than if open source inference providers are an alternative, because then I can just switch my API endpoint.
Eh, this is our species first contact with that type of technology. A good number of voices see how deleterious these things are, and it’s all still very new. Future humans will tell parables about the evil tech bros and their silly obsessions, and the unequal accumulation of capital. This will be seen as a dark and stupid time, but I think we’ll persevere - the tech bro set is much weaker than they imagine, and certainly than they project.
We love our open source models. GLM 5.2 came out recently and the timeline for closing the gap to closed source shrank to something like 2 weeks by popular measurements.
Edit: Mis remembered the timeline I saw not 2 weeks, 3 months, but still I think my point stands.
Just to make sure I understand your argument. Are you saying that today's open source models are on par with frontier closed models of two weeks ago? By what criteria?
Sorry definitely mis-remembered on my part, its about 3 months
https://x.com/yaroslavvb/status/2067367657272422584 https://x.com/voratiq/status/2067667800643268928 https://arena.ai/leaderboard/agent
HN thread from 2 days ago:
https://news.ycombinator.com/item?id=48567759
Commenters there were saying GLM 5.2 was roughly equivalent to Opus 4.8 in coding prowess, based on personal experience of the people commenting. Opus 4.8 came out on May 28 this year (so more like 3 weeks ago), GLM 5.2 came out 2 days ago.
I haven't tried this specific model, but you can understand that a lot of HN testimonials are bound to swing extremely pro open source. I certainly hope they are as good. But I have personally tried a lot of models that are supposedly as good as the frontier ones and have found them lacking.
After a while, you do start to start to skip a couple rounds of open source models until there's a notable release. That, and the resources needed to run them are increasingly bought up by the owners of frontier models
That's when you load up qwen3.7_67b_21a_heretic_ablated_ablitereted_uncensored_agi_sparse_attention_MTP_SuperHOT_q6_AGI_FINAL.gguf
The biggest negative I'm seeing is that people are moving _too fast_ to evaluate the stuff they're slinging. When you move too fast, you do not take the time to build up your taste or appreciate the nuances of different approaches to the same problem.
I am using LLMs quite a lot, but the amount of time I spend sitting on some slopped out code is I think on average much longer than a lot of my peers. What I've found is that while the original thing "works", it usually winds up being another 2-3 cycles of iterating on the original idea after I've let it settle in my head before I actually feel confident about merging.
As a result, when I add it all up, for actual "this is important" design-level concerns, I do not feel significantly more productive.
This is hardly surprising. Think about Staff+ engineers: their work is a lot like commanding AI. Most climb the corporate ladder through hard work, excellent engineering skills, soft skills, and a bit of luck. But as their companies grow, they gradually spend less time coding, debugging, or doing deep design.
Instead, they act more like highly technical product managers. They help VPs plan and write high-level product requirements sprinkled with technical terms. They draw boxes on whiteboards and create pretty slides. They write polished documents that keep their leadership happy, and they are either in meetings or on their way to the next one. When they have a technical idea, they dispatch a team to test it out.
Naturally, they still feel deeply technical, until the day they have to resolve a production issue, pass a technical interview, or write extensive code. That is when they realize their skills have grown rusty.
I point this out not to criticize, but to highlight a genuine career challenge. As an engineer, I would rather hone my technical skills. Yet, if you want to climb the corporate ladder, you have to take on more organizational work. The only solution I can think of is to become more like a researcher or a professor. Over the years, good professors spend less time writing papers or deriving formulas. However, their insights are so deep that they still produce amazing results by advising PhD students. But that path is much easier said than done.
Welcome software engineers to the world of automated production. Your skills will drop, the usefulness of your skills will drop, the pay for your skills will drop, the quality of output you produce will drop, everything will be worse for everybody except for the person that legally claims ownership over the machine and what it produces.
100 years from now people will wonder at how anybody did anything manually writing all their code, but at the same time will have their mind boggled that a game like Roller Coaster Tycoon is possible without 1024 CPU cores running a few terabytes of memory in order to control every single person walking around the park.
Of course. If you don’t use something it atrophies not only in non use but in losing interest in keeping up with the state of the art in said tech.
What we gain though is for people don’t possess that knowledge in the first place, now have this superpower. I know several individuals who have vast experience in specific disciplines and they are now able to solve real problems where there were previously struggling and having to make existing solutions work.
In the context of software engineering it allows people that have great institutional knowledge bypass the software market and construct stuff on their own - or at least prototype something and turn it over to an SE if the situation dictates.
I’ve been using CC for several months now and have noticed an increasing quality of output - Fable 5 I think was 85% there. At 95% SE’s are going to be increasingly looking for work to do.
To the title though, I’ve noticed while my desire to actually write code is decreasing CC is forcing me to improve my high level thought processes in the context of overarching goals in a project through discussion with CC. The software often introduces things that had escaped me or just think more outside the box.
My concerns are that this technology will be restricted at some point and the people making the restrictions will have a lot of control - and we know how that works out. But I believe they are inevitable, first obvious example being Fable 5. Are guardrails needed - yeah sure. Common sense says that I don’t want someone able to concoct an easily transmittable Ebola virus that has a 90 day incubation period in their kitchen but I do want an entrepreneur to be able to build a competitor to MS Office, or a cure for Ebola, for example.
> What we gain though is for people don’t possess that knowledge in the first place, now have this superpower.
Indeed. The great innovation of AI is giving people with wealth access to vast amounts of knowledge, while limiting the amount of wealth that people with knowledge can access.
It's completely bass-ackwards.
All the benefits you're describing apply to the present moment; people with knowledge, self-discipline and expertise can leverage LLMs to great effect.
How many people like this will exist in a decade? Two?
Meanwhile the people with mountains of money can leverage them to enough effect to make even bigger mountains of money, by hiring less-than-skilled people who will do way more with it than they could without it.
Oh ho ho....looks like I'll need to update my essay! The evidence is mounting higher and higher.
https://larsfaye.com/articles/agentic-coding-is-a-trap
Is this a situation where AI will go away and we will regret the loss of skills? At worse, we will be forced to use open weight models instead of the cutting edge, so I don't think it's a big deal. I'm sure people got worse at arithmetic after the invention of the calculator.
Did the entire history of computing prior to 2023 disappear? Are we really that boned?
I don't think the real threat is at the individual level, but at the societal level.
Building skills over time leads to insights that lead to innovation.
AI does many interesting things, but it doesn't innovate (yet).
The real threat isn't that we'll all lose our skills (possibly) and then lose access to AI (unlikely), it is that AI will remain at roughly currently levels and we'll dull our skills due to reliance on it and innovation will stall because we've offloaded too much of the thinking to the non-innovative machine.
I'm not saying this is what definitely will happen, but it does seem like a very possible outcome.
The number of people now involved in software development has now increased because of a lower barrier to entry. I know many people who would previously use a no-code tool or hire offshore devs, or simply not have their problem solved, who are now vibe coding. Many of these people couldn't write very much code manually if they had to, but they're closer to understanding software than they were previously.
Yeah, this describes me perfectly. I can't really program, but I'm now building a bunch of different projects and submitting PRs that people seem to appreciate.
https://github.com/kristofferR
One of them, a Home Assistant integration for controlling adjustable beds, would be borderline impossible to do well manually - I've vibe-reverse engineered the Bluetooth protocols of more than a 100 Android apps.
> I've vibe-reverse engineered the Bluetooth protocols of more than a 100 Android apps
Impressive coverage, but how do you verify that the reverse engineering matches what the hardware expects? Do you find people who own those beds and get them to test stuff for you?
> The number of people now involved in software development has now increased because of a lower barrier to entry.
I don't see what this has to do with what I posted.
Yes, AI lowers the floor on software development, and there are positive aspects of that, but that doesn't change the possibility of an innovation stall.
re: concerns at a societal level. More people are making software than before. People who have been coding for a long time have moved up an abstraction layer and are further from the code. But many people are actually closer to the code than before.
> But many people are actually closer to the code than before.
More people are coding, I wouldn't say they are closer to the code if they are vibe coding. Are any of them going to produce the next breakthrough in computer language/framework/method of development/etc?
The risk of AI is that we dull the skills of enough people at the high end of the state of the art of the nuts and bolts of software development that we slow down innovation on that end. That's the concern.
Previously-non-programmers vibe coding CRUD apps they never could have before is all well and good but really has nothing to do with this concern. They may create wonderful and successful businesses but they are irrelevant to computer science related innovation.
Congrats you’re the only poster in this whole thread who gets it.
The standard response is ‘Dw AI will recursively improve and that’ll take care of all the emerging issues”.
Lmao ok bro
> I'm sure people got worse at arithmetic after the invention of the calculator
For LLMs, we can see this sentence but replace "arithmetic" with a variable X
I'm sure people got worse at X after the invention of LLMs"
The problem isn't that X skills atrophy necessarily
The problem is that for LLMs, X is "basically all knowledge and communication skills"
Can we really tolerate a society where "basically all knowledge and communication skills" are atrophying?
At worse you suffer from cognitive atrophy. Become more dumb and lazy.
There are a lot of mundane coding skills I consciously put off learning in case they'd ever become obsolete, and now I'm glad I did. Like sure learning React was good, but Angular? That boilerplate is Claude's job now. Ruby? Forget that.
For me personally, holding a dialogue with several agentic dev streams in parallel keeps me very much on my toes mentally. It's quite exhausting, and certainly not letting my brain atrophy.
I'm probably losing some coding skills, but replacing them with different ones, and honing some others.
I used to manage dev teams of 20+ people inside high pressure, high stakes projects.
I've been coding all my adult life, on big things and small.
To me, agentic engineering is a deja vue of managing teams, except in real time.
One thing I have found out immensely useful for is exploring new niche areas.
One example is I got it to setup a new terminal (wezterm) and configure it quickly to my liking. This would have been a lot of googling and reading before, generally too much time for me to invest.
I don't think it's valuable knowledge for me to have apart from a high level to evaluate the possibilities.
Humans will become individually and independently less skilled while having access to tools that allow them to do far more than even the most skilled human could, before having access to these tools.
I'm not sure if we'll become less intelligent. I think our sacks of neurons are gonna keep on making associations, just across a different set of topics.
MDs NPs and PAs currently are offered free access to a medical AI app called Open Evedidence. Was just discussing this with our chief medical officer and while he is quite enthusiastic about it, he could already see this same dumbing down effect emerging among his providers. The solution I think, is 5 or 10 ten year re-certification exams for MDs, currently reqd for "Physician Extenders" but not the Physicians who supervise them. Continuing compulsory education and re-certification works, I suspect, in all highly skilled fields that are both augmented by AI but also degraded by it.
My compiler writing skills atrophied with the advent of high-level languages, but in exchange I got more done. There is still a very well paid market for compiler writers, but the fact that not everyone needs to be one has made the world richer overall.
Comparison never made sense to me
1/ When dealing with High level language I am not seeing assembly or the language it compiles to. It's not a leaky abstraction
2/ It's deterministic
The day my markdown file is the thing I deploy on AWS your analogy will stand
You also don't need to directly maintain or evolve the generated machine code to fix bug or add new features.
I like it when AI metaphors presume determinism.
As jatins says: Why bother with source code? just deploy prompts and specs. It’s rare for C programmers to care about assembly, or python programmer to care about C. Why do you care that much about source code?
You clearly don't know what you're talking about if you think compilers are a good analogy.
They really are though. There seems to be some confusion about "determinism" as if everyone writes perfect code and every compiler interprets that code perfectly. There's bugs and nondeterminism in every step of the chain with human coders and, say, a .NET compiler too. The idea that we've got coding down to a pristine and exact science and AI is some random chaos element that muddies it makes no sense to me after seeing how most human created code works.
Colonoscopy detection has gotten modestly better with AI assistance.
https://pubmed.ncbi.nlm.nih.gov/39216648/
https://www.cancer.gov/news-events/cancer-currents-blog/2023...
https://www.nejm.org/doi/full/10.1056/NEJMoa1309086
https://info.asge.org/083024-colon-asge/acg-quality-task-for...
New tool that does task better than worker leads to workers being less good at task. Net outcome for patient is positive. Next?
Programming: "for a given task, if you take a shortcut then you will not have the familiarity and expertise that someone who took the veritable and righteous path would have".
The question is then, what did you do with the extra time. If it's fuck all, then yes, that's a liability.
Like any technology, it comes down to the disposition of any given person in how they plan on applying it.
Not trying to say it's all going to be awesome. Definitely maybe the opposite. These arguments are weak tho.
If your argument boils down to "individuals just need to use it the right way" then I ask you to kindly use that same framing for driving cars, owning guns, etc. and then rethink your position.
It turns out that when the use of a tool has external consequences for misusing it, it's important that there are structures in place for penalizing the misuse of that tool.
I don't feel like AI coding has ruined my skills, and I could go back to manual coding any time.
However, I cannot build a good mental model of a software component that I didn't write myself, and that can affect future maintenance if that component is not properly decoupled from the rest of the system.
Personally? Yes (and I hate it).
Autocomplete of entire functions and methods. Nice, but also really boring. Takes the fun out it. It's all about fixing sup-par code now, a line here or there.
It's just boring. I tried writing some code by hand today after a few months hardly thinking about things and it was really hard to do even the simplest stuff.
I suppose tractors and cars ruined our skills at feeding and caring for horses.
Losing a specific skill to automation isn't necessarily a bad thing. Losing the ability to learn things would be however, and that would be my fear with AI, but I'm not sure it's well-founded. Humans learn naturally by interacting with the world.
Some effect is real, but it's likely overstated by poor metrics.
"Currency" in all fields relates to the recency and frequency with which you dealt with a particular issue. Whether flying on auto-pilot or coding with AI, automated reduces some currency. But is that a reduction in capability?
Measuring concrete tasks makes currency the operative skill; that's why it works to cram for standardized and mid-level tests.
(Indeed, the 2010's interviewing "wisdom" about people being quick to answer simple questions veered into measuring currency, not skill.)
I think this effect is strongest in time-impacted professionals. Doctors doing dozens of endoscopies a week and developers churning out code will use what tool leverage they can, and forget as much as possible to focus on what they need to. I suspect the effect is weaker in personal or research projects.
People riding bikes won't be able to run long distances - because they won't have to, and will be able to outdo any runners. That's only a problem if the supply of bikes is someone constrained. So the risk is not skill loss, but losing control of the means of production.
How do you measure the bundle of "skills" that comprises critical thinking? And if that's your analogy's distance running, what's the "riding a bike" analogue?
No, AI isn't doing anything. When someone gets stabbed, do you say "are knives stabbing people?" No, just like you don't give credit to the knife for cutting up your vegetables.
"Just being aware that this phenomenon exists hopefully provokes some self-reflection about which skills people want to maintain and which they’re willing to outsource” to AI tools. Right. Obviously.
So we need to be teaching that core lesson to children -- they don't retain skills that they don't practice. And we need to be careful to decide what skills and verify they are learning them. We also should absolutely be using AI to provide personalized instruction to every single student.
Blaming the tools for things that humans do is incredibly stupid and dangerously misguided. Because it shirks responsibility onto the technology, when technology is the best lever humans and society have to improve things! It just happens to also be the best lever available to make things worse.
This negative view of improving technology starts from a warped and very unrealistic concept of the state of the world, where it has been, and the role technology has played.
1. Technologies, starting with fire, the printing press, etc. have been critical in raising life expectancy, standard of living, etc.
2. The world is still a profoundly unequal and exploitive place.
3. AI and robotics have the potential to provide everyone on earth who wants it with extremely inexpensive labor to help them with anything they need or can imagine. This will be a dramatic shift in quality of living.
Human society is the source of our problems, not technology. Part of this is that I think deep down people believe that any tools or developments that arise will just be used to exploit and suppress them more, and there is no alternative. In this case, I guess people think the best outcome is to go back to feudalism or some nonsense because technology just makes things worse.
But why stop there? Why not go back to, I don't know.. fire? Or maybe no one should ever eat any red fruit?
Why are you comparing it to technologies with a positive outcome? Why not to opium, tobacco, or alcohol?
> So we need to be teaching that core lesson to children
Yeah, that has always worked out well.
AI should demand tokens of attention from proffesionals little challenges to continue, turn the toolise into an adventure. The gold standard would be the primer from diamond age
Wow so technology doing tasks for people has an effect of diminishing their skill at those tasks. Truly unprecedented.
The bit about computer science (behind the paywall) starts:
> To investigate whether skills are being lost in the field of computer science, researchers at the AI firm Anthropic in San Francisco, California, designed a randomized controlled trial in which 52 software engineers were asked to perform a basic coding task
That's this study here: https://arxiv.org/abs/2601.20245 - also written about on the Anthropic research site here: https://www.anthropic.com/research/AI-assistance-coding-skil...
Of course it's ruining skills. Anyone who has a really well honed mental skill knows that we get rusty really damned fast (e.g. musician at a professional musician level, which happens to be my reference point). You keep shedding, you stay sharp. You take more than a week off, it's going downhill.
There's no way this isn't happening with other skills. No goddamned way. Anyone who tells you otherwise is grifting.
Luckily for me, I was never truly that intelligent nor skilled to begin with.
Back in the mid-90s I was doing desktop support. It was a lot of work because PCs were relatively new (and they were garbage!), and people broke shit all the time. Sometime around '96 a disk-cloning utility called Ghost was released. It was great - one could provision a fully working PC with all required apps and config settings in minutes! Sounds lame now, but back then it was revolutionary. It had a dark side, though. After about a year most people I worked with had lost the ability to troubleshoot even the most basic problems. The solution to every problem was to just re-apply the standard Ghost disk image (we called it 'Ghosting' back then) ... Can't print? Ghost it! Not receiving emails? Ghost it! Word is too slow? Ghost it!
https://en.wikipedia.org/wiki/Ghost_(disk_utility)
Happened with all of tech support, really, or at least in my corner of the world: you take your PC to a technician and you receive in return a fresh Windows install, a folder with most of your files (including a copy of "C:\Windows"), and none of the programs or shortcuts you had before.
> To investigate whether skills are being lost in the field of computer science, researchers at the AI firm Anthropic in San Francisco, California, designed a randomized controlled trial in which 52 software engineers were asked to perform a basic coding task3. During the exercise, all 52 participants could search the web and access instructions on how to do the task. Half of the participants were prompted to use an AI assistant as well.
> Afterwards, all of the software engineers were asked to complete a quiz about what they had learnt from the task. The participants who had used an AI assistant did significantly worse on the quiz than those who hadn’t: the average score was 50% in the AI group versus 67% in the non-AI group.
This doesn't strike me as a great test? Most engineers aren't going to learn anything from a basic coding task anyway, so I do wonder exactly what they were testing there. If it was just recall about what the issue was, then it doesn't really strike me as a problem - using AI to handle simple problems that it's clearly capable of dealing with is the right way to use it, and of course you're not going to spend time poring over the details because then you haven't saved any time by using AI.
There are other examples that don't strike me as particularly problematic, like GPS eroding people's sense of direction. It's totally reasonable to let a skill atrophy that you no longer really need because you have an ever-present tool to handle it. I'm a lot worse at doing long division than I was when I was <whatever grade one learns long division in>.
The whole skill atrophy thing seems like much less of a problem than it's made out to be. We've been letting skills atrophy for good reason long before the advent of AI. If you start at McDonald's as a fry cook and work your way up to regional manager, if you suddenly have to work a shift on the fry station you're going to be worse than you were when you were doing it all the time. MDs at investment banks almost certainly can't put together a pitch deck as well as the junior bankers who are doing that task regularly. These things are fine - part of moving up in the world and having a broader impact is being able to successfully delegate tasks, and when you delegate tasks your skill at those tasks will atrophy. No real difference whether you're delegating them to AI or not.
To be clear, there are of course cases where skill atrophy is bad. iLoveOncall posted about senior engineers in their org who have lost all of those skills and their judgment along with them. That's definitely bad! If you delegate so much that you lose the ability to even judge good work, now you can't even delegate effectively any more.
I think the real lesson with AI is that you need to be self-aware about what skills you should practice and retain vs. what skills you can let atrophy, since it's easier than ever to hand things off. I've lost most of my ability to write a SQL query, but that's fine because it was only a skill I used intermittently and AI can always do the job fine at the level of complexity I need. I have not let my skill of writing product specs atrophy (I am a PM, in case you haven't read my username), because that's critical to using AI correctly in the first place.
How many of you feel this, anecdotally? I had to come up with a class hierarchy for a roguelike game the other day. This should be something that's dead simple, off the top of my head, no problem.
And suddenly I was stuck! It was like thoughts weren't forming properly. My instinct was to use Claude to help brainstorm, but I resisted. 5 minutes later, I finally broke free and instantly came up with the plan.
What the hell?
I realized I'd offloaded my planning onto AI. I would ask it for plans and then choose the best one, but that's a different skill than coming up with the plans in the first place. My skills were rotting.
I don't know what to tell you bro. I type much less code but I haven't stopped thinking altogether. I still write to myself and to others. I read longform. I walk and introspect. I think about high level problems. Any general cognitive decline I ever notice, I attribute to having a bad day or getting old...
I think it’s like shoes. Shoes are generally better than callouses at protecting your feet, providing traction, etc. Almost everyone accepts the trade-off and wears shoes as opposed to going barefoot. To be sure, your feet lose some of their toughness. You may become reliant to some degree on the tool. I know that when my kid started wearing shoes, walking became actually a bit harder and be fell a few times. The shoes were a little clunky and lacked the same sensation. He changed he gait to accommodate, but over-compensated. They took a while to get used to wearing. AI also often feels a bit clunky and lacks the same feel. However, I think some people are confusing the shoes for the feet. The shoes don’t do the walking. The LLM is the words, but not really the ideas. If you stand or pose in shoes, move your legs around and wait for the shoes to take you somewhere… that’s not going to work.
At least for writing, I think AI is mostly useful for the types of writing that aren’t particularly interesting or worthwhile in having to begin with.
In concrete terms, AI isn’t all that useful for writing a personal blog, because no one wants to read obvious AI slop. But it is useful for creating boilerplate product pages, FAQs, and other types of writing that weren’t very interesting pre-AI.
So it’s not really a huge deal to me that my skill for writing descriptive product page text or FAQs is atrophying, assuming that it is.
Not worthwhile feels a bit strong (a good FAQ is definitely worthwhile!), but I definitely agree that there is a big difference between any kind of art (writing, playing music, creation of images/videos/etc.) for its own sake and for commercial purposes. AI is terrible for the former but perfectly fine for the latter.
There will always be value in a human writing fiction or a memoir or even a Substack. The human perspective is inherently valuable there. Much less so with ad copy that's just going to get A/B tested ad infinitum until a winner is picked out based entirely on data.
Same with visual art. Art painters aren't going to lose their jobs to AI, but once you've got a robot that can paint a house reliably, house painters are done for.
Is C ruining our memory allocation skills? Early results are in - and they're not good
https://archive.ph/NS4Ae
Technology like many things is about _how_ you use it.
If social media is consuming first, or primarily consuming, anyone can scroll their way to a negative rabbit hole that never ends.
If creation is the use it's something else entirely.
AI in the form of interactive chats, can be a novel kind of consumption.
You can have passive conversations in terms of asking a magic genie, or more active ones.
It is not said here much given the AI psychosis, but firms like Anthropic do not allow you to use AI in their interviews for a reason.
They know that this is one of the biggest de-skilling programmes they have seen.
So expect the return of in person Leetcodes and whiteboard challenges.
What is most troubling for me is seeing kids just switch off when LLMs are available. Doing homework they will have zero interpretation or contemplation, just enter the question as a prompt and record the result. LLMs appear to have the ability to interfere with the most basic aspects of attention and executive function.
That’s got nothing to do with llm’s and it’s got nothing go do with ‘how technology is used’.
I’m not even sure why this has to be said.
Universal fact 1) most people lack the self discipline to choose what really benefits them and takes care of their well being. We already know this - obesity rates, excessive screen time etc.
The solution is to control the actions of people. But most of you ain’t ready for that conversation.
AI has allowed me to keep shipping features and system even when holding a normally managerial position, so if anything it preserved some of my coding skills. I'd not have seen any code otherwise (writing code is a huge time sink compared to managing things around an org).
I pity those who need to contend with that as ICs, though.
Hopefully this is not your case but as an IC the “manager, who didn’t write code for a couple of years” that decides to come back and “help” is one of the worst experiences. Code is usually subpar, uses old idioms, and comes with added pressure. With AI this can become much worse, because of the sheer volume of code that can be thrown at the IC.
Nothing like that.
My current position is more on the operational side and AI allowed me to create a pretty significant software system for our technical ops, that the wider org liked and given me the resources to recruit flesh-made engineers to support.
A rare case of AI creating jobs.
AI/AR Glasses that show you every piece of knowledge about whatever your looking at might not help, especially if the AI is wrong. Otherwise everyones a know it all?
It's going to get a lot worse than AI being unintentionally wrong. We're only at the start of the curve for this, but it's obvious that controlling what LLMs show in response to politically-charged topics is huge lever of power for those in charge.
So now imagine you're using Chineses AI/AR glasses that you've come to rely on for "knowledge" and you look at the famous picture from Tianemen square: "Doesn't look like anything to me".
Are employees ruining managers' skills? Late results are in - and they're not good.
That's a great analogy - everybody knows you forget pretty much everything after a few years away from the line. It should be seen like one of those studies that proves smiling is correlated with happiness. :-)
The two senior engineers in my org (in a FAANG) who vibe-code the most have lost literally all of their skills. Their code has become terrible and their judgment even worse.
A very similar topic was discussed here: https://news.ycombinator.com/item?id=48392004 and I make the exact same conclusion:
All of this makes me selfishly excited for my own future. It's glaringly obvious that anyone who's a heavy user of LLMs is atrophying their skills in real-time. I have yet to meet a single person for whom it's not the case. But I essentially completely stopped using them for software engineering (why isn't really relevant, but it's not because od this skill atrophy). So as the skills of everyone else is diminishing, mine is proportionally raising.
It has never been easier to get better than others. You don't need to put in more effort, just the same effort as you always have, and others will do the job of losing their skills for your own benefit.
Idk man my system design game is better than its ever been because I put in the effort to use these tools and recognize they can't do software design better than I can and because I've increased the scope of what I'm building I often have to think more deeply about the problem up front. A typical speccing sessions lasts a few hours for me on big work before I have AI start writing that work where I'm just going back and forth on what I want, points of consideration for performance, usability, structure etc pushing back on where AI (always) chooses the most naive way it wants to do something.
Every time I see an anecdote like this, A it reaffirms my belief that FAANG devs are fairly mediocre on the whole (not saying this is you, obviously there are good FAANG devs) and B it reaffirms my belief that the developers who kind of give up their thinking like this are really using the tool wrong or didn't really care about the work before AI either so its now just a quick means to an end.
+100 on this. In addition, if you don't outsource your thinking and you're willing to go through all this, you absolutely don't need the top tier models.
> Every time I see an anecdote like this, A it reaffirms my belief that FAANG devs are fairly mediocre on the whole
I think another (partially causal?) problem is how they're managed. The whole perf circus is just ridiculous, especially the stuff recently reported about facebook. But they're all more or less like that. Steeped in that cocktail of incentives, who even knows what might happen to an otherwise excellent engineer.
But also just numerically, they can't be much above average, on average, because there are so many.
> But also just numerically, they can't be much above average, on average, because there are so many.
There are as many CS grads every 2 years in the US alone as total software engineers employed by FAANGs...
I stand corrected! That's very interesting.
I strongly believe that you cannot evaluate how good a system design is if you don't implement it by hand.
LLMs will implement what you ask them to, even if it is the wrong approach. They can be lazy and take shortcuts all the time, but they do not feel PAIN (obviously they don't feel anything and aren't lazy, I'm just personifying them but you get the point). Only when you implement by hand can you feel if the implementation of your design is painful or not, and only this signal can tell you if your design if truly good or not.
I do think LLMs are useful for design work, they are good at asking clarifications and probing questions which actually do push you to approach problems differently, but leaving implementation of designs to LLM is a recipe for disaster, and judging your own design skills when you're not implementing such designs is seriously laughable. And to be clear, it already was before LLMs, when "software architects" were just designing and then had peons implement for them.
LLMs are enabling a whole new level of bad code that is best describe by the following Jurassic Park quote: “Your scientists were so preoccupied with whether or not they could, they didn't stop to think if they should.”.
> A it reaffirms my belief that FAANG devs are fairly mediocre on the whole
Off-topic but having worked in other companies as well, I can guarantee you that this is not the case. The skill of engineers in FAANGs and other "top tier" companies is much higher than average.
> I strongly believe that you cannot evaluate how good a system design is if you don't implement it by hand.
Fwiw we both agree that LLMs should not design systems. I do the design, but otherwise I don't get how this is true, the success of a design is indicated by long term success in the system it built. You can measure this against success in the task it was deployed for via performance metrics for one. And then from a developer standpoint how easy it was to maintain later on. Success of a system is a measurement over time, but it's not some quality that can only be measured by those who built it.
> Off-topic but having worked in other companies as well, I can guarantee you that this is not the case. The skill of engineers in FAANGs and other "top tier" companies is much higher than average.
I have first hand knowledge of this so I agree to disagree. Being surrounded by google, aws, and meta folks my understanding is the best people leave faang when they get the itch to do something better with their time.
> my understanding is the best people leave faang when they get the itch to do something better with their time.
Your understanding is wrong. The extreme majority of people who leave FAANGs don't leave willingly, and are not the best, unless it's to go to even better companies like Jane Street and the like, but those are rare.
Turns out for most people, there's not much better to do during your 9 to 5 than solving problems for half a million a year.
> Turns out for most people, there's not much better to do during your 9 to 5 than solving problems for half a million a year.
Lol. I don't know I was talking with a guy yesterday who left a FAANG clearing 1.4 million / yr who's now running his own successful startup. My sample is successful founders backed by top tier VCs or exited founders who have done much better than FAANG. If you can't do more you stick around the FAANG, those who can go and do it.
You're talking about what happens to engineers who decide they want to be small business owners rather than engineers. That's orthogonal to this thread about where engineers with the best technical skills end up.
> Only when you implement by hand can you feel if the implementation of your design is painful or not, and only this signal can tell you if your design if truly good or not.
I think this touches obliquely on a point I keep coming back to, that one of the most important things a codebase does is to communicate ideas about how a process should work. Yes, it also produces some binary that runs on a bunch of servers or whatever, but that's a really temporary, ephemeral artifact. The lasting thing is the idea. Making your ideas (expressed in code) easy to understand, easy to work with, and easy to evolve in time is the art of software engineering. I 100% agree, from my own experimentation with LLMs, glancing at something a model has produced and checking that it has some test coverage isn't enough to know whether it's well-engineered. You'll only find out later when you try to work with the code.
I recently did a manual exercise to force myself to keep my skills from decaying too much after about a year of using agents exclusively. My ability to go from a blank slate to software was indeed in the toilet, but my ability to reason over and edit code seems to be surviving fine. I suspect that your LLM-pilled coworkers' judgment issues are related to laziness that LLMs have enabled, rather than an inherent property of LLM use.
> My ability to go from a blank slate to software
> my ability to reason over and edit code
You can't be able to do one but not the other, they're the same thing.
> I suspect that your LLM-pilled coworkers' judgment issues are related to laziness that LLMs have enabled, rather than an inherent property of LLM use.
Does it matter? The end result is the same. Maybe the studies that the article mentions are simply showcasing the exact same effect that you're suggesting here, but it doesn't change the fact that there is indeed a negative outcome.
Not true, blank page syndrome, inertia and write vs edit flows are a known thing across disciplines. You might not think about it as much because you have a command in your package manager or framework to bootstrap projects, but imagine if you had to hand code _everything_. I'll guess there are languages/frameworks where you consider yourself quite proficient but where you couldn't bootstrap a project without checking the docs if not for the tooling.
It matters. If LLMs are like alcohol, useful but potentially dangerous, they should be treated very differently than if they're just poison. If someone's a drunk and they can't keep their lives together, it's not the alcohol's fault, it's the person's.
You're making the right call there. I've felt it, seen it around me personally. That study is terrifying.
I'm no code ninja at the best of times. It's scary to hear that's happening to top engineers.
I need an exit strategy. Anyone else come off AI?
Yes, feeling like I had to relearn to walk. The first week was rough, everything was wired for LLM usage and autocomplete. Couldn't even type right anymore.
Yeah, I had the same and decided to cancel all my subscriptions and remove all my local models ~1 months ago ago.
So far I’m very happy with my decision.
I wrote about it here: https://news.ycombinator.com/item?id=48083162
> I had the same experience over the past year with early coding harness at the beginning of the year, then Claude code since its release date. But after 1+year going that direction I really don’t want to continue. The novelty is gone, dealing with AI now feels frustrating and boring, I miss engaging deeply with the actual lower level technical challenges. I do not want to manage fleets of agents. I do not want to rediscover for the hundredth time that in fact all this time an agent took shortcuts for acceptance tests I rely upon and didn’t catch. Or once again get the agent to understand why and what I want it to do after its context got bloated and it start to drift completely. While I got artifacts I can use (libraries, tools, docs), including some things that I’m pretty confident are SoA I do not feel satisfied anymore knowing that I used a model to generate them, even if I was the one designing every part of it. I do feel that I’m lying anytime I come to a colleague to share a new cool tool I have made.
> YMMV but I’m personally feeling burnt out with AI coding agents and ready to go back to the old ways for my next personal project
And also here (specifically to human communication): https://sam.elborai.me/articles/no-more-llm-comms/.
Are you actually seeing any signs that we’re going back to how software was written before, and needing those skills in the same way? Because I sure am not seeing that right now. As someone who vibe codes 100% and has become managements favorite, while being more or less allowed to break the platform every other release I know my skills are atrophying. But it’s taking me different places in my career entirely. There’s a path to managing other engineers now that opened years before it would have previously. Even writing this makes it sound ridiculous, but that’s what’s infront of me right now. There is an entirely other set of skills that I’m interested in sharpening now. Definitely no more sitting down several hours per day and meeting about system design and integrations with others.
It's not really about going back. Evolution happens within a pattern of ebb and flow, back and forth. We never get anything perfectly right. We overdo, then course correct, rinse and repeat. Right now, we're embracing AI, but we're also noticing atrophy of skill as an effect. These may be the last generations of such craftspeople that can notice, compare, and inform as to whether there's actual loss. That future you're seeing for yourself is still being written. Stay tuned.
> two senior engineers in my org (in a FAANG) who vibe-code the most have lost literally all of their skills.
Unfortunately FAANG incentives this behaviour with their token leaderboards and general push for velocity over anything else (other than goog maybe)
The article is saying that using AI degrades certain skills when AI is not available. You're claiming that AI is making people less effective even when they have access to AI. I'm skeptical of your claim.
The article's claim is probably true, but not really an argument against AI. Using keyboards degrades my ability to write by hand but that's not a good argument against keyboards. AI will become another tool that allows us to operate more effectively and at a higher level of abstraction. Just like keyboards and Python.
Now, we still occasionally need people who can write assembly (and do calligraphy). But mostly we don't.
This is the most important point: it seems like AI tool use undermines one's ability to use the AI tool.
I did stop because I noticed the severe atrophy after using them for a year. Never again.
As someone who was self taught as a programmer and has a reasonable high level understanding of some CS concepts but not lots of experience applying them, and no good mentor, I’ve found working with an LLM really englightening. Asking Claude to think about “good ways to structure this” or asking how similar problems get solved in industry or high profile projects has really helped me design better solutions and avoid painfully reinventing wheels (recent eg was for a plugin type architecture).
I think a lot of academics and researchers who code but aren’t software engineers or CS majors are going to benefit, provided they take the time to prove what the model does and are curious about whether it’s doing something sensible!
Relative to a 1% coder hand rolling something then yes it’s AI slop etc. but it’s prob still raising the bar generally.
> Asking Claude to think about “good ways to structure this” or asking how similar problems get solved in industry or high profile projects has really helped me design better solutions
I think this highlights the difference between the “how do I make a ham sandwich?” approach of chat vs the “sudo make me a ham sandwich” of agentic coding.
Another tactic is to use LLMs to help you learn. That's another way to approach "It has never been easier to get better than others."
Avoiding tool use because you're afraid you won't be able to use the tool responsibly is not likely to be a winning strategy in the end. Learning to use the tool well is much more effective.
But they're also unreliable in what they present, they still hallucinate. I rather do my own research or listen to a real human on the topic who actually has an internal concept and structure of what they're talking about.
My friends can't ride horses as well as they used to, they've been using cars to get around too much.
Sure horses are more efficient, but cars are faster and more convenient, and allow you to get a lot more done.
Also cars will get better in our lifetime, horses are horses.
Just try to take those same friends who drive everywhere on a (walking) trail hike inside the city, then come back to report.
Your analogy doesn't make very much sense.
Perhaps it would be better to compare somebody who drives a car vs someone who used to drive but now uses Uber.
"My friends can't reason as well as they used to, they've been using LLMs to do their reasoning for them." is going to be the eventual outcome. What's the point in getting more done if it ruins your mind in the process?
Writing code seems more like walking to me; at least it is the most manual way of getting a computer program. Horses might be more like one of those low-code/no-code solutions (it really fits, they are useful but very opinionated, so not always cooperative). And, the situation with AI seems a bit worrying for them.
To continue the modified analogy, if your friends lost the ability to walk, you’d be quite worried, right?
It’s absolutely atrocious in the Fortune 500 tech sector. Multiple times a day I see people debating each other on teams about things they know nothing about and it’s obvious that it is just copilot v copilot. Crazy.
I call it Alien Slop Intelligence (ASI) that gradually turns your mind into slop. AI is sloppy on details, so at first you polish its slop by hand, then you start ignoring small imperfections, and eventually you lose taste and skills to evaluate AI output. At that point your mind has become slop.
A skill which is now done better by a machine is no longer a skill, it is technology. It is just a matter of time before most of our logical and language reasoning skills are replaced by frontier model-agents, which will at some point be far superior (if not already) to human capability.
So I totally disagree with this premise that human skills are being ruined by the use of AI technology. No, many human skills are being made obsolete. That's a good thing for economic productivity as a whole, but for those who only have skills that are being automated, their labor value decreases (which is usually bad for them as individuals).
You miss the point, the point is that by using AI, our skills (let say "coding in Rust") are diminishing and even without reading this article, we can feel it to some extent already if we aren't lying to ourselves, especially very heavy AI users.
We do however create new skills, skills that might be more relevant for the future, but still, it is controversial.
For the people peddling LLMs this is actually good news:
1. Force AI down everyone's throats claiming it's going to boost productivity
2. See people lose valuable skills because they rely too much on AI
3. Peddle more AI to make up for the lack of skills in professionals
It's not good news if the AI companies have to raise their prices 10x to deliver the same service while paying down the crippling debt they've incurred getting everyone hooked on AI.
If my mechanic started charging 10x more to fix my car, I'd learn to fix my car.