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.

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.

  if this same effect happens for very wealthy and/or very senior executives?

Yes, I'm 100% sure. It's for the same reason: if you don't actively use a skill - it atrophies; there is no workaround.

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.

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.

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.

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 learning new things or are you just reading LLMs output? there might be some overlap, but those are very distinct activities

>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.

In the eyes of many ceo’s, you are slowly becoming useless.

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.

> 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.

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.

What's an example of "bigger and better" and what projects did you have before?

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.

> 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.

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.

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

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.

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.

> 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.

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'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.

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?

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 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.

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.

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.

We never learned how to use IBM punch cards. Writing functions by hand for modern languages is fast becoming anachronistic.

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.

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 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 "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.

Isn't this effectively the same argument an addict would put forward?

Did the entire history of computing prior to 2023 disappear? Are we really that boned?

> 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.

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?

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.

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?

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.

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.

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.

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.

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.

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".

> 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.

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. :-)

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.

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.

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.

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.

> AI's greatest feature is that it obfuscates provenance.

maybe because people see AI not just as a clever packet manager, but also, to some degree, as a problem solving engine. Similar to humans.

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.

> producing more, more, more, faster, faster, faster. And for what? What’s it all for?

If you wanted a literal answer, it is to accelerate revenue as much as possible to make the very rich people who own most of the company even unfathomably richer. No benefit to you for making more faster. Other than the burnout, that's all for you to enjoy.

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...

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.

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.

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.

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?

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.

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 clearly don't know what you're talking about if you think compilers are a good analogy.

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.

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.

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'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?

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.

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.

The difference is about slop tolerance not "process over product".

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.

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"

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.

That's when you load up qwen3.7_67b_21a_heretic_ablated_ablitereted_uncensored_agi_sparse_attention_MTP_SuperHOT_q6_AGI_FINAL.gguf

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.

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.

You also don't need to directly maintain or evolve the generated machine code to fix bug or add new features.

+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.

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.

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.

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.

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/.

> 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.

The tools aren't owned by a cabal of state-adjacent parties. Specific implementations are. But open-source models have saved the day here.

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.

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.

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.

> 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.

> 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.

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.

> 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.

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.

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.

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?

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.

> 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.

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 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.

> 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...

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

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.

> 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?

I stand corrected! That's very interesting.

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.

> 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.

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.

> 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.

> 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

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/

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

> 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.

> 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 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...)

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

Spot on analogy.

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.

You probably just built a shanty town according to a professional.

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?

Most of the best engineers I know leave faang because the work at faang sucks. It is not orthogonal. The majority of faang work is not highly impactful work like trivially because the amount of work outstrips the amount of interesting work by an easy oom.

FAANG doesn't straightforwardly retain the best technical talent, and the median task there is routine.

What's your point? I could just not ask an LLM and not do anything. Is that better somehow?

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.

Yes. You could not do it. Doing stuff isn’t essential. Especially when you have no idea what you are doing.

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.

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"

> 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?

But it's not going to do that. It's literally designed to give the best and most accurate answer that it can. It's not perfect and I don't expect it to be perfect. According to my personal experience it's definitely good enough for a lot of things like recipes, coding, hobby stuff, and home repairs.

You're expecting me not to trust my own personal learned experience using AI as if somehow all my continued successes are worth nothing because you can invent some wild hypothetical. The more I use it, the more I understand its limitations and avoid them.

> But it's not going to do that. It's literally designed to give the best and most accurate answer that it can.

This is not true. This is not how LLM's work. They have no concept of accuracy or "best".

> But it's not going to do that. It's literally designed to give the best and most accurate answer that it can.

No, no, no and no. This is the biggest mistake I see people consistently make with LLMs. It is not designed to give you the most accurate answer, it is designed to give you the most likely series of words following your prompt.

If an LLM is trained on 10 jackasses thinking bleach is a medicinal drink and 1 doctor who disagrees, it will by virtue of probability tell you to drink bleach. Companies add additional safeguards or system prompts to try and keep it 'on rails' but it's all probability based on what you prompt it. It is by the literal functionality of how it works to do so. A function depending on what said company ingests during training, many of which include the entire corpus of the internet.

If you do not understand how LLMs work under the hood then yes, you shouldn't trust your own personal learned experience because you've already demonstrated that you're wrong.

You missed my point.

The system designers (OpenAI, Anthropic, etc.) are absolutely trying to build a system that gives the best and most accurate answers possible.

The model itself does not have goals, intentions, or an internal concept of "best" or "accurate."

> It is not designed to give you the most accurate answer, it is designed to give you the most likely series of words following your prompt.

That's not specifically true either; training is more complex than that. ChatGPT had to be trained, for example, to answer questions in a chat format.

> If an LLM is trained on 10 jackasses thinking bleach is a medicinal drink..

Again with the hypotheticals! You literally cannot discuss this subject without hallucinating things that don't exist. LLMs are trained on huge corpus of information from books to videos to reddit posts. Ultimately, statistically, it's going to predict the most common answer to something. Yes, that might be wrong but the vast majority of the time it's going to be right. And you know what, in the real non-hypothetical world, it works great. As much you don't want it to. You can hypothetically hallucinate as many weird unlikely scenarios as you want but that doesn't make it true.

I see there is no point trying to get through to you. These are not hypotheticals, as an example just recently ChatGPT advised a user to mix bleach and vinegar. I have seen Googles AI hallucinate many details that I had to double check and would've been disastrous had I assumed it was truthy. There's a reason why models have the legal clauses like 'this is for entertainment purposes only' because they are trying to shield themselves from legal obligations when it makes catastrophic mistakes.

The people least willing to understand how the system works are also the most willing to blindly believe it.

I understand how the system works. I also know that the free LLMs are dumb as rocks. But they're not all dumb as rocks -- it's a continuum. You're the person who doesn't seem to understand that it's not all or nothing.

AI companies are trying to build a system that gives the best and most accurate answers possible -- that's the whole point of it all.