Ie. most managers can't help their team find a hard bug that is causing a massive outage.
Note: I'm a manager, and I spend a lot of time pondering how to spend my time, how to be useful, how to remain relevant, especially in this age of AI.
But the feeling of skill atrophy is very real. The other day I needed to write a function that recursively traverses a Python dictionary, making modifications along the way. It's the perfect task to give an LLM. I can easily see that if I always give this task to an LLM, a few years down the road I'll be really slow in writing it on my own, and will fail most coding interviews.
Also, while there is a high in producing a lot in a short amount of time, there is no feeling of satisfaction that you get from programming. And debugging its bugs is as much fun as debugging your coworkers' bugs - except at least with the coworkers you can have a more intelligent conversation on what they were trying to do.
This happens a lot in the natural world in ecosystems. For example, many people plant trees and add a drip system. The trees grow to depend on the drip system, and never stretch and develop their roots -- and the relationship they have with the soil microbiome. They make the trees prone to get knocked down when an unusually strong gust of wind come through.
And you know what? That's just peachy keen. I don't need to write x86 assembly anymore. In fact, these days I do a lot of coding for ARM platforms, but I never learned ARM assembly at all. So it would take more than just a refresher course to bring me up to speed in that. I don't anticipate any such need, fortunately.
So... if I still need to write C in 10 years, why in the world would I consider that a good thing? Things in this business are supposed to progress toward higher and higher levels of abstraction.
Perhaps the biggest difference is the lack of feedback AI gives that humans can give: a subordinate can communicate if they feel like their manager is being too hands off. AI never questions management style.
Managers grow the skills needed for their organization. Their team affects them.
A process-oriented team with quality/validation mindset has replaceable roles; the action is in the process. An expert team has people with tremendous skills and discretion doing what needs doing; the action is in selection and incentives. Managers adapt to their team requirements, in ways positive and negative: becoming rule-bound, privileging favorites, etc.
With AI this might be a positive insofar as it forces people to state the issues clearly, identifying relevant context, constraints, objectives, etc.
Agile benefitted software development by changing the granularity of delivery and planning -- essentially, helping people avoid getting lost in planning fantasies.
Similarly, I believe that the winner of the AI-for-development race (copilot et al) will not just produce good code, but build good developers by driving them to state requirements clearly and simply. A good measure here is the number of iterations to complete code.
An anti-pattern here, as with agile, is where planning devolves into exploring and exploring into incremental changes for no real benefit - polishing turds. Again, a good measure is the number of sessions to completion; too many and you know you don't know what you're doing, or the AI cannot grasp the trees for the forest.
Strategy vs tactics. Managers aren't there to teach their reports skills, they hire them because they already have them. They're there to set priorities and overall direction.
While the manager doesn't need those skills, you still do.
Do all the work Farts_Mckensy does at half the price using AI.
The seductive promise of solving all your problems is the issue. By reaching for it to solve any problem at an almost instinctual level you are completely failing to cultivate an intrinsically valuable skill - that of critical reasoning.
That act of manipulating the problem in your head—critical thinking—is ultimately a craft. And the only way to become better at it is by practicing it in a deliberate, disciplined fashion.
This is why it's pretty baffling to me when I see attempts at comparing LLMs to the invention of the calculator. A calculator is still used IN SERVICE of a larger problem you are trying to solve.
With that said, I do worry that losing the ability to craft sentences (or code) is more problematic than losing the ability to do mental math.
Losing the ability to do calculations by hand on a piece of paper with a pencil probably actually is a big deal
When I went to school we still had to do a lot of calculations by hand on paper. Thus, if I use a calculator to get an answer, I'm capable of reproducing the answer by hand if necessary
With math, at least when I was learning it, we seemed to understand that the calculator is a useful tool that doesn't replace the need to develop underlying skills
I'm seeing the exact opposite behavior and mentality from the AI crowd. "You don't need to learn how to do that correctly anymore, you can just have the AI do it"
"Vibe Coding", literally the attitude that you don't need to understand your code anymore. You just have the AI generate it and go off vibes to decide if it's right or not
Yeah, I don't know how my car engine works. But I trust that the people who engineered it do, and the mechanics that fix it when it breaks do. There's no room for "Vibe Bridge Building" in reality
Anyone advocating for "Vibe coding" is an admission that it doesn't actually matter if the thing they build works or not
Unfortunately that seems to be a growing portion of software
In the US :-)
And those skills are entirely context dependent. You're likely saying this from a SW engineer's point of view. Whereas I've worked in teams with physicists and electrical engineers. When you're in a meeting, and there is a technical discussion, and everyone can calculate in their head the effects after integrating a well known function and how that will impact the physical system, while you have to pull out a calculator/computer, you'll be totally lost.
You can argue that you could be as productive as the others if you were given the time (e.g. doing this on your own at your own cubicle), but no one will give you that time in meetings.
https://www.reddit.com/r/interestingasfuck/comments/13jhckh/...
Source?
Fortunately, my ADHD-addled brain doesn't need some fancy AI to make its cognition "Atrophied and Unprepared"; I can do that all on my own, thank you very much.
Nah, it was already at zero before ChatGPT came to public attention.
Google helps you find things that you process later on with your brain.
With AI your brain shuts off as you offload all thinking to asking questions. And asking questions is not thinking. Answering them is.
I'm willing to bet that there were a lot of Google searches, pre-ChatGPT, that effectively were questions. Lots of "huh, I wonder" during conversations and the first result was taken as "the truth".
Whereas, you can ask an LLM to speak to you in e.g. Spanish, about whatever topic you're interested in, and be able to stop and ask it to explain any idioms or vocabulary or grammar in English at any time.
I found this to be more like a "cognitive gym". Maybe we're just not using the tools beneficially.
I think that's really just it, and I agree with you. There are many other areas LLMs can, and should, be more useful and effort put toward both assisting and automating.
Instead, the industry is focusing on creative arts and software development because human talent for that is both limited and expensive, with a third factor of humans generally being able to resist doing morally questionable things (e.g., what if hiring for weapons systems software becomes increasingly difficult due to a lack of willingness to work on those projects, likewise for increasingly invasive surveillance tech, etc.)
We're rushing into basically the opposite of what AI should do for us. Automation should work to free us up to focus more on the arts and sciences, not take it away from us.
Greed at its finest.
the topics are constrained
I don't depend on AI for anything. I am not doing corporate work. Could it be that what people are experiencing is that they are becoming less suitable for corporate work as AI and robots replace them? Isn't this a good thing? Shouldn't the focus be on using AI to bring out the innate talents of humans that aren't profit driven?
The current "AI" tech is in fact developed FOR the profit. There is 0 concern at the capital investing level for enhancing any innate human talents. In fact the effort is explicitly intended to REPLACE humans with automation in tasks that have traditionally required those innate human skills.
I do believe you find learning and research to be enhanced, and I agree in general that the tech has a great deal of possible benefit, but sadly that's not what ownership hopes to use it for (statistically speaking that is. not all ownership is created equally).
This is the luddite problem all over again: it's not the looms that are the problem, it's ownership's interest in shifting funds that would be paid to workers into the coffers of corporate profit.
Could advances in technology benefit ownership, and still insure that labor can earn a decent living? Of course it COULD! It's just that given the precedent of human history, that's not how it WILL be used...
For the same reason, we were required to have graphing calculators, but not TI-92 or similar models in calculus class. While the utility is fine for people who have already learned the concepts, attempting to learn advanced algebra or other math with a symbolic solver available cripples your long term retention.
The question of "what is motivating the task" doesn't really factor very well into "how does this tool affect a novice", at least not in any similar circumstance I have seen.
I just don't know how much is actually being replaced though. I think of corporate jobs I have done in that past. I can't think of anything I have ever been paid to do that would be replaced by a language model. It was either something that could have been automated without a language model but was not for various reasons or the output would just be amplified by a language model. In some cases my work would have been enormously amplified and better but not "automated".
For some reason we don't seem to like this idea of a cybernetic relationship with a machine that benefits the human even though that is exactly what we have been doing for at least a 150 years. Maybe it is something in our brains that can't turn off a type of predator/prey model. Then on top of that is the mass appeal of this infantile and collectivist idea that AI will do all the work while we collect our UBI trust fund allowance from artificial daddy.
Or, since LLMs seem to be addictive, it's like getting rid of the spinach farms and replacing them with opium poppies. (I really hate this tech.)
Let's... not do that for brainrot.
Yes, far different, because we can still go to the gym and throw medicine balls around or swing kettle bells and do dead lifts and squats if we want to stay fit.
There is no substitute for exercising our ability to logically construct deterministic, hardened, efficient data flow networks that process specific inputs in specific environments to produce specific changes and outputs.
Maybe I'm the only one who understood the most important factor the eminent Leslie Lamport explained in grisly detail the other day, that, namely, logical thinking is both irreplaceable and essential. I'll add that that nerdiest of skillsets is also withering on the vine.
"Enjoy." --Daniel Tosh
Factorio?
“[A] key irony of automation is that by mechanising routine tasks and leaving exception-handling to the human user, you deprive the user of the routine opportunities to practice their judgement and strengthen their cognitive musculature, leaving them atrophied and unprepared when the exceptions do arise,” the researchers wrote.
But these have always been issues that humans commonly struggle with so idk.
That said, the counter to my own counter is "do I really need to memorize that?" Yes yes no internet and I'm screwed... but that's such a rare edge case. I am able to quickly find the command and knowing that it is stored somewhere else may be enough knowledge for me rather than memorization. I can see Gen AI falling into a similar design, I don't need to know explicitly how to do something, just that that task can be resolved through an LLM prompt.
Granted, we're still trying to figure out how to communicate with LLMs and we only really have 3 years of experience. Most of our insights have come from blog posts and a handful of research articles. I agree that Gen AI laziness is a growing issue, but I don't think it needs to go full Idiocracy sensationalist headline.
Communicating with 'guess-the-next-token' machines is just ELIZA version X.Y.
You can also ask your dog it they want another treat, or if they want to play Quake. They're merely listening for key words and tone of voice and reacting to them according to their experience with that matrix. And LLMs don't even understand tone of voice, and they never will.
My favorite part of AI discourse is the confident assertions that AIs will never be able to do things that they've been doing for months.
> In this paper, we aim to address this gap by conducting a survey of a professionally diverse set of knowledge workers ( = 319), eliciting detailed real-world examples of tasks (936) for which they use GenAI, and directly measuring their perceptions of critical thinking during these tasks
So, they asked people to remember times they used AI, and then asked them about their own perceptions about their critical thinking when they did.
How are we even pretending there is serious scientific discussion to be had about these "results"?
> Abstract The rise of Generative AI (GenAI) in knowledge workflows raises questions about its impact on critical thinking skills and practices. We survey 319 knowledge workers to investigate 1) when and how they perceive the enaction of critical thinking when using GenAI, and 2) when and why GenAI affects their effort to do so. Participants shared 936 first-hand examples of using GenAI in work tasks. Quantitatively, when considering both task- and user-specific factors, a user’s task-specific self-confidence and confidence in GenAI are predictive of whether critical thinking is enacted and the effort of doing so in GenAI-assisted tasks. Specifically, higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking. Qualitatively, GenAI shifts the nature of critical thinking toward information verification, response integration, and task stewardship. Our insights reveal new design challenges and opportunities for developing GenAI tools for knowledge work.
It is be presented at CHI Conference https://chi2025.acm.org/
https://en.wikipedia.org/wiki/Conference_on_Human_Factors_in...
Basically, we might need to standardize 10-20% of work time being used to “keep up” automatable skills that once took up 80+% of work time in fields where AI-based automation is making things more efficient.
This could even be done within automation platforms themselves, and sold to their customers as an additional feature. I suspect/hope that most employers do not want to see these automatable skills atrophy in their employees, for the sake of long-term efficiency, even if that means a small reduction in short-term efficiency gains from automation.
I wish you were right, but I don't think any industry is realistically trending towards thinking about long term efficiency or sustainability.
Maybe it's just me, but I see the opposite, constantly. Everything is focused on the next quarter, always. Companies want massive short term gains and will trade almost anything for that.
And the whole system is set up to support this behavior, because if you can squeeze enough money to retire out of a company in as short a time as possible, you can be long gone before it implodes
I find this is similar in my experience with AI: I pick up tidbits and tricks from AI when it's doing something I'm familiar with, but if I have it working with a completely novel framework or language it quickly races ahead and I'm essentially steering it blind, which inevitably fails.
Submitters: please don't post paywalled articles unless there are workarounds (such as archived copies).
You are tremendously better off getting a bad grade doing your own work than getting a good one using ChatGPT.
I work in a high school, so I've seen this first hand. To be fair, this mindset isn’t entirely their fault. Their parents, their future universities, and society as a whole place a high value on getting top grades, too..
In a system where college admissions are highly competitive, and where cheating with AI offers a high reward and low risk, even students who genuinely care about their learning will feel pressured to follow suit. Just to remain in the game.
Common activities provided by these gyms include fixing misconfigured printers, telling a virtual support customer to turn their PC off and back on again, and troubleshooting mysterious NVIDIA driver issues (the company has gone bankrupt 5 years ago, but their hardware is still in great demand for frustration tolerance training).
I've also turned to AI in side projects, and it's allowed me to create some very fast MVPs, but the code is worse than spaghetti - it's spaghetti mixed with the hair from the shower drain.
None of the things I've built are beyond my understanding, but I'm lazy and it doesn't seem worth the effort to use my brain to code.
Probably the most use my brain gets every day is wordle
With using GenAI (and/or "being a manager") aren't they somewhat inversely related?
I find implementation level programmers to generally be poor at stating specifications. They often phrase problems in terms of lacking their desired solutions. They jump straight to implementation.
But a manager has to get skilled at giving specification: being clear about what they expect, without stating how to do it. And that's a skill that needs to be quickly developed to use GenAI well as well. I think getting good at specifying is definitely worthwhile, and I think GenAI is helping a lot of people get better at that quickly.
Overall, it seems that should very much be considered part of "critical thinking".
While much is made of the 'diminished skill for independent problem-solving' caused by over-reliance, is there a more salient KPI than some iteration of this 'Synthetic Thinking Effort' by which to baseline and optimise the cost/benefit of AI usage versus traditional cognition?
"Impact of Gen AI on Critical Thinking: Reduction in Cognitive Effort, Confidence"
"Impact of AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort"
"The Impact of Generative AI on Critical Thinking: Reductions in Cognitive Effort"
Actual title of the paper:
"The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers"
Previous discussion:
10 Feb 2025 17:01:08 UTC https://news.ycombinator.com/item?id=43002458 (1 comment)
10 Feb 2025 22:31:05 UTC https://news.ycombinator.com/item?id=43006140 (0 comments)
11 Feb 2025 11:14:06 UTC https://news.ycombinator.com/item?id=43011483 (0 comments) [dead]
11 Feb 2025 14:13:36 UTC https://news.ycombinator.com/item?id=43012911 (1 comment)
12 Feb 2025 01:47:16 UTC https://news.ycombinator.com/item?id=43020846 (0 comments) [flagged] [dead]
14 Feb 2025 15:54:57 UTC https://news.ycombinator.com/item?id=43049676 (1 comment)
15 Feb 2025 12:06:01 UTC https://news.ycombinator.com/item?id=43057907 (101 comments)
Right now I’m curious to see how long I can keep up with those using AI for more mundane assistance. So far, so good.
If you don't understand what's happening, you have no way to know if the system is working as intended. And understanding (and deciding) exactly how the system works is the really hard part for any sufficiently complex project.
I only tend to do that when I am tired or annoyed, but when I do it I can feel myself getting dumber. And it’s a weirdly satisfying feeling.
I just need a chair that doubles as a toilet and I’ll be all set.
Critical thinking is a skill that requires practice to improve at and maintain it. Using LLMs pushes the task that would require critical thinking off to something/someone else. Of course the user will get worse at critical thinking when they try to do it less often.
I also believe, however, that humans who are able to reason properly would become much more valuable, because of this same thing.
We have the cognition science to make it happen - or at least learn how to structure it.
Some discussion on the study: https://news.ycombinator.com/item?id=43057907
I don't build or rely on pre-prompted agents to automate specific problems or workflows. Rather, I only rely on services like ChatGPT or Claude for their generic reasoning, chat, and "has read the entire web at some point" capabilities.
My use-cases break down into roughly equal thirds:
---
1. As natural-language, iteratively-winnowing-the-search-space versions of search engines.
Often, I want to know something — some information that's definitely somewhere out there on the web. But, from 30+ years of interacting with fulltext search systems, I know that traditional search engines have limitations in the sorts of queries that'll actually do anything. There are a lot of "objective, verifiable, and well-cited knowledge" questions that are just outside of the domain of Google search.
One common example of fulltext-search limitations, is when you know how to describe a thing you're imagining, a thing that may or may not exist — but you don't know the jargon term for it (if there even is one.) No matter how many words you throw at a regular search engine, they won't dredge up discussions about the thing, because discussions about the thing just use the jargon term — they don't usually bother to define it.
To find answers to these sorts of questions, I would have previously ask a human expert — either directly, or through a forum/chatroom/subreddit/Q&A site/etc.
But now, I've got a new and different kind of search engine — a set of pre-trained base models that, all by themselves, perform vaguely as RAGs over all of the world's public-web-accessible information.
Of course, an LLM won't have crystal clarity in its memory — it'll forget exact figures, forget the exact phrasing of quotations, etc. And if there's any way that it can be fooled or misled by some random thing someone made up somewhere on the web once, it will be.
But ChatGPT et al can sure tell me the right jargon term (or entire search query) to turn what was previously, to me, almost deep-web information, into public-web information.
---
2. As a (fuzzy-logic) expert system in many domains, that learned all its implications from the public information available on the web.
One fascinating thing about high-parameter-count pre-trained base models, is that you don't really need to do any prompting, or supply any additional information, to get them to do a vaguely-acceptable job of diagnosis — whether that be diagnosing your early-stage diabetic neuropathy, or that mysterious rattle in your car.
Sure, the LLM will be wrong sometimes. It's just a distillation of what a bunch of conversations and articles spread across the public web have to say about what are or aren't the signs and symptoms of X.
But those are the same articles you'd read. The LLM will almost always outperform you in "doing your own research" (unless you go as far as to read journal papers — I don't know of any LLM base model that's been trained on arXiv yet...). It won't be as good at medicine as a doctor, or as good at automotive repair as an automotive technician, etc. — but it will be better (i.e. more accurate) at those things than an interested amateur who's watched some YouTube videos and read some pop-science articles.
Which means you can just tell LLMs the "weird things you've noticed lately", and get it to hypothesize for you — and, as long as you're good at being observant, the LLM's hypotheses will serve as great lines of investigation. It'll suggest which experts or specialists you should contact, what tests you can perform yourself to do objective differential diagnostics, etc.
(I don't want to under-emphasize the usefulness of this. ChatGPT figured out my house had hidden toxic mold! My allergies are gone now!)
---
3. As a translator.
Large-parameter-count LLM base models are actually really, really good at translation. To the point that I'm not sure why Google Translate et al haven't been updated to be powered by them. (Google Translate was the origin of the Transformer architecture, yet it seems to have been left in the dust since then by the translation performance of generic LLMs.)
And by "translation", I do literally mean "translating entire documents from one spoken/written human language to another." (My partner, who is a fluently-bilingual writer of both English + [Traditional] Chinese, has been using Claude to translate English instructions / documents into Chinese for her [mostly monolingual Chinese] mother to better understand them; and to translate any free-form responses her mother is required to give, back into English. She used to do these tasks herself "by hand" — systems like Google Translate would provide results that were worse-than-useless. But my partner can verify that, at least for this language pair, modern LLMs are excellent translators, writing basically what she would write herself.)
But I also mean:
• The thing Apple markets as part of Apple Intelligence — translation between writing styles (a.k.a. "stylistic editing.") You don't actually need a LoRA / fine-tune to do this; large-parameter-count models already inherently know how to do it.
• Translating between programming languages. "Rewrite-it-in-Rust" is trivial now. (That's what https://www.darpa.mil/research/programs/translating-all-c-to... is about — trying to build up an agentive framework that relies on both the LLM's translation capabilities, and the Rust compiler's typing errors on declaration change, to brute-force iterate across entire codebases, RiiRing one module at a time, and then recursing to its dependents to rewrite them too.)
• Translating between pseudocode, and/or a rigorous description of code, and actual code. I run a data analytics company; I know far more about the intricacies of ANSI SQL than any man ought to. But even I never manage to remember the pile of syntax features that glom together to form a "loose index scan" query. (WITH RECURSIVE, UNION ALL, separate aliases for the tables used in the base vs inductive cases, and one of those aliases referenced in a dependent subquery... but heck if I recall which one.) I have a crystal-clear picture of what I want to do — but I no longer need to look up the exact grammar the SQL standard decided to use yet again, because now I can dump out, in plain language, my (well-formed) mental model of the query — and rely on the LLM to translate that model into ANSI SQL grammar.
The same is true of managers. I have had managers who yelled at me to do things they did not understand. They rotted on the inside. Other managers learned every trick I brought to the company. They grew.
At the end, I spent probably more time and learnt nothing.. My initial take was that this is the kind of thing I don't care much for so giving it to a llm is OK... However, by the end of it I ended up more frustrated and lost it in the simulation of working things out aa well
Comedians' ability diminishes as they take time off.
Ahnold wasn't lounging around all day.
We should understand that fixing crap, unsensible code is not a productive skillset. As Leslie Lamport said the other day, logically developing and coding out proper abstractions is the core skillset, and not one to be delegated to just anything or anyone.
It's ok; the bright side for folks like me is that you're just happily hamstringing yourselves. I've been trying to tell y'all, but I can only show y'all the water, not make you drink.