I have some sympathy for these kids. If LLMs were around when I was a student, I would've also used them to "speed up" my homework assignments then proceed to fail all my tests. Now I work mostly with PhDs who were at the top of every academic environment they've ever been in. And yet I can see their thinking skills rapidly declining as well; many of them can no longer brainstorm, code, think deeply, or write without an LLM present doing 90% of the work. Many of them can no longer sit quietly for even 30 minutes just thinking on their own, which is a required skill for producing original thought. For adults the cognitive decline won't be as measurable since there's no exams, and overall output volume will still be fine due to LLM help. But I do believe it's already happening absolutely everywhere around us. Honestly, I wanted to be in denial about it before but it's too obvious to ignore now.

Rest doesn't make sense either with the parent's tautologies and self-counters. They seem to argue for it and raise the challenges. ~"Speed doesn't matter unless you need it." ~"LLMs can be good, but if you don't use them properly™, then they become a crutch." It's hard to deny that "cognitive offloading" via LLMs is becoming a more acute problem [0]. The intelligentsia were supposed to be immune. [0] https://www.bbc.com/future/article/20260417-ai-chatbots-coul...

This is the right balance for me as well. I use an agent to generate a first-pass attempt, and then (deadlines willing), I manually read every line at least once so I understand what the code actually does. Then I manually fix the inevitable slop that is mixed in with the good stuff, and only once the code is up to my personal standards do I send it. This probably reduces my “AI performance boost” to 30-50% instead of the huge gains reported by others. But I retain the ability to reason about the codebase and use AI much more precisely when I’m trying to troubleshoot production outages or subtle bugs — something I notice the rest of my team struggles with, since adopting “agentic workflows” everywhere. I think actively working to retain some cognitive flexibility and “muscle memory” around coding tasks is going to be rather advantageous in the long run.

Pure copium, but what can you do with the deadlines.

I don’t buy it. Properly leveraging LLMs to generate stable and extendable systems is mentally exhausting (i.e. highly demanding of intelligent thought), especially given the poor quality and churn within the harness ecosystem. Now, if you’re creating trivial, unstable, or nonextendable systems maybe this doesn’t apply. And maybe I have long overestimated the work that SWEs have done.

LLMs have killed my facility but not my knowledge. I can still read code and write it, I just need to look back at docs a lot more, when I used to just know things. I also have to sit and try to recall how to do things and what abstractions are involved more. I also have more "writer's block" when starting with a fresh program/document if trying not to get AI to seed it with a baseline implementation, where I have to sit for a while thinking about what I really want to build.

The other day I just wanted to loop through characters in a std::string to copy data to a new string with a few escape characters (sending to peripheral device). Simple enough task for AI. I got a coroutine monstrocity back, with copies to std::array and a range based iterator, since I specified C++23. If I specified C++11, I would have received a: char p = src.data(); while ( p) { … p++; } I had the experience to keep calling out AI to simplify and downgrade the solution to something primitive, which ended up smaller, faster, easier to maintain. Juniors with real world experience would not bother, they’ll take the first working AI result.

Perhaps, but I'm not sure my boss would appreciate that though.

I have observed this in myself when I began to over-leverage AI in my workflows. I've since become more deliberate with what kinds of tasks I will use it for, although I still slip up. With writing: Things like brainstorming a plot line for a book with a custom GPT or Claude project that has all of my prior books in its knowledge? Works great. Things like asking it to write a paragraph or chapter for me - I can rapidly feel my own writing skill, motivation, vocabulary, and ability to grasp/remember the resulting plotlines deteriorating. I don't use it for that anymore. With studying: I've been taking a couple of evening uni courses and the thing I found so great is that I've been forcing myself to think through the problems, and take my own notes in every lecture. I may then still get ChatGPT to help explain and reason through some of the concepts with me. And I have it review and 'grade' my assignments. But I refuse to ask it to start drafting answers. With programming: This one is tougher. When I am not very personally invested in a problem or codebase it becomes too easy to offload more parts to Claude, and when the company encourages 'vibing' to speed up velocity and you're reviewing and writing a higher influx of lower quality PRs, investment goes down. I still sometimes catch myself committing solutions I only _mostly_ grasp and the rest is hand-waving. A big part of it is a work culture thing. For my own projects I make sure to understand and have a back-and-forth with the planning agent for each task, or write the first plan myself to go off of. When it comes to producing the code, I have to admit it is much easier to properly review parts of the codebase I am extra interested and knowledgeable in (backend in my case). The frontend I'm less well versed in and also admittedly less interested in, so I do sometimes fall into the trap of "Ehh it works, just commit it" with the goal of doing a thorough quality pass before actual release. With all of the above, I can feel my ability to think, plan, reason, focus (and my vocabulary) suffer if I go over the line too much into agent offloading. For me keeping that balance is as much about maintaining my own long-term brain health as it is about producing good output. I imagine younger people growing up with AI today won't even know what that more capable (in my opinion) brain state feels like - to them, the AI-using brain will be the norm.

I think this would be fine as an adult, if it meant using LLM to churn out the boring work required of you at a corporate gig, to spend more brain cycles on something you actually want to work on. The problem is that it sounds like many people are just using it for everything.

Do any of said PhDs gain anything positive from LLM usage? Or does it only lead to declining thinking skills in your view?

Yes, I can churn out a lot more stuff as can most of my peers. Experiments etc are all way faster to run with coding agents. But I think the overall creativity and originality is a lot lower. I think this is what many people are facing, if you don't use LLMs your short term productivity is worse.

There’s the saying that we overestimate what we can achieve in the short term, but underestimate what we can achieve in the long term. Optimizing for the short term is therefore counterproductive if it impairs us for the long term.

They're incredibly more productive. LLMs are amplifiers, so where they'd have branched and tried out N things, they can easily try 5N pathways of RnD. LLMs are extending the frontiers of science fast -- math -> phy -> chem -> bio in that order.

In my own experience, the only path I truly gain intellectual benefits is the one where I work closely with the LLM, test very narrow hypotheses, and leverage it for learning over producing. Trying 5N paths is useful and sometimes yields interesting insights I’ll retain, but it’s not the rich, challenging, deeply engaging kind of process I find I need in order to develop useful knowledge and skills. So yes it’s an accelerant for people who want stuff from me, but that doesn’t map directly to learning and building skills. I think that mismatching is really important.

I'm hearing different from PhDs. The bottleneck with much research isn't "trying out ideas" so much as it's all the bureaucratic minutiae, grants, mentoring PhD candidates, collaboration with other researchers, etc. I've heard LLMs can be helpful in limited targeted ways. But not as some kind of "game changing" accelerant.

To me it is crazy that you are being downvoted. My experience in academia was that an incredible amount of time was devoted to data cleansing analysis, coding, etc., which were completely non-core to the actual underlying academic pursuit.

There's an unnecessary feeling of fear that permeates any factual conversation on LLM's impact on science and engineering. You can just view the practitioner over the shoulder and see all the things they're able to do in a minute that would have taken days. The downvotes are just a sign of the times. It's also something to observe and think about..

I think it varies tremendously from one role to the next. I'm a senior software engineer and LLMs, the way I'm using them, improve almost everything I do. I use them to write most of my code now, but first I spent twenty years writing code before LLMs came into existence and second writing code is like 5% of my job. Most of my job is research, investigation, and architecture. I treat LLMs just like a junior engineer. I give them clearly defined jobs that I could do on my own just fine, that I already spent years doing. The problem here is that students are using LLMs to automate everything BEFORE they become proficient at it themselves. Letting college students use LLMs for homework is like letting kindergarteners use calculators instead of counting on their fingers.

You cannot tell me that letting anyone do something for you does not affect the skills that you outsourced, unless you are some sort of a superhuman. As an example, I have been drawing portraits for quite a few years now, and whenever I go on a hiatus and come back after a few months, I can notice my skill not being anywhere close to where it was before I stopped using it. Sure, after 2 or 3 portraits they mostly come back because of the previous experience, but skill rust is a real thing, and if you think your coding skills are the same because you used to code 20 years but haven't coded for some time, you are probably just lying to yourself.

This likely varies person by person or the way people adapted AI. For me AI replaced the boring part of writing code, but has not replaced the fun part of thinking about code and problem solving.

> Now I work mostly with PhDs who were at the top of every academic environment they've ever been in. And yet I can see their thinking skills rapidly declining as well; many of them can no longer brainstorm, code, think deeply, or write without an LLM present doing 90% of the work. Many of them can no longer sit quietly for even 30 minutes just thinking on their own, which is a required skill for producing original thought. This was my experience even pre-LLMs though (about my own PhD thinking skills too). I blame the amount of random stuff work now involves more than LLMs.

> Yes, if they are "responsible" for the code delivered, where responsible means they understand the code, the architecture, the decisions made, etc. This is a good principle to maintain, I think. I'm not a professor, but I manage a team of about a dozen people. The maxim I have is: "You're responsible for anything that hits git." Don't care if the LLM generated it, or the LLM told you if it's a good idea. If you commit it, you are endorsing it as a good idea - so you're the one I'm going to ask about it. I see the same principle at work in your pedagogy. > I'm not anti-AI (and really, what could I do if I were?) since I use it myself, I'm just anti-slop, especially from my students. This hits. Especially this part: > and really, what could I do if I were? My completely unsolicited opinion: you're doing a responsible thing by teaching these students how to use AI as a reference, and keeping them honest about not using it as a substitute for their own critical thinking.

You are confusing faculty with records. The ability to navigate by sense of direction. The ability to memorize numbers. The ability to think clearly by yourself. Watcha gonna do if big tech takes away your access to the outsourced brain, dear?

My experience (n=1) is that while I'm definitely lazier on certain tasks, AI has opened up some much more complex tasks. There are many tasks which I still carry out which I don't trust AI with. Maybe it's a result of the codebase I work with being fairly complicated and math heavy, but I'd say the overall outcome for me has been: lazier application on the easy tasks, mind opening on the harder tasks.

Even as a software engineer myself, I'm feeling a bit of cognitive decline having AI doing some/most of the thinking for me The solution? I'm not sure but possibly use AI as more of a collaborate partner to discuss with rather than letting it give you the answers

All of this makes me selfishly excited for my own future. It's glaringly obvious that anyone who's a heavy user of LLMs is atrophying their skills in real-time. I have yet to meet a single person for whom it's not the case. But I essentially completely stopped using them for software engineering (why isn't really relevant, but it's not because od this skill atrophy). So as the skills of everyone else is diminishing, mine is proportionally raising. It has never been easier to get better than others. You don't need to put in more effort, just the same effort as you always have, and others will do the job of losing their skills for your own benefit.

It’s only going to get worse. The second things like Claude Cowork get opened up to non-technical teams you start to see the influx of emails and Slack message written with LLM’s for absolutely no productivity gain (in fact probably a loss given how unnecessarily wordy the messages are). Too many people want to give up any and all responsibility.

I will add, too, although less relevant to education than just generally - for all the talk that these tools must be useless and incorrect, that just plainly does not map to my experience using these tools. AI can chew through a debug log on a custom system and pick out root causes on behaviors very effectively, in my experience. It is just hard to reconcile that denigration of AI with the typical experience I have using these tools in the real world. It is not omnipotent or God, but it can effectively assist in work. There is a certain cognitive dissonance I feel when I walk away from using the tool to help accomplish particular tasks, then hear over and over people say the technology is fundamentally useless and fundamentally does not work. I guess I am just not enough of an academic to understand how something can accomplish work yet fundamentally isn't, somehow.