A new and growing fear I have is that AI cuts off some of the most well-trodden pathways to intellectual growth. Every time the AI solves a problem that I would not have been able to efficiently solve alone, it has replaced an interaction that previously would have happened with a mentor/supervisor/code reviewer. This is faster for the person seeking the micro-assistance, but there is a flipside: fewer human-to-human acts of mentorship and shared problem solving. We lose a lively ingredient of team formation, expert formation, and a source of joy for all involved. As this scales up it feels plausible that we end up with a wider deagradation of intellectual standards. I like the Cognitive Comons framing.
I think this has already happened to some extent, with social media etc. crowding out reading books. Even higher quality stuff like HN pales in comparison with what a good book brings, and it's a daily struggle to keep reading.
At least we're (mostly) not scrolling feeds all day at work. AI is going to be like this for all our professional competencies.
>Every time the AI solves a problem that I would not have been able to efficiently solve alone, it has replaced an interaction that previously would have happened with a mentor/supervisor/code reviewer.
I suppose that Socrates would strongly agree, seeing how the same argument could be applied to learning from a book.
The pseudo-academic style of the article is awful. All abstractions and buzzwords, no examples. Also, the tragedy of the commons concept is far older than 1968. The term "common" refers to common grazing land upon which anyone could graze animals.
That's an English term. The American equivalent is "open range".
Getting past that, the author has a point. There's a loss of shared expertise when there aren't people around learning and doing something. In the US, we've seen this in manufacturing. The number of Americans who know how to set up a good production plant is much lower than it was in the 1980s. That's the consequence of the hollowing out of American manufacturing. The author talks about AI vs. white collar work, but fails to make the connection with outsourcing vs. blue collar work.
The loss of this expertise has recently been made very clear in the US as attempts are made to scale up weapons production for the US's various wars. Progress is very slow, as has been seen with both artillery ammo and air-defense missiles.
For anyone that does LLM supported work it's crystal clear that skill decline is real. Surprisingly, the leadership at my work doesn't give rats ass about it. They published AI engineering manifesto, pushing for loops, workflows, advocating for producing many times more code. All of this ignoring the elephant in the room.
Cynically they may be banking (however misguided it turns out to be) on model skill in aggregate growing faster than the skills of their staff decline at which point, bin the staff, use the model.
The hype from non-technical people about AI would make you think they already discovered AGI.
I don't know, I find AI still misses the mark so much for stuff that's not happening inside a well constrained framework with very specific guardrails and boundary conditions. Every time I have agents build something half complex that I would be easily able to pull off myself (albeit not at their pace) I always ask if everything is correct and working and I get a resounding "yes!", then I test it and it's essentially a steaming pile of garbage which I then tell the agent and after which it goes "you're absolutely right to push back, this isn't good at all!" as if it was clear from the beginning. I don't know if we'll achieve real intelligence soon with these things but what we have now isn't there yet. It's incredibly useful but I don't think you can replace a good programmer with it, you can make a good programmer into an excellent one though!
Same goes for writing, anyone who has written long complex texts with LLMs knows that a ton of editing is required to make it half decent.
It doesn't seem clear to me at all. e.g. I've been having a fun time learning Godot this last week with codex laying down groundwork for logic, generating assets, etc. This gives me something concrete that I can work with and ground myself within instead of starting from a blank canvas. I don't much care for the actual scripting logic because it's all basically trivial to me with a few decades of programming experience, and I'm not going to be able to develop my artistic abilities to a useful capacity anytime within the next few years, so I can instead focus on learning architecture/organization patterns and game systems that I'm interested in while I read through docs.
So basically, if you use it to wave away things you've mastered or explicitly don't want to learn at the moment, you can stay focused on things you are figuring out. I don't see it as any different from writing "by induction" in a math proof without writing all the details because you and the reader know you could easily work them out. In this way I can learn things that I simply would not have the time to dig into before, increasing my skillset. As is was before AI, metacognition is the most widely useful skill to have.
I don’t think it’s as clear cut as you make it out to be.
Maintenance <> growth
Once you got a skill it’s quite hard to actually lose it. It’s like riding a bike as they say.
I find all the reading of those walls of code exercises my mind better than the writing did. I’m sure I lost memories of how to read a file by hand or balance a tree but who cares?
I am thinking about architecture and design a lot more these days and every decision I make has to be actually argued for even to myself. I can no longer lean back and say “that’s how we always do it” or “too expensive to change now” as I find many people in practice actually do. They were just coasting on premade architectural choices and their “skills” consisted of knowing arcane incantations and syntactic details completely unrelated to the (business) problem at hand.
I am not convinced many developers actually have the skills they think they have. They could wrestle syntax and mess around with tooling, but could they abstract properly? Define clear semantic boundaries? Have proper civil discussions about responsibilities and where they should lie on the right level of abstraction? Nothing has changed in that regard. If anything that part has been amplified. (“taste”)
I think as an existing expert in my field (robot devops) what I feel most strongly is not that I'm losing my skills but rather than I'm losing my nerve.
Like, problems that I would previously dig into on my own, poking at this and that log file to try to understand what happened to get a system into a particular state, now claude is mediating most of those kinds of interactions and it is the one doing the first pass surfacing of "okay I discovered X, Y, and Z things that are hinky, and I'm not totally sure yet what this all means, but let's look together."
Nowadays I feel almost naked looking at a terminal where I'm typing each character myself. A lot of the old instincts around tab completion and grepping through the --help output of every tool, all that stuff has atrophied somewhat. Maybe that is genuine skill decline?
I think this is a really important idea. Good problem solvers, in general, are people who are prepared to go in and poke around, to see what happens when you disable this or add that.
The only way I've found to make juniors develop that instinct is to make them do it until it comes naturally--to continually demand they verify their assumptions, test out theories while debugging. If you don't practice being comfortable with the unfamiliar, you lose the knack.
It's fine, of course, to lose the knack to AI, as long as you'll never need to solve a problem the AI can't (or can't access). I don't think that's a good assumption for everyone to make.
> More people that are not devs, are using AI to create things, and are skilling up to a mediocre level?
What skills are being developed by non-devs prompting AI to churn out code they don't understand? Note that I'm not debating if the code is good or works or whatever, I'm asking if any real skills are being developed merely by prompting AI towards some goal.
If they are producing code, there is some knowledge being absorbed. Nobody is using AI to create an app is learning zero.
Its just what they are learning, is not as much as they think, and really random, not-structured. Like reading just a few random chapter of a book across multiple classes. They get a smattering of ad-hoc tidbits of knowledge.
I am really worried if using AI is reducing someone's already gained knowledge. Moving someone backwards. That has me a bit more scared.
> If they are producing code, there is some knowledge being absorbed. Nobody is using AI to create an app is learning zero.
I think the knowledge being absorbed is 'how to use AI to make apps', not necessarily anything about the code itself. I've seen lots of people who are very effective with AI suddenly poleaxed in interviews without it. Going from 'able to explain exactly what needs to happen' to 'staring at a blank page with no ideas'.
>I am really worried if using AI is reducing someone's already gained knowledge. Moving someone backwards. That has me a bit more scared.
It absolutely is. It removes a lot of friction from pathways that kept previous knowledge firm.
This isn't unique to AI. When calculators came around, people had a new tool that decreased the need to do arithmetic in your head. This has lead to a decline in the percentage of people who knew arithmetic, but can no longer do it adequately without the external tool. Another example would be assembly skills declining amongst software engineers after compiled languages become dominant.
What I would argue is unique to AI is two-fold. The first being that the scale of things being automated is enormous. Keeping with just software, AI isn't only a tool that decreases the need to write code, it's also a tool that decreases the need to do your own research, debugging, system design, version control, etc. The list of tasks a regular engineer is responsible for that have not been consumed by AI is small. To make matters more frightening, this is just a single industry. The same applies to multiple other knowledge heavy fields (e.g. mathematics).
The second thing I'd argue is unique about AI is the addictive nature of it. These LLMs are deliberately coerced into being sycophants that blow smoke up your ass unless you deliberately tell the model not too (and it may still do it anyways!). A calculator didn't compliment your insight for asking it to calculate the square root of -1.
Furthermore, many people have talked about the resemblance of LLMs to slot machines--there's a feeling that anything could be possible if you only prompt the model correctly. Your first prompt is wildly off the mark, but your next prompt is better though still not quite right, so you continue forward. You keep iterating your instructions, your word choices, your tone, all in search of the desired response. In this way, an LLM is a slot machine, your prompt the lever you pull, and the output is the jackpot.
All this taken together, we have a machine that decreases the need to engage with the critical faculties of your own brain, which has been trained to keep your attention through flattery, and whose probabilistic nature preys on our love for gambling.
AI is good enough to generate a first pass for an increasingly large number of projects. An engineer using AI can create an even larger amount.
But muscle memory and expertise are not fixed. They fade over time without use. If you're no longer writing code yourself, you'll get rusty on syntax in the short term. In the long term, you'll get rusty on code structure and layout.
But many engineers are no longer reviewing code either. Reviewing code written by an AI is now the bottleneck, so you're expected to allow AI to review it as well.
So we have a group of people who are no longer engaging in either the writing of the code, or the analysis of the code that's written.
Obviously, this would lead to skill loss.
Many would argue that they're not truly losing skills, because they're more engaged in the grander architecture of the code. To that I would say--your job title says engineer, not architect.
Similar to the argument made by John Blow several years ago. Though, Blow blamed frameworks/engines and layers of abstractions. I wonder if the authors came to this framing themselves, or listened him.
My approach is to write a bit of totally AI-free code every day. So after a day of Claude, I'll spend at least 30 mins wrestling with something. The gnarlier the better, e.g.leet code or Project Euler-type stuff. I think of it as like lifting weights for the mind. The more I struggle at the edge of my knowledge and skill the better.
The other thing is to give your agent a skill not to solve certain key problems unless explicitly prompted. Write the scaffolding sure, but leave the juicy parts alone. And if I get stuck, I have it enter into a dialogue with me, nudging me towards understanding.
I solved a Leetcode Medium on my first try in about 10 minutes the other day and it felt so amazing. Thinking through architecture, project composition, composability, etc when working with agents felt fine for a while but it made me feel disconnected from what I'm actually building.
Actually engaging my brain to solve the lower-level, on-the-ground code allows me to think of better ways to do things while I'm writing them. It's like writing anything. You start with something you want to convey, a thesis, and then it evolves and becomes better as you write it. An agent will just write it with no thought, as in it will reflect one of the LLM 'ghosts' as Andrej Karpathy puts it, doing something in the same way that someone in the training data has done it on a similar or different problem. This is why I get conniptions now when I am sent generated text or am expected to read it on a public forum. It's disrespectful of the time of every person expected to read it.
That said, I wouldn't like to go back to the before times without having the agentic option available. Ideally, businesses should not mandate how LLM's are to be used at their company, and just let the devs find their own flow. That is, if quality is even a factor that any company optimizes for anymore.
for a lot of the article i was mentally replacing "ai" with "calculator" and going "yeah we will just need longer education times. more schooling etc."
but the phrasing of "who pays for the increased schooling times?" is a good one.
i think "debt" can be a decent way to conceptualize the cost and repayment of training someone. feels evil to say, but viewing people as firms you can invest in and expect returns upon. you know not all loans will be repayed, but hopefully they'll average to a profit. (risk management etc.)
student loans are. a decent example. the government/private enterprise gives money to pay for education, then this is repayed, providing a financial incentive for paying for someone else's longer schooling timelines. firms investing in training can be viewed as an extension of student loans. but then ah, there are countless stories of how debtor/creditor relationships can be exploited. indentured servitude etc. there are a lot of complications coming to mind. also "altruistic" people who give without expectation of repayment. or the divide between like, communal vs individualistic cultures. (individualism, i argue, encourages the formalization of debt, as opposed to a more communal culture where the expectation of repayment is informal.) you could do math on how many people pay vs how many people benefit, who is the biggest stakeholder, etc.
but i am on my lunch break and need to get back to my work. good article tho. good topic to bring up.
Perhaps, but I find when I use AI (for fixing things), I learn a few things here and there as the AI-provided answers are often wrong. And when I work around these errors, the learning occurs.
At least we're (mostly) not scrolling feeds all day at work. AI is going to be like this for all our professional competencies.
I suppose that Socrates would strongly agree, seeing how the same argument could be applied to learning from a book.
Except there's mass directives to eat as much seed corn as you can
Getting past that, the author has a point. There's a loss of shared expertise when there aren't people around learning and doing something. In the US, we've seen this in manufacturing. The number of Americans who know how to set up a good production plant is much lower than it was in the 1980s. That's the consequence of the hollowing out of American manufacturing. The author talks about AI vs. white collar work, but fails to make the connection with outsourcing vs. blue collar work.
The loss of this expertise has recently been made very clear in the US as attempts are made to scale up weapons production for the US's various wars. Progress is very slow, as has been seen with both artillery ammo and air-defense missiles.
The hype from non-technical people about AI would make you think they already discovered AGI.
How people in tech are aware they're training their replacements?
Or are we quickly approaching an apex where the people running these companies realize AI cannot completely replace human developers?
Same goes for writing, anyone who has written long complex texts with LLMs knows that a ton of editing is required to make it half decent.
So basically, if you use it to wave away things you've mastered or explicitly don't want to learn at the moment, you can stay focused on things you are figuring out. I don't see it as any different from writing "by induction" in a math proof without writing all the details because you and the reader know you could easily work them out. In this way I can learn things that I simply would not have the time to dig into before, increasing my skillset. As is was before AI, metacognition is the most widely useful skill to have.
Maintenance <> growth
Once you got a skill it’s quite hard to actually lose it. It’s like riding a bike as they say.
I find all the reading of those walls of code exercises my mind better than the writing did. I’m sure I lost memories of how to read a file by hand or balance a tree but who cares?
I am thinking about architecture and design a lot more these days and every decision I make has to be actually argued for even to myself. I can no longer lean back and say “that’s how we always do it” or “too expensive to change now” as I find many people in practice actually do. They were just coasting on premade architectural choices and their “skills” consisted of knowing arcane incantations and syntactic details completely unrelated to the (business) problem at hand.
I am not convinced many developers actually have the skills they think they have. They could wrestle syntax and mess around with tooling, but could they abstract properly? Define clear semantic boundaries? Have proper civil discussions about responsibilities and where they should lie on the right level of abstraction? Nothing has changed in that regard. If anything that part has been amplified. (“taste”)
Is this -> More people that are not devs, are using AI to create things, and are skilling up to a mediocre level?
or
Is this -> Existing experts, are actually loosing there skills? Using AI is reducing someone already known skills.
Like, problems that I would previously dig into on my own, poking at this and that log file to try to understand what happened to get a system into a particular state, now claude is mediating most of those kinds of interactions and it is the one doing the first pass surfacing of "okay I discovered X, Y, and Z things that are hinky, and I'm not totally sure yet what this all means, but let's look together."
Nowadays I feel almost naked looking at a terminal where I'm typing each character myself. A lot of the old instincts around tab completion and grepping through the --help output of every tool, all that stuff has atrophied somewhat. Maybe that is genuine skill decline?
The only way I've found to make juniors develop that instinct is to make them do it until it comes naturally--to continually demand they verify their assumptions, test out theories while debugging. If you don't practice being comfortable with the unfamiliar, you lose the knack.
It's fine, of course, to lose the knack to AI, as long as you'll never need to solve a problem the AI can't (or can't access). I don't think that's a good assumption for everyone to make.
What skills are being developed by non-devs prompting AI to churn out code they don't understand? Note that I'm not debating if the code is good or works or whatever, I'm asking if any real skills are being developed merely by prompting AI towards some goal.
If they are producing code, there is some knowledge being absorbed. Nobody is using AI to create an app is learning zero.
Its just what they are learning, is not as much as they think, and really random, not-structured. Like reading just a few random chapter of a book across multiple classes. They get a smattering of ad-hoc tidbits of knowledge.
I am really worried if using AI is reducing someone's already gained knowledge. Moving someone backwards. That has me a bit more scared.
I think the knowledge being absorbed is 'how to use AI to make apps', not necessarily anything about the code itself. I've seen lots of people who are very effective with AI suddenly poleaxed in interviews without it. Going from 'able to explain exactly what needs to happen' to 'staring at a blank page with no ideas'.
Maybe not actually absolutely zero, but it's much closer to zero than what they would learn creating the app themselves
It absolutely is. It removes a lot of friction from pathways that kept previous knowledge firm.
This isn't unique to AI. When calculators came around, people had a new tool that decreased the need to do arithmetic in your head. This has lead to a decline in the percentage of people who knew arithmetic, but can no longer do it adequately without the external tool. Another example would be assembly skills declining amongst software engineers after compiled languages become dominant.
What I would argue is unique to AI is two-fold. The first being that the scale of things being automated is enormous. Keeping with just software, AI isn't only a tool that decreases the need to write code, it's also a tool that decreases the need to do your own research, debugging, system design, version control, etc. The list of tasks a regular engineer is responsible for that have not been consumed by AI is small. To make matters more frightening, this is just a single industry. The same applies to multiple other knowledge heavy fields (e.g. mathematics).
The second thing I'd argue is unique about AI is the addictive nature of it. These LLMs are deliberately coerced into being sycophants that blow smoke up your ass unless you deliberately tell the model not too (and it may still do it anyways!). A calculator didn't compliment your insight for asking it to calculate the square root of -1.
Furthermore, many people have talked about the resemblance of LLMs to slot machines--there's a feeling that anything could be possible if you only prompt the model correctly. Your first prompt is wildly off the mark, but your next prompt is better though still not quite right, so you continue forward. You keep iterating your instructions, your word choices, your tone, all in search of the desired response. In this way, an LLM is a slot machine, your prompt the lever you pull, and the output is the jackpot.
All this taken together, we have a machine that decreases the need to engage with the critical faculties of your own brain, which has been trained to keep your attention through flattery, and whose probabilistic nature preys on our love for gambling.
AI is good enough to generate a first pass for an increasingly large number of projects. An engineer using AI can create an even larger amount.
But muscle memory and expertise are not fixed. They fade over time without use. If you're no longer writing code yourself, you'll get rusty on syntax in the short term. In the long term, you'll get rusty on code structure and layout.
But many engineers are no longer reviewing code either. Reviewing code written by an AI is now the bottleneck, so you're expected to allow AI to review it as well.
So we have a group of people who are no longer engaging in either the writing of the code, or the analysis of the code that's written.
Obviously, this would lead to skill loss.
Many would argue that they're not truly losing skills, because they're more engaged in the grander architecture of the code. To that I would say--your job title says engineer, not architect.
https://www.youtube.com/watch?v=q3OCFfDStgM
The other thing is to give your agent a skill not to solve certain key problems unless explicitly prompted. Write the scaffolding sure, but leave the juicy parts alone. And if I get stuck, I have it enter into a dialogue with me, nudging me towards understanding.
Actually engaging my brain to solve the lower-level, on-the-ground code allows me to think of better ways to do things while I'm writing them. It's like writing anything. You start with something you want to convey, a thesis, and then it evolves and becomes better as you write it. An agent will just write it with no thought, as in it will reflect one of the LLM 'ghosts' as Andrej Karpathy puts it, doing something in the same way that someone in the training data has done it on a similar or different problem. This is why I get conniptions now when I am sent generated text or am expected to read it on a public forum. It's disrespectful of the time of every person expected to read it.
That said, I wouldn't like to go back to the before times without having the agentic option available. Ideally, businesses should not mandate how LLM's are to be used at their company, and just let the devs find their own flow. That is, if quality is even a factor that any company optimizes for anymore.
but the phrasing of "who pays for the increased schooling times?" is a good one. i think "debt" can be a decent way to conceptualize the cost and repayment of training someone. feels evil to say, but viewing people as firms you can invest in and expect returns upon. you know not all loans will be repayed, but hopefully they'll average to a profit. (risk management etc.)
student loans are. a decent example. the government/private enterprise gives money to pay for education, then this is repayed, providing a financial incentive for paying for someone else's longer schooling timelines. firms investing in training can be viewed as an extension of student loans. but then ah, there are countless stories of how debtor/creditor relationships can be exploited. indentured servitude etc. there are a lot of complications coming to mind. also "altruistic" people who give without expectation of repayment. or the divide between like, communal vs individualistic cultures. (individualism, i argue, encourages the formalization of debt, as opposed to a more communal culture where the expectation of repayment is informal.) you could do math on how many people pay vs how many people benefit, who is the biggest stakeholder, etc.
but i am on my lunch break and need to get back to my work. good article tho. good topic to bring up.